CMU Advanced NLP 2024 (18) Knowledge and Language Models/transcript.vtt

WEBVTT

00:00:00.120 --> 00:00:04.880
everyone I today I'd like to talk about

00:00:02.760 --> 00:00:07.399
uh learning from knowledge bases uh

00:00:04.880 --> 00:00:11.440
learning from in for knowledge bases

00:00:07.399 --> 00:00:14.799
this is kind of a a shift uh from a lot

00:00:11.440 --> 00:00:16.480
of the stuff that we've done so far uh

00:00:14.799 --> 00:00:18.439
and I'm going to be talking about like a

00:00:16.480 --> 00:00:20.480
different information Source some

00:00:18.439 --> 00:00:21.960
relatively different algorithms compared

00:00:20.480 --> 00:00:26.080
to the stuff that we talked about up

00:00:21.960 --> 00:00:28.880
until this point so um you know it might

00:00:26.080 --> 00:00:32.360
be uh interesting it might be different

00:00:28.880 --> 00:00:35.640
so uh get started with

00:00:32.360 --> 00:00:37.360
that so I'm going to be talking about

00:00:35.640 --> 00:00:40.000
knowledge bases and knowledge bases are

00:00:37.360 --> 00:00:43.039
basically a structured databases of

00:00:40.000 --> 00:00:46.079
knowledge and they can contain a lot of

00:00:43.039 --> 00:00:48.559
things but most commonly when people are

00:00:46.079 --> 00:00:50.600
talking about them they are talking

00:00:48.559 --> 00:00:53.160
about relational knowledge bases that

00:00:50.600 --> 00:00:55.559
include things like entities which are

00:00:53.160 --> 00:00:57.399
nodes in a graph and relations which are

00:00:55.559 --> 00:01:00.239
edges between

00:00:57.399 --> 00:01:02.079
nodes and

00:01:00.239 --> 00:01:03.879
I'll I'll talk about some examples of

00:01:02.079 --> 00:01:05.479
this in a little bit to make that a

00:01:03.879 --> 00:01:08.040
little bit more concrete and then some

00:01:05.479 --> 00:01:11.240
of the questions that we ask about these

00:01:08.040 --> 00:01:14.400
are how can we learn to create and

00:01:11.240 --> 00:01:16.799
expand knowledge bases with uh you know

00:01:14.400 --> 00:01:18.439
neural network based methods and then

00:01:16.799 --> 00:01:20.200
the second question is how can we learn

00:01:18.439 --> 00:01:22.600
from the information in knowledge bases

00:01:20.200 --> 00:01:24.720
to improve like neural network models or

00:01:22.600 --> 00:01:27.560
uh use them in effective

00:01:24.720 --> 00:01:31.479
ways and how can we use uh structured

00:01:27.560 --> 00:01:31.479
knowledge to answer questions

00:01:32.200 --> 00:01:37.159
so the first uh thing I'd like to talk

00:01:35.000 --> 00:01:40.960
about a little bit is types of knowledge

00:01:37.159 --> 00:01:43.079
bases and they come in several different

00:01:40.960 --> 00:01:46.119
varieties the first one I'd like to talk

00:01:43.079 --> 00:01:48.560
about is a very uh classical one called

00:01:46.119 --> 00:01:50.960
wordnet has anyone actually ever used

00:01:48.560 --> 00:01:53.479
wordnet

00:01:50.960 --> 00:01:55.520
before I see at least one person raising

00:01:53.479 --> 00:01:57.640
their hand so it's not entirely uh

00:01:55.520 --> 00:02:00.119
hasn't entirely disappeared has anyone

00:01:57.640 --> 00:02:03.240
heard of wordnet before

00:02:00.119 --> 00:02:05.079
okay more more people um so basically

00:02:03.240 --> 00:02:06.960
this used to be a really big thing in in

00:02:05.079 --> 00:02:10.440
natural language processing it's not So

00:02:06.960 --> 00:02:12.319
Much Anymore um but I I want to explain

00:02:10.440 --> 00:02:14.800
about it because I want to explain why

00:02:12.319 --> 00:02:17.360
this is maybe like less necessary to use

00:02:14.800 --> 00:02:19.599
but actual knowledge bases are still

00:02:17.360 --> 00:02:23.160
more necessary to

00:02:19.599 --> 00:02:26.280
use and so wordnet is a large database

00:02:23.160 --> 00:02:29.560
of words and specifically what it does

00:02:26.280 --> 00:02:32.720
is each word or something they call a

00:02:29.560 --> 00:02:37.120
syn set is a node and then there are

00:02:32.720 --> 00:02:42.560
relationships between nodes and the

00:02:37.120 --> 00:02:44.319
nodes can correspond to nouns um and or

00:02:42.560 --> 00:02:45.920
verbs or

00:02:44.319 --> 00:02:48.360
adjectives

00:02:45.920 --> 00:02:49.959
and nouns have different types of

00:02:48.360 --> 00:02:53.360
relations between them so they have

00:02:49.959 --> 00:02:56.280
things like an is a relation so like a

00:02:53.360 --> 00:03:00.040
hatchback is a type of car they are part

00:02:56.280 --> 00:03:02.840
of relations uh where a wheel is a part

00:03:00.040 --> 00:03:05.720
of a car um and they also make

00:03:02.840 --> 00:03:09.799
distinctions between types and instances

00:03:05.720 --> 00:03:12.400
so like Joe Biden is an instance of a

00:03:09.799 --> 00:03:16.560
president and president is the

00:03:12.400 --> 00:03:19.239
type so um verb relations are ordered by

00:03:16.560 --> 00:03:22.680
specificity so like communicate is more

00:03:19.239 --> 00:03:25.799
broad than talk so talk is you know

00:03:22.680 --> 00:03:27.519
generally a sub class of communicate and

00:03:25.799 --> 00:03:30.720
then whisper is generally a subass of

00:03:27.519 --> 00:03:33.159
talk so it's ordered in this way

00:03:30.720 --> 00:03:35.920
and then adjective relations are mostly

00:03:33.159 --> 00:03:37.720
antonyms so like wet and wet versus dry

00:03:35.920 --> 00:03:43.599
and other things like

00:03:37.720 --> 00:03:47.080
this um when I said sinets uh actually

00:03:43.599 --> 00:03:50.239
the each node is not a word despite the

00:03:47.080 --> 00:03:53.239
name word net it's a set of words that

00:03:50.239 --> 00:03:56.200
all have the same meaning so you might

00:03:53.239 --> 00:03:59.120
have artifact and thing would both

00:03:56.200 --> 00:04:00.879
correspond to this um node because they

00:03:59.120 --> 00:04:02.599
both mean basically the same thing so

00:04:00.879 --> 00:04:04.159
it's like sets of synonyms and this is

00:04:02.599 --> 00:04:07.599
also important when we talk about other

00:04:04.159 --> 00:04:09.920
types of uh knowledge bases as well and

00:04:07.599 --> 00:04:13.920
so what was this used for um this was

00:04:09.920 --> 00:04:17.160
used for for example uh trying to figure

00:04:13.920 --> 00:04:22.400
out whether trying to find all the cars

00:04:17.160 --> 00:04:24.440
that were mentioned in like a in a large

00:04:22.400 --> 00:04:27.440
set of text so you would go through you

00:04:24.440 --> 00:04:30.280
would identify all

00:04:27.440 --> 00:04:32.120
sinets or you would identify all words

00:04:30.280 --> 00:04:34.120
that corresponded to these sunsets and

00:04:32.120 --> 00:04:35.720
then you would take a step up and find

00:04:34.120 --> 00:04:38.800
motor car and you would know that like

00:04:35.720 --> 00:04:42.320
all of those were mentions of cars so

00:04:38.800 --> 00:04:45.520
like why don't we use wordnet very much

00:04:42.320 --> 00:04:45.520
anymore any

00:04:49.160 --> 00:04:52.840
ideas what would what would you do

00:04:51.080 --> 00:04:55.560
instead if I told you find all the cars

00:04:52.840 --> 00:04:55.560
in a big piece of

00:04:55.960 --> 00:05:00.160
text yeah just do something with the

00:04:58.280 --> 00:05:02.880
embeding just do something with

00:05:00.160 --> 00:05:04.560
embeddings yeah so you might get um you

00:05:02.880 --> 00:05:06.720
might get something and find all things

00:05:04.560 --> 00:05:10.360
that were close in embedding space to a

00:05:06.720 --> 00:05:10.360
car what what's another thing you might

00:05:11.560 --> 00:05:15.520
do like what I would do is I would

00:05:13.639 --> 00:05:17.080
download mistol and say does this

00:05:15.520 --> 00:05:19.880
sentence talk about a car and it would

00:05:17.080 --> 00:05:22.199
say yes or no and I I would you know or

00:05:19.880 --> 00:05:23.479
I would say find all the cars in this uh

00:05:22.199 --> 00:05:25.319
that are mentioned in the sentence and

00:05:23.479 --> 00:05:28.720
it would get them and sure that's like

00:05:25.319 --> 00:05:31.319
expensive but it's really easy so um you

00:05:28.720 --> 00:05:32.919
know there are other options that might

00:05:31.319 --> 00:05:36.720
be less expensive but that could solve a

00:05:32.919 --> 00:05:39.520
lot of the things so word not you know

00:05:36.720 --> 00:05:41.039
started out with more and more it it

00:05:39.520 --> 00:05:42.600
started out being very popular in

00:05:41.039 --> 00:05:44.039
natural language processing but now it's

00:05:42.600 --> 00:05:45.440
less so because we can get a lot of it

00:05:44.039 --> 00:05:47.639
from embeddings we can get a lot of it

00:05:45.440 --> 00:05:50.520
from language models

00:05:47.639 --> 00:05:52.759
itself um another thing that started

00:05:50.520 --> 00:05:55.759
maybe before wordnet or even around the

00:05:52.759 --> 00:05:58.840
same time as wordnet was this uh data

00:05:55.759 --> 00:06:00.800
base called psych and it was a manually

00:05:58.840 --> 00:06:04.160
curated database attempting to encode

00:06:00.800 --> 00:06:06.280
all common sense knowledge um and the

00:06:04.160 --> 00:06:08.759
project itself lasted for about 30 to 40

00:06:06.280 --> 00:06:11.840
years it might even still

00:06:08.759 --> 00:06:13.319
exist um and so they had this huge uh

00:06:11.840 --> 00:06:15.199
like hierarchy of all the different

00:06:13.319 --> 00:06:17.680
types of knowledge you could have it

00:06:15.199 --> 00:06:19.680
encoded knowledge about like events and

00:06:17.680 --> 00:06:21.479
like which events happened before other

00:06:19.680 --> 00:06:26.840
events and all these other stuff like

00:06:21.479 --> 00:06:29.039
this um but the problem with this is uh

00:06:26.840 --> 00:06:31.000
this was just too ambitious basically it

00:06:29.039 --> 00:06:35.680
was not possible to encode all of this

00:06:31.000 --> 00:06:37.440
manually by hand so people um like it it

00:06:35.680 --> 00:06:38.840
did it got part of the way there but

00:06:37.440 --> 00:06:40.240
that part of the way there was not

00:06:38.840 --> 00:06:42.560
enough for it to be really useful in

00:06:40.240 --> 00:06:45.199
Practical systems so it isn't this sort

00:06:42.560 --> 00:06:47.800
of method is not used as frequently

00:06:45.199 --> 00:06:51.240
now

00:06:47.800 --> 00:06:56.000
um a a followup one

00:06:51.240 --> 00:06:57.479
um which is it's successor is now uh the

00:06:56.000 --> 00:06:59.879
the most widely used knowledge Bas is

00:06:57.479 --> 00:07:03.240
something called dbpedia and the basic

00:06:59.879 --> 00:07:06.120
idea behind dbpedia is that while Psych

00:07:03.240 --> 00:07:07.840
is too difficult because they had people

00:07:06.120 --> 00:07:12.400
on the psych project who would go in and

00:07:07.840 --> 00:07:12.400
curate rules um for

00:07:13.280 --> 00:07:19.080
machines Wikipedia basically they have a

00:07:17.160 --> 00:07:21.080
very very large number of humans

00:07:19.080 --> 00:07:23.639
curating this structured data about

00:07:21.080 --> 00:07:25.199
entities in the world for humans they're

00:07:23.639 --> 00:07:27.879
creating it for humans because then you

00:07:25.199 --> 00:07:29.599
can put it on a Wikipedia page and you

00:07:27.879 --> 00:07:31.440
can look and see it says cardig melan

00:07:29.599 --> 00:07:34.160
University it has the former names of

00:07:31.440 --> 00:07:36.919
Carnegie melon um it has the motto of

00:07:34.160 --> 00:07:38.759
Carnegie melon the type of entity who it

00:07:36.919 --> 00:07:41.360
was established by and when and other

00:07:38.759 --> 00:07:42.840
stuff like that and because people are

00:07:41.360 --> 00:07:44.280
no longer creating it for machines

00:07:42.840 --> 00:07:46.280
they're creating it for humans people

00:07:44.280 --> 00:07:47.840
are like motivated to do this so like

00:07:46.280 --> 00:07:49.960
lots of people will do it for free so

00:07:47.840 --> 00:07:51.960
you can actually get a reasonably sized

00:07:49.960 --> 00:07:53.639
amount of data from this and actually

00:07:51.960 --> 00:07:55.720
cover you know like most of the entities

00:07:53.639 --> 00:07:57.080
in the world or not most of the entities

00:07:55.720 --> 00:08:00.120
in the world but most of the notable

00:07:57.080 --> 00:08:03.319
entities in uh part of the world that

00:08:00.120 --> 00:08:03.319
have high participation in

00:08:03.479 --> 00:08:09.800
Wikipedia um so now the the thing that a

00:08:08.039 --> 00:08:13.319
lot of people use is something called

00:08:09.800 --> 00:08:14.919
Wiki data this is not this name is a

00:08:13.319 --> 00:08:17.039
little bit of a misnomer because it's

00:08:14.919 --> 00:08:18.960
not actually that closely connected to

00:08:17.039 --> 00:08:20.639
Wikipedia they extract data from

00:08:18.960 --> 00:08:21.720
Wikipedia but they also extract it from

00:08:20.639 --> 00:08:24.400
lots of other

00:08:21.720 --> 00:08:27.520
sources and this is a curated database

00:08:24.400 --> 00:08:30.360
of entities um it's linked it's

00:08:27.520 --> 00:08:33.959
extremely large scale and it's

00:08:30.360 --> 00:08:38.080
multilingual and um this is an example

00:08:33.959 --> 00:08:39.680
of a thing from Richard fean um where

00:08:38.080 --> 00:08:42.680
people can go in and they can actually

00:08:39.680 --> 00:08:45.320
like add information and stuff like that

00:08:42.680 --> 00:08:47.440
um and you know it gives information

00:08:45.320 --> 00:08:50.959
about education and all kinds of other

00:08:47.440 --> 00:08:52.600
stuff so um for fun I can go to the wiki

00:08:50.959 --> 00:08:55.040
data

00:08:52.600 --> 00:08:59.360
site does anyone have an entity they'd

00:08:55.040 --> 00:08:59.360
like to know more about

00:09:01.640 --> 00:09:07.320
any any ideas maybe something that has

00:09:03.959 --> 00:09:07.320
been in the news recently

00:09:10.680 --> 00:09:16.160
or nobody brave enough to come up with

00:09:13.040 --> 00:09:18.360
an entity yeah

00:09:16.160 --> 00:09:20.640
Mamba that's a good one I'm actually not

00:09:18.360 --> 00:09:23.800
sure if that one's going to be in here

00:09:20.640 --> 00:09:27.720
um there's lots of mambas but I don't

00:09:23.800 --> 00:09:27.720
know about that particular Mamba let me

00:09:27.839 --> 00:09:31.200
see do you want to know about a

00:09:29.720 --> 00:09:33.399
different Mamba do you want about know

00:09:31.200 --> 00:09:36.040
about Mamba the research

00:09:33.399 --> 00:09:38.399
group so Mamba is a research group it's

00:09:36.040 --> 00:09:41.800
the modeling and Analysis for medicine

00:09:38.399 --> 00:09:44.800
research group um it focuses on

00:09:41.800 --> 00:09:48.000
mathematical biology and it's in the uh

00:09:44.800 --> 00:09:51.120
in this National Center for scientific

00:09:48.000 --> 00:09:52.519
research in France um the chairperson is

00:09:51.120 --> 00:09:55.360
this person and stuff like that so you

00:09:52.519 --> 00:10:00.200
can see it has all of these things so

00:09:55.360 --> 00:10:03.920
Mamba this Mamba is a node in the graph

00:10:00.200 --> 00:10:06.839
and then the edges are pointing um the

00:10:03.920 --> 00:10:09.440
edges are labeled with like instance of

00:10:06.839 --> 00:10:11.200
and then the next note is research group

00:10:09.440 --> 00:10:13.000
so research group is like another note

00:10:11.200 --> 00:10:17.120
in the graph and so you can click

00:10:13.000 --> 00:10:18.680
through this and it has its own ID and

00:10:17.120 --> 00:10:21.200
other things like

00:10:18.680 --> 00:10:22.839
this also you'll notice that research

00:10:21.200 --> 00:10:24.160
group is translated into lots of

00:10:22.839 --> 00:10:27.440
different languages in the world so you

00:10:24.160 --> 00:10:30.120
can use it multi multilingually and um

00:10:27.440 --> 00:10:33.880
and other things like that

00:10:30.120 --> 00:10:37.000
um even minor entities like Graham

00:10:33.880 --> 00:10:40.160
nuig are included in this and it has a

00:10:37.000 --> 00:10:42.240
little bit of um like information about

00:10:40.160 --> 00:10:45.480
me like my PhD was in Kyoto University

00:10:42.240 --> 00:10:45.480
in 2012 I am a

00:10:45.600 --> 00:10:52.079
human I I am male uh and first name last

00:10:50.519 --> 00:10:53.720
name University teacher computer

00:10:52.079 --> 00:10:56.279
scientist natural language processing

00:10:53.720 --> 00:10:58.639
this is all right um because this is

00:10:56.279 --> 00:11:00.240
mostly hand curated it even has the IDS

00:10:58.639 --> 00:11:04.240
of my advisor

00:11:00.240 --> 00:11:06.519
advisers um the reason why it has all of

00:11:04.240 --> 00:11:09.839
this stuff actually is because like 15

00:11:06.519 --> 00:11:12.160
years ago or like 10 years ago I entered

00:11:09.839 --> 00:11:14.399
in my uh my information into the

00:11:12.160 --> 00:11:16.240
mathematical genealogy project uh which

00:11:14.399 --> 00:11:18.880
is this project about who your advisers

00:11:16.240 --> 00:11:20.680
were because I wanted to see like who my

00:11:18.880 --> 00:11:22.800
mathematical like siblings were and

00:11:20.680 --> 00:11:24.519
stuff like that and uh somehow they

00:11:22.800 --> 00:11:27.360
managed to pull that out and keep this

00:11:24.519 --> 00:11:28.760
like 10 years later so um basically

00:11:27.360 --> 00:11:30.519
they're pulling information from like

00:11:28.760 --> 00:11:32.800
many many different structured data

00:11:30.519 --> 00:11:34.160
sources that they can use so uh they can

00:11:32.800 --> 00:11:37.480
pull it in there I don't know where they

00:11:34.160 --> 00:11:39.440
got that I'm human uh but maybe that was

00:11:37.480 --> 00:11:43.240
inferred from some piece of data

00:11:39.440 --> 00:11:44.760
somewhere online or something cool um

00:11:43.240 --> 00:11:46.839
another good thing about this that

00:11:44.760 --> 00:11:52.680
actually I didn't mention directly in

00:11:46.839 --> 00:11:52.680
the um in the lecture note or

00:11:54.680 --> 00:12:01.120
slides is that there's a query language

00:11:57.360 --> 00:12:04.320
for this yeah and a query language this

00:12:01.120 --> 00:12:06.839
query language is called Sparkle so

00:12:04.320 --> 00:12:10.680
there's a sequel for querying relational

00:12:06.839 --> 00:12:14.399
databases and Sparkle is for querying

00:12:10.680 --> 00:12:15.240
these uh knowledge bases and let me see

00:12:14.399 --> 00:12:18.279
if I

00:12:15.240 --> 00:12:22.560
can I asked chat

00:12:18.279 --> 00:12:24.560
GPT to write me a sparkle query to find

00:12:22.560 --> 00:12:26.839
all presidents of Carnegie melon

00:12:24.560 --> 00:12:31.160
University so let's see if Chad GPT is

00:12:26.839 --> 00:12:31.160
capable of doing that um

00:12:35.639 --> 00:12:39.680
okay that's a problem let me

00:12:41.279 --> 00:12:47.000
see okay there's there's an errand there

00:12:43.880 --> 00:12:48.360
but like if uh uh if I could find a I

00:12:47.000 --> 00:12:50.160
don't want to waste time in class like

00:12:48.360 --> 00:12:52.079
finding a working query but basically

00:12:50.160 --> 00:12:53.399
you can put it in a query and it allows

00:12:52.079 --> 00:12:56.120
you to do a lot of things that are

00:12:53.399 --> 00:13:00.519
similar to what you can do in SQL so you

00:12:56.120 --> 00:13:02.720
can find like all of the edges of nodes

00:13:00.519 --> 00:13:05.279
that satisfy a particular relation so

00:13:02.720 --> 00:13:07.360
you could say I want for Carnegie melon

00:13:05.279 --> 00:13:10.160
University to find all things that

00:13:07.360 --> 00:13:13.519
followed the like president of relation

00:13:10.160 --> 00:13:14.959
and that would give me all um you know

00:13:13.519 --> 00:13:18.680
all presidents of Carnegie melon

00:13:14.959 --> 00:13:20.440
University you can also like filter um

00:13:18.680 --> 00:13:22.160
filter by their start date and end date

00:13:20.440 --> 00:13:24.120
so find all of the preceden between a

00:13:22.160 --> 00:13:25.839
certain time and a another time or

00:13:24.120 --> 00:13:30.480
things like

00:13:25.839 --> 00:13:34.199
that so this is good if you want to get

00:13:30.480 --> 00:13:36.600
like high reli high reliability data um

00:13:34.199 --> 00:13:39.839
in a scalable way because like if I ask

00:13:36.600 --> 00:13:41.920
chat GPT like one of my favorite um one

00:13:39.839 --> 00:13:45.720
of my favorite queries for chat GPT is

00:13:41.920 --> 00:13:48.600
like name all of the name all of the

00:13:45.720 --> 00:13:51.959
presidents that were born uh east of the

00:13:48.600 --> 00:13:53.880
Mississippi River um and I've never

00:13:51.959 --> 00:13:56.519
successfully gotten chat GPT to be able

00:13:53.880 --> 00:13:57.800
to do this um because there's lots of

00:13:56.519 --> 00:13:59.560
presidents who were born east of the

00:13:57.800 --> 00:14:02.320
Mississippi River and it starts counting

00:13:59.560 --> 00:14:04.079
them it can't distinguish what position

00:14:02.320 --> 00:14:05.639
is east of the Mississippi and what

00:14:04.079 --> 00:14:09.120
position is the west west of the

00:14:05.639 --> 00:14:11.279
Mississippi but if you write a uh like a

00:14:09.120 --> 00:14:14.759
sparkle query it's not that hard to do

00:14:11.279 --> 00:14:16.480
that so there are um you know there are

00:14:14.759 --> 00:14:18.639
certain types of questions especially

00:14:16.480 --> 00:14:20.399
information aggregation and complex

00:14:18.639 --> 00:14:22.839
relations and stuff that uh language

00:14:20.399 --> 00:14:26.600
models are not very good

00:14:22.839 --> 00:14:28.120
at cool um so that's kind of an intro to

00:14:26.600 --> 00:14:31.240
knowledge bases why you might want to

00:14:28.120 --> 00:14:33.759
think about them any questions so far

00:14:31.240 --> 00:14:33.759
for

00:14:34.759 --> 00:14:39.720
discussion okay um I will move on next

00:14:38.320 --> 00:14:41.199
so the next thing I'd like to talk about

00:14:39.720 --> 00:14:43.839
is learning representations for

00:14:41.199 --> 00:14:45.519
knowledge bases um so knowledge bases

00:14:43.839 --> 00:14:48.000
are great but one problem is they're

00:14:45.519 --> 00:14:51.040
like inherently

00:14:48.000 --> 00:14:55.040
incomplete and even with extremely large

00:14:51.040 --> 00:14:58.279
scale uh it becomes impossible to have

00:14:55.040 --> 00:15:00.360
them be complete and the reason why is

00:14:58.279 --> 00:15:03.639
uh for examp example in Freebase which

00:15:00.360 --> 00:15:05.480
was the predecessor to Wiki data um 71%

00:15:03.639 --> 00:15:08.560
of humans didn't have a date of

00:15:05.480 --> 00:15:10.560
birth um and probably every human

00:15:08.560 --> 00:15:12.079
actually has a date of birth right um

00:15:10.560 --> 00:15:15.880
you know we're pretty much guaranteed

00:15:12.079 --> 00:15:17.639
for that to be the case so the issue is

00:15:15.880 --> 00:15:19.160
like for very famous entities you want

00:15:17.639 --> 00:15:21.040
lots of detailed information like you

00:15:19.160 --> 00:15:24.000
can know absolutely everything about Joe

00:15:21.040 --> 00:15:25.759
Biden or Barack Obama but you know at

00:15:24.000 --> 00:15:26.880
the same time for Less major entities

00:15:25.759 --> 00:15:28.079
you still want them in the knowledge

00:15:26.880 --> 00:15:30.079
base but you're not going to be able to

00:15:28.079 --> 00:15:31.519
get all that information or should you

00:15:30.079 --> 00:15:35.600
for privacy

00:15:31.519 --> 00:15:36.680
purposes and so the idea is um for

00:15:35.600 --> 00:15:38.079
information that's written on the

00:15:36.680 --> 00:15:40.600
internet somewhere can you perform

00:15:38.079 --> 00:15:42.759
relation extraction which essentially

00:15:40.600 --> 00:15:44.600
allows you to extract this information

00:15:42.759 --> 00:15:46.360
and create your own knowledge bases and

00:15:44.600 --> 00:15:47.680
stuff like this and this can also be

00:15:46.360 --> 00:15:50.079
useful if you want to create it for like

00:15:47.680 --> 00:15:52.199
a specialized domain or um or other

00:15:50.079 --> 00:15:55.000
stuff like

00:15:52.199 --> 00:15:59.519
that so there's a bunch of ways that

00:15:55.000 --> 00:16:03.079
people do this um and one kind of

00:15:59.519 --> 00:16:06.120
popular way that people have tried to do

00:16:03.079 --> 00:16:09.199
relation extraction is through uh

00:16:06.120 --> 00:16:12.560
leveraging consistency in embedding

00:16:09.199 --> 00:16:15.319
space and so this is the most famous

00:16:12.560 --> 00:16:17.959
example from word de uh what seems like

00:16:15.319 --> 00:16:21.880
ages ago uh in

00:16:17.959 --> 00:16:23.920
2013 and in the word Toc paper one of

00:16:21.880 --> 00:16:26.279
the big you know exciting things was

00:16:23.920 --> 00:16:28.639
essentially they demonstrated that

00:16:26.279 --> 00:16:30.120
vectors in embedding space had kind of

00:16:28.639 --> 00:16:31.839
in

00:16:30.120 --> 00:16:33.160
you know meaning and actually the

00:16:31.839 --> 00:16:34.600
vectors in embedding space could

00:16:33.160 --> 00:16:37.639
correspond to relations between

00:16:34.600 --> 00:16:39.480
embeddings so like uh we would have man

00:16:37.639 --> 00:16:41.000
pointing to woman in approximately the

00:16:39.480 --> 00:16:42.920
same direction that we had Uncle

00:16:41.000 --> 00:16:46.600
pointing to Aunt and King pointing to

00:16:42.920 --> 00:16:49.680
Queen and so um then you could do things

00:16:46.600 --> 00:16:51.440
like you could take Kings subtract out

00:16:49.680 --> 00:16:53.560
the vector that corresponded to

00:16:51.440 --> 00:16:58.360
plurality uh add the vector that

00:16:53.560 --> 00:17:00.839
corresponded to um you know uh to going

00:16:58.360 --> 00:17:04.319
from masculine to feminine words and

00:17:00.839 --> 00:17:05.559
then um like read the vector to that

00:17:04.319 --> 00:17:07.160
were plural and you'd be able to

00:17:05.559 --> 00:17:09.439
identify the plural by just knowing

00:17:07.160 --> 00:17:11.000
these two uh vectors the plural of green

00:17:09.439 --> 00:17:14.000
by just knowing those two

00:17:11.000 --> 00:17:14.000
vectors

00:17:14.160 --> 00:17:21.880
um but it turns out that you can either

00:17:18.199 --> 00:17:21.880
learn embeddings

00:17:22.720 --> 00:17:28.240
from like uh you can either learn

00:17:25.000 --> 00:17:30.400
embeddings from text or you can use the

00:17:28.240 --> 00:17:32.039
fact that you have a big knowledge base

00:17:30.400 --> 00:17:34.880
that was curated by humans like Wiki

00:17:32.039 --> 00:17:36.120
data to improve the embeddings of a

00:17:34.880 --> 00:17:39.559
neural model

00:17:36.120 --> 00:17:41.799
itself and so another pretty large uh

00:17:39.559 --> 00:17:43.600
research area that a lot of people have

00:17:41.799 --> 00:17:47.120
focused on is how do you get good

00:17:43.600 --> 00:17:48.720
embeddings of a Knowledge Graph and this

00:17:47.120 --> 00:17:50.600
is important if you want to do any sort

00:17:48.720 --> 00:17:52.799
of like Knowledge Graph Search or other

00:17:50.600 --> 00:17:54.160
things like this like for example one of

00:17:52.799 --> 00:17:56.799
the really nice things about knowledge

00:17:54.160 --> 00:17:58.880
graphs is they have information about a

00:17:56.799 --> 00:18:00.200
whole bunch of really sparse entities

00:17:58.880 --> 00:18:03.240
that aren't mentioned very much on the

00:18:00.200 --> 00:18:05.679
internet for example and so because of

00:18:03.240 --> 00:18:07.440
that you can um you can leverage the

00:18:05.679 --> 00:18:10.720
knowledge graph structure together with

00:18:07.440 --> 00:18:10.720
text to learn better embeddings

00:18:11.240 --> 00:18:18.520
overall and so this particular paper is

00:18:15.280 --> 00:18:20.960
one example of it um and the way they do

00:18:18.520 --> 00:18:23.280
this is they express uh Knowledge Graph

00:18:20.960 --> 00:18:25.919
triples is additive

00:18:23.280 --> 00:18:28.480
Transformations and they minimize the

00:18:25.919 --> 00:18:31.640
distance uh of existing triples with a

00:18:28.480 --> 00:18:35.039
margin based loss so the way they do

00:18:31.640 --> 00:18:38.240
this is they have the head um in the

00:18:35.039 --> 00:18:40.799
tail and L is the vector corresponding

00:18:38.240 --> 00:18:42.679
to like the link between the things that

00:18:40.799 --> 00:18:47.960
corresponds to a

00:18:42.679 --> 00:18:52.159
relation and so you go uh you have H and

00:18:47.960 --> 00:18:53.559
T and here um like this is L but here

00:18:52.159 --> 00:18:55.640
it's written as are because I got this

00:18:53.559 --> 00:18:58.120
from a different paper and basically you

00:18:55.640 --> 00:18:59.480
you try to go from H to T um according

00:18:58.120 --> 00:19:00.919
to the relation

00:18:59.480 --> 00:19:05.120
uh Vector

00:19:00.919 --> 00:19:07.200
are and you use a hinge loss where um

00:19:05.120 --> 00:19:10.039
for the hinge loss you you have a hinge

00:19:07.200 --> 00:19:12.640
parameter and then you try to upweight

00:19:10.039 --> 00:19:15.760
the example of a true triple and

00:19:12.640 --> 00:19:17.960
downweight the example of a of a false

00:19:15.760 --> 00:19:19.880
triple so this could be one that was

00:19:17.960 --> 00:19:22.080
like randomly sampled to be incorrect

00:19:19.880 --> 00:19:22.080
for

00:19:23.760 --> 00:19:29.080
example um one interesting thing about

00:19:26.880 --> 00:19:31.559
knowledge graph embeddings is like a lot

00:19:29.080 --> 00:19:33.600
of famous AI researchers got their start

00:19:31.559 --> 00:19:36.000
in Knowledge Graph embeddings and so

00:19:33.600 --> 00:19:39.760
Richard soer is one of them if you know

00:19:36.000 --> 00:19:44.320
he's the CEO of vi.com search engine now

00:19:39.760 --> 00:19:46.679
um and uh this was a first attempt at

00:19:44.320 --> 00:19:49.679
predicting relations they basically

00:19:46.679 --> 00:19:55.400
created a um MLP that tries to predict

00:19:49.679 --> 00:19:58.880
whether a relation exists so they have

00:19:55.400 --> 00:20:00.760
a matrix for the left side of the

00:19:58.880 --> 00:20:03.320
relation a matrix for the right side of

00:20:00.760 --> 00:20:05.080
the relation and then they feed in the

00:20:03.320 --> 00:20:07.559
embeddings of each of the entities in

00:20:05.080 --> 00:20:08.919
the relation they have a nonlinearity

00:20:07.559 --> 00:20:11.799
and then they have another Vector that

00:20:08.919 --> 00:20:14.720
tries to predict the um the probability

00:20:11.799 --> 00:20:16.679
of the uh actual relation being correct

00:20:14.720 --> 00:20:18.960
so you would run this through a sigmoid

00:20:16.679 --> 00:20:21.000
and then uh if it was one the relation

00:20:18.960 --> 00:20:24.039
was likely to exist if it was Zero then

00:20:21.000 --> 00:20:25.480
the relation was likely to not exist and

00:20:24.039 --> 00:20:27.799
then they also propos something called a

00:20:25.480 --> 00:20:31.480
neural tensor Network and this adds a

00:20:27.799 --> 00:20:34.000
bilinear feature extractor um and so

00:20:31.480 --> 00:20:37.440
basically what this is saying is we have

00:20:34.000 --> 00:20:40.000
the embedding here the embedding here we

00:20:37.440 --> 00:20:41.840
have a matrix and then we calculate the

00:20:40.000 --> 00:20:43.080
dot product between the embedding after

00:20:41.840 --> 00:20:45.799
transformation it looks a lot like

00:20:43.080 --> 00:20:47.720
attention actually in a way um because

00:20:45.799 --> 00:20:50.000
we had the bilinear attention so it's

00:20:47.720 --> 00:20:53.640
similar to that as well and then we also

00:20:50.000 --> 00:20:56.840
have the MLP so this part corresponds to

00:20:53.640 --> 00:21:00.320
MLP and then we have a bias

00:20:56.840 --> 00:21:02.200
term and um this is a powerful model but

00:21:00.320 --> 00:21:05.400
it's a bit overparameterized so we

00:21:02.200 --> 00:21:08.120
actually later um uh this kind of fell

00:21:05.400 --> 00:21:10.360
out of uh favor towards these more

00:21:08.120 --> 00:21:14.520
simple models that we're using uh kind

00:21:10.360 --> 00:21:14.520
of just linear projections between the

00:21:17.600 --> 00:21:22.279
two so there's um there's a lot of

00:21:20.120 --> 00:21:25.320
methods like this these methods are

00:21:22.279 --> 00:21:27.039
basically assuming that we have either

00:21:25.320 --> 00:21:29.080
Knowledge Graph

00:21:27.039 --> 00:21:30.799
embeddings um and we want to learn

00:21:29.080 --> 00:21:32.480
relations or they're assuming that we

00:21:30.799 --> 00:21:34.320
don't have any information at all about

00:21:32.480 --> 00:21:36.840
the knowledge graph and we want to learn

00:21:34.320 --> 00:21:40.039
the knowledge graph embedding themselves

00:21:36.840 --> 00:21:42.400
it's been used for both of them but um I

00:21:40.039 --> 00:21:44.000
I'd say now it's probably most useful

00:21:42.400 --> 00:21:45.520
for learning Knowledge Graph embeddings

00:21:44.000 --> 00:21:50.480
if you want to do any sort of Knowledge

00:21:45.520 --> 00:21:50.480
Graph based modeling uh which can be

00:21:51.240 --> 00:21:55.919
useful um cool any questions about these

00:21:57.360 --> 00:22:01.679
ones okay

00:21:59.520 --> 00:22:04.360
next um actually this part might be a

00:22:01.679 --> 00:22:06.600
little bit simpler than the uh than the

00:22:04.360 --> 00:22:09.000
like knowledge graft based approaches so

00:22:06.600 --> 00:22:10.960
another method for relations extraction

00:22:09.000 --> 00:22:13.440
is learning from text

00:22:10.960 --> 00:22:16.120
directly

00:22:13.440 --> 00:22:19.080
and the first question about this is how

00:22:16.120 --> 00:22:22.200
do you get training data to learn uh

00:22:19.080 --> 00:22:24.480
about relation learn relation extraction

00:22:22.200 --> 00:22:26.720
and so there was this very influential

00:22:24.480 --> 00:22:28.279
paper a distant supervision for relation

00:22:26.720 --> 00:22:31.120
extraction I would say it's almost one

00:22:28.279 --> 00:22:32.880
of the first or certainly one of the

00:22:31.120 --> 00:22:34.559
most influential papers on like data

00:22:32.880 --> 00:22:35.960
augmentation or synthetic data for

00:22:34.559 --> 00:22:38.400
natural language

00:22:35.960 --> 00:22:40.440
processing and basically the idea is you

00:22:38.400 --> 00:22:44.279
already have a knowledge base that has

00:22:40.440 --> 00:22:47.440
some entries in it like Wiki data and so

00:22:44.279 --> 00:22:50.919
then given in entity relation entity

00:22:47.440 --> 00:22:52.919
triples um can you extract all text that

00:22:50.919 --> 00:22:54.799
matches this particular relation type

00:22:52.919 --> 00:22:56.480
and use it to train a relation extractor

00:22:54.799 --> 00:22:59.640
a supervised relation

00:22:56.480 --> 00:23:01.880
extractor so the way this works

00:22:59.640 --> 00:23:04.039
is like let's say we have this is an old

00:23:01.880 --> 00:23:06.120
paper so the examples are also old but

00:23:04.039 --> 00:23:08.039
um let's say we have Steven Spielberg

00:23:06.120 --> 00:23:10.159
being a director of the film Saving

00:23:08.039 --> 00:23:12.840
Private Ryan and that's included in our

00:23:10.159 --> 00:23:14.840
uh our knowledge base so what it would

00:23:12.840 --> 00:23:17.080
do is it would find all sentences that

00:23:14.840 --> 00:23:19.400
have Steven Spielberg and Saving Private

00:23:17.080 --> 00:23:22.080
Ryan included in them and it would label

00:23:19.400 --> 00:23:24.159
this as like a positive example of that

00:23:22.080 --> 00:23:28.240
relation so this

00:23:24.159 --> 00:23:30.760
is in general often it's okay it it

00:23:28.240 --> 00:23:34.480
works reasonably well but the problem

00:23:30.760 --> 00:23:37.200
with this is there are also um negative

00:23:34.480 --> 00:23:38.840
examples of this so like for example

00:23:37.200 --> 00:23:40.480
here I think the first one is kind of a

00:23:38.840 --> 00:23:43.240
negative example for the director

00:23:40.480 --> 00:23:45.880
relation because Steven Spielberg's film

00:23:43.240 --> 00:23:48.120
Saving Private Ryan doesn't actually

00:23:45.880 --> 00:23:50.000
tell you he's the director it just tells

00:23:48.120 --> 00:23:52.520
you that he's somehow affiliated with it

00:23:50.000 --> 00:23:54.840
he could be the writer or he could be uh

00:23:52.520 --> 00:23:57.679
the actor or or something else like that

00:23:54.840 --> 00:24:00.440
so this is a nice way to create data for

00:23:57.679 --> 00:24:03.640
basically free but at the same time uh

00:24:00.440 --> 00:24:06.159
you can like create noisy examples and

00:24:03.640 --> 00:24:06.159
that can be a

00:24:07.159 --> 00:24:14.600
problem so um there's been a lot of work

00:24:11.400 --> 00:24:16.000
about this um relationship uh relation

00:24:14.600 --> 00:24:17.840
classification with neural networks

00:24:16.000 --> 00:24:20.840
there's a lot of uh different methods

00:24:17.840 --> 00:24:23.159
that could be uh doing this most of them

00:24:20.840 --> 00:24:24.919
work by extracting features and then

00:24:23.159 --> 00:24:27.039
classifying somehow although there are

00:24:24.919 --> 00:24:29.960
some uh large language model based

00:24:27.039 --> 00:24:33.120
methods now um one one thing about

00:24:29.960 --> 00:24:35.440
relation extraction or not kind of like

00:24:33.120 --> 00:24:36.799
information extraction in general is

00:24:35.440 --> 00:24:38.559
that very often you want to run this

00:24:36.799 --> 00:24:40.200
over like a huge Corpus you want to run

00:24:38.559 --> 00:24:42.320
it over the whole internet or other

00:24:40.200 --> 00:24:45.000
things like that so from that point of

00:24:42.320 --> 00:24:47.159
view like I I said I could just ask

00:24:45.000 --> 00:24:49.480
mistol to give me the answer about like

00:24:47.159 --> 00:24:52.440
whether cars are included in sentences

00:24:49.480 --> 00:24:55.120
but if you want to run you know gp4 over

00:24:52.440 --> 00:24:56.799
the whole internet that's a pretty big

00:24:55.120 --> 00:25:00.159
budget and you might want to reconsider

00:24:56.799 --> 00:25:02.440
that so there are so um there is also

00:25:00.159 --> 00:25:04.440
some you know benefit in having cheap

00:25:02.440 --> 00:25:07.200
and lightweight

00:25:04.440 --> 00:25:09.159
methods so basically what this

00:25:07.200 --> 00:25:11.279
particular paper did is it extracted

00:25:09.159 --> 00:25:12.760
features in in classified so it

00:25:11.279 --> 00:25:15.600
extracted lexical features of the

00:25:12.760 --> 00:25:20.240
entities themselves and features of the

00:25:15.600 --> 00:25:22.360
whole span and so like the way I uh most

00:25:20.240 --> 00:25:26.960
modern methods for this do this is they

00:25:22.360 --> 00:25:29.399
basically um extract features from the

00:25:26.960 --> 00:25:31.679
first part of the first entity the

00:25:29.399 --> 00:25:33.760
second part of the the first entity the

00:25:31.679 --> 00:25:36.360
first part of the second entity and the

00:25:33.760 --> 00:25:37.720
last part of the uh second entity and

00:25:36.360 --> 00:25:39.600
take all of those embeddings feed them

00:25:37.720 --> 00:25:41.440
into like an MLP or something like that

00:25:39.600 --> 00:25:44.039
and then make a prediction about whether

00:25:41.440 --> 00:25:45.760
that relation exists so if you have an

00:25:44.039 --> 00:25:47.840
embedding model this is relatively easy

00:25:45.760 --> 00:25:50.360
to do you feed it through like uh

00:25:47.840 --> 00:25:51.919
Roberta or you feed it through mistol

00:25:50.360 --> 00:25:54.559
and get the embeddings for each of the

00:25:51.919 --> 00:25:55.840
tokens and um and then you make a

00:25:54.559 --> 00:25:58.840
prediction based on those four

00:25:55.840 --> 00:25:58.840
embeddings

00:26:00.600 --> 00:26:04.840
um the details of that are like not

00:26:03.520 --> 00:26:07.320
super important unless you're going to

00:26:04.840 --> 00:26:09.279
go in and implement it yourself so you

00:26:07.320 --> 00:26:10.919
can um like if you're actually going to

00:26:09.279 --> 00:26:12.120
be doing relation extraction obviously

00:26:10.919 --> 00:26:14.279
the details are important but I'm

00:26:12.120 --> 00:26:16.000
assuming that most people won't be uh

00:26:14.279 --> 00:26:19.720
you know doing that as your final

00:26:16.000 --> 00:26:21.240
project but um one really interesting

00:26:19.720 --> 00:26:22.919
thing that is relevant even if you're

00:26:21.240 --> 00:26:26.360
not doing relationship relation

00:26:22.919 --> 00:26:29.360
extraction is how you can model noise

00:26:26.360 --> 00:26:32.600
because this um as I said they're

00:26:29.360 --> 00:26:35.720
creating lots of like semi noisy data

00:26:32.600 --> 00:26:38.919
and a lot of the work in getting good

00:26:35.720 --> 00:26:40.360
bottles for relation extraction has been

00:26:38.919 --> 00:26:41.799
how do we deal with this distant

00:26:40.360 --> 00:26:43.799
supervision noise and I'm just going to

00:26:41.799 --> 00:26:45.760
give one example here but there's like a

00:26:43.799 --> 00:26:49.120
series of papers after this that also

00:26:45.760 --> 00:26:50.600
tried to do similar things so the idea

00:26:49.120 --> 00:26:53.600
is that there's noise in the distant

00:26:50.600 --> 00:26:56.559
supervision labels um and so we want to

00:26:53.600 --> 00:27:01.039
model and mitigate that noise and the

00:26:56.559 --> 00:27:03.919
way this paper does this is they have an

00:27:01.039 --> 00:27:06.679
encoder and from the encoder you

00:27:03.919 --> 00:27:10.960
calculate embeddings and make

00:27:06.679 --> 00:27:14.279
predictions and so you have a small set

00:27:10.960 --> 00:27:16.080
of like very high quality data and this

00:27:14.279 --> 00:27:17.760
small set of very high quality data you

00:27:16.080 --> 00:27:19.880
can basically trust that all of the data

00:27:17.760 --> 00:27:22.320
is not noisy like maybe it's manually

00:27:19.880 --> 00:27:23.720
annotated data and you have like 5,000

00:27:22.320 --> 00:27:25.000
examples of it or something like that

00:27:23.720 --> 00:27:26.880
and then separately from that you have

00:27:25.000 --> 00:27:28.440
like 5 million examples of automatically

00:27:26.880 --> 00:27:30.799
labeled data that might be good might

00:27:28.440 --> 00:27:32.679
not be good and so what they do is

00:27:30.799 --> 00:27:34.200
essentially at the beginning they take

00:27:32.679 --> 00:27:36.520
this encoder get embeddings make

00:27:34.200 --> 00:27:38.000
predictions over the high quality data

00:27:36.520 --> 00:27:40.320
and then they have a separate noise

00:27:38.000 --> 00:27:43.440
modeling layer where what this noise

00:27:40.320 --> 00:27:46.919
modeling layer does is it has a

00:27:43.440 --> 00:27:50.039
transition Matrix which says given that

00:27:46.919 --> 00:27:53.279
this given that we made a particular

00:27:50.039 --> 00:27:55.159
prediction over classes because this is

00:27:53.279 --> 00:27:59.919
essentially a multiclass classification

00:27:55.159 --> 00:28:01.519
problem they transform the

00:27:59.919 --> 00:28:03.159
sorry I don't remember if they transform

00:28:01.519 --> 00:28:04.640
the probabilities or the low Jets I

00:28:03.159 --> 00:28:07.320
think it's the probabilities but they

00:28:04.640 --> 00:28:12.799
transform the probabilities and get a

00:28:07.320 --> 00:28:14.720
final uh distribution after noise and so

00:28:12.799 --> 00:28:17.399
that means that you can basically smooth

00:28:14.720 --> 00:28:19.240
out this uh distribution and account for

00:28:17.399 --> 00:28:20.880
the fact that the labels may be noisy or

00:28:19.240 --> 00:28:24.399
may may not be

00:28:20.880 --> 00:28:26.600
noisy um then they add additional

00:28:24.399 --> 00:28:28.559
normalization on this transition Matrix

00:28:26.600 --> 00:28:32.440
using something called Trace normal

00:28:28.559 --> 00:28:35.840
ization to move this Matrix closer to

00:28:32.440 --> 00:28:38.480
the identity function which says that

00:28:35.840 --> 00:28:40.720
the predictions are probably not wrong

00:28:38.480 --> 00:28:43.159
all the time uh the predictions are

00:28:40.720 --> 00:28:45.360
probably correct you know a lot of the

00:28:43.159 --> 00:28:46.600
time they're not correct all the time uh

00:28:45.360 --> 00:28:49.720
so then you have that Trace

00:28:46.600 --> 00:28:51.880
normalization competing with um this uh

00:28:49.720 --> 00:28:55.440
trying to give you like a more smooth

00:28:51.880 --> 00:28:58.760
distribution and and reduce your uh L

00:28:55.440 --> 00:29:00.320
like reduce your loss so um I I think

00:28:58.760 --> 00:29:02.559
this is actually a pretty interesting

00:29:00.320 --> 00:29:04.480
idea and it can be used not just for

00:29:02.559 --> 00:29:08.600
relation extraction but also in cases

00:29:04.480 --> 00:29:08.600
where um you might have noisy labels

00:29:08.799 --> 00:29:14.320
overall um so are there any questions

00:29:12.360 --> 00:29:15.720
about this or any of the things that are

00:29:14.320 --> 00:29:18.480
going on

00:29:15.720 --> 00:29:20.279
here um even if you're completely

00:29:18.480 --> 00:29:21.960
uninterested in relation extraction I'd

00:29:20.279 --> 00:29:23.720
encourage you to think about like what

00:29:21.960 --> 00:29:26.159
are

00:29:23.720 --> 00:29:27.360
some examples of things that you are

00:29:26.159 --> 00:29:29.519
interested in where you could get

00:29:27.360 --> 00:29:31.840
potentially labels and how could you for

00:29:29.519 --> 00:29:34.880
theise there like that might be uh you

00:29:31.840 --> 00:29:34.880
know a thing to

00:29:35.679 --> 00:29:39.919
about okay so this was a very very brief

00:29:38.320 --> 00:29:42.679
overview of how we create knowledge

00:29:39.919 --> 00:29:44.080
bases uh from textual data or from

00:29:42.679 --> 00:29:47.159
Knowledge Graph data structured

00:29:44.080 --> 00:29:48.840
Knowledge Graph data um so now I like to

00:29:47.159 --> 00:29:51.519
talk a little bit about how to use

00:29:48.840 --> 00:29:53.960
knowledge bases to inform neural

00:29:51.519 --> 00:29:56.159
models and there's a bunch of different

00:29:53.960 --> 00:29:59.519
ways to do this

00:29:56.159 --> 00:30:02.600
um the

00:29:59.519 --> 00:30:06.960
the first way um is to

00:30:02.600 --> 00:30:09.840
improve embeddings uh

00:30:06.960 --> 00:30:11.960
with existing lexicons and this example

00:30:09.840 --> 00:30:14.679
is using non-contextual embeddings like

00:30:11.960 --> 00:30:16.240
not the not the ones we get from neural

00:30:14.679 --> 00:30:17.919
language models but once we get from

00:30:16.240 --> 00:30:20.919
just running a embedding model like word

00:30:17.919 --> 00:30:22.960
toac or something like this um and what

00:30:20.919 --> 00:30:25.640
they did in this paper is they

00:30:22.960 --> 00:30:27.600
essentially um retrofitted embeddings to

00:30:25.640 --> 00:30:30.840
existing lexicons by doing post Hawk

00:30:27.600 --> 00:30:34.080
trans of the embeddings so that they

00:30:30.840 --> 00:30:36.840
matched the um the knowledge graph for

00:30:34.080 --> 00:30:39.080
lexon better and so the way they did

00:30:36.840 --> 00:30:41.880
this is

00:30:39.080 --> 00:30:43.720
um they started out with pre-trained

00:30:41.880 --> 00:30:45.399
embeddings and they had a double

00:30:43.720 --> 00:30:47.240
objective of making the transform

00:30:45.399 --> 00:30:49.120
embeddings close to the neighbors and

00:30:47.240 --> 00:30:52.519
close to the original

00:30:49.120 --> 00:30:58.840
embedding and the way they did this is

00:30:52.519 --> 00:30:58.840
they essentially had um this

00:30:59.799 --> 00:31:03.720
this regularization term over here so

00:31:01.880 --> 00:31:06.200
this regularization term is basically

00:31:03.720 --> 00:31:08.279
saying um I don't want you to move your

00:31:06.200 --> 00:31:09.360
embeddings too far away from how they

00:31:08.279 --> 00:31:11.679
were

00:31:09.360 --> 00:31:14.799
initialized and then at the same time I

00:31:11.679 --> 00:31:17.279
would like you to make these uh

00:31:14.799 --> 00:31:19.600
embeddings closer to each other if they

00:31:17.279 --> 00:31:21.240
are synonyms of each other so they did

00:31:19.600 --> 00:31:23.600
this using word net and they basically

00:31:21.240 --> 00:31:26.200
took the words uh that were synonyms to

00:31:23.600 --> 00:31:28.679
each other in sinets with each other and

00:31:26.200 --> 00:31:30.000
they tried to regularize the synonyms to

00:31:28.679 --> 00:31:32.120
be closer together but also the

00:31:30.000 --> 00:31:33.639
embeddings to be closer to how they

00:31:32.120 --> 00:31:35.960
started

00:31:33.639 --> 00:31:38.799
out and there were also examples of

00:31:35.960 --> 00:31:40.720
forcing anms away from each other so

00:31:38.799 --> 00:31:42.480
like if you're um this is a little bit

00:31:40.720 --> 00:31:44.799
of an older work so it was working on

00:31:42.480 --> 00:31:47.600
non-contextualized embeddings but we

00:31:44.799 --> 00:31:49.399
could do something very similar for um

00:31:47.600 --> 00:31:52.000
more modern models in like Knowledge

00:31:49.399 --> 00:31:55.320
Graph embeddings for example so let's

00:31:52.000 --> 00:31:58.960
say we had

00:31:55.320 --> 00:32:03.240
um a model that ident

00:31:58.960 --> 00:32:06.600
entities and then different examples of

00:32:03.240 --> 00:32:06.600
those entities across different

00:32:07.159 --> 00:32:11.480
contexts um let's go back to the wiki

00:32:20.639 --> 00:32:26.840
data and so um if we had lots of

00:32:23.960 --> 00:32:29.360
examples of Joe Biden um Joe Biden is

00:32:26.840 --> 00:32:35.159
referred to in a number ways like Joe

00:32:29.360 --> 00:32:44.440
Biden Joseph Biden Joseph R Biden um J

00:32:35.159 --> 00:32:47.880
jrb I guess um pus 48 46 sorry um and uh

00:32:44.440 --> 00:32:50.799
so you could find different examples of

00:32:47.880 --> 00:32:52.799
things that match these strings um and

00:32:50.799 --> 00:32:55.360
even do entity linking uh which I'll

00:32:52.799 --> 00:32:57.200
I'll talk about in a little bit and then

00:32:55.360 --> 00:32:58.760
encourag the embeddings for all of these

00:32:57.200 --> 00:33:01.360
different instances is to be closer

00:32:58.760 --> 00:33:04.039
together to make your model like disting

00:33:01.360 --> 00:33:06.799
uh distinguish them less and Ure that

00:33:04.039 --> 00:33:08.399
they uh they get closer edings and that

00:33:06.799 --> 00:33:11.639
could improve like question answering

00:33:08.399 --> 00:33:11.639
look up other stuff like

00:33:12.960 --> 00:33:19.880
that

00:33:14.919 --> 00:33:23.399
cool um yeah I have a question about

00:33:19.880 --> 00:33:25.399
this so what happens if you do like subw

00:33:23.399 --> 00:33:28.000
modeling and then you don't have like

00:33:25.399 --> 00:33:30.440
the embedment for that entire string

00:33:28.000 --> 00:33:32.320
that is supposed to be Clos yeah what

00:33:30.440 --> 00:33:34.279
happens if you do subword modeling and

00:33:32.320 --> 00:33:35.480
you don't have the embedding uh you

00:33:34.279 --> 00:33:37.159
don't have a single embedding that

00:33:35.480 --> 00:33:40.360
corresponds to an entity so that's a

00:33:37.159 --> 00:33:42.559
really good question um let me

00:33:40.360 --> 00:33:44.240
check I don't think I actually have

00:33:42.559 --> 00:33:46.600
these on the slide so I might have to

00:33:44.240 --> 00:33:46.600
open a

00:33:53.639 --> 00:33:59.720
paper yeah okay so there's a lot of

00:33:56.440 --> 00:33:59.720
different ways to handle this

00:34:11.520 --> 00:34:18.079
so there there's two papers um the first

00:34:14.879 --> 00:34:20.000
paper is uh a really nice paper very

00:34:18.079 --> 00:34:22.359
influential on the subject of

00:34:20.000 --> 00:34:25.359
co-reference resolution and co-reference

00:34:22.359 --> 00:34:27.240
resolution um is essentially trying to

00:34:25.359 --> 00:34:30.000
identify when two spans correspond to

00:34:27.240 --> 00:34:32.320
each other so like if I say Joe B Joe

00:34:30.000 --> 00:34:34.359
Biden early in a document and then later

00:34:32.320 --> 00:34:35.480
in a document it just says Biden we want

00:34:34.359 --> 00:34:38.839
to know that those two things are

00:34:35.480 --> 00:34:40.919
referring to each other and then um we

00:34:38.839 --> 00:34:42.839
had a paper later where we generalized

00:34:40.919 --> 00:34:44.839
this and applied you know very similar

00:34:42.839 --> 00:34:48.079
methodology to like lots and lots of

00:34:44.839 --> 00:34:50.760
different analysis tasks but I can um I

00:34:48.079 --> 00:34:53.839
can show the beginning here and

00:34:50.760 --> 00:34:59.320
basically the methodology that they use

00:34:53.839 --> 00:35:02.440
here um is they add

00:34:59.320 --> 00:35:04.440
a and this is specifically for modeling

00:35:02.440 --> 00:35:08.240
spans and getting embeddings out of

00:35:04.440 --> 00:35:09.040
spans of uh tokens and what they did is

00:35:08.240 --> 00:35:13.079
they

00:35:09.040 --> 00:35:14.920
essentially have a model where you take

00:35:13.079 --> 00:35:16.440
the thing from the beginning the

00:35:14.920 --> 00:35:18.760
embedding from the beginning of the span

00:35:16.440 --> 00:35:22.040
the embedding from the end of the span

00:35:18.760 --> 00:35:24.280
and the average embedding of all of the

00:35:22.040 --> 00:35:26.280
embeddings in the span and that gives

00:35:24.280 --> 00:35:27.480
you three vectors for any span right

00:35:26.280 --> 00:35:30.160
because you can always get the beginning

00:35:27.480 --> 00:35:33.280
that and in the mean and then based on

00:35:30.160 --> 00:35:36.560
that they feed that through um like a

00:35:33.280 --> 00:35:37.800
neural network and get a new edting so

00:35:36.560 --> 00:35:40.000
they feed that through a transformation

00:35:37.800 --> 00:35:42.520
and get a new edting and so that's the

00:35:40.000 --> 00:35:44.200
method that they used and I think our

00:35:42.520 --> 00:35:46.640
paper actually has a

00:35:44.200 --> 00:35:49.640
better

00:35:46.640 --> 00:35:52.640
um a better figure of how you can

00:35:49.640 --> 00:35:56.680
actually use that actually maybe it

00:35:52.640 --> 00:35:58.160
doesn't okay but anyway um yeah because

00:35:56.680 --> 00:36:00.240
uh yeah here's the figure

00:35:58.160 --> 00:36:01.520
so then you can use that for a number of

00:36:00.240 --> 00:36:03.040
things you could use that to like look

00:36:01.520 --> 00:36:06.359
up something in a knowledge base you

00:36:03.040 --> 00:36:08.599
could also use that to um decide whether

00:36:06.359 --> 00:36:10.440
two spans are co-referent by feeding in

00:36:08.599 --> 00:36:12.800
like the first span and the second Span

00:36:10.440 --> 00:36:14.960
in and then predicting whether those two

00:36:12.800 --> 00:36:19.640
spans cor correspond to each other or

00:36:14.960 --> 00:36:21.240
not so this general idea of modeling

00:36:19.640 --> 00:36:22.960
spans and then modeling relations

00:36:21.240 --> 00:36:24.520
between the spans allows you to solve

00:36:22.960 --> 00:36:26.119
like lots of different tasks like part

00:36:24.520 --> 00:36:27.920
of speech tagging or named entity

00:36:26.119 --> 00:36:30.319
recognition or relation extraction or

00:36:27.920 --> 00:36:31.920
other stuff like that so um yeah

00:36:30.319 --> 00:36:34.040
actually I realized now that I should

00:36:31.920 --> 00:36:35.079
have probably talked about these in the

00:36:34.040 --> 00:36:36.560
slides where I was talking about

00:36:35.079 --> 00:36:38.599
modeling but that that would be my

00:36:36.560 --> 00:36:42.319
recommended way of doing

00:36:38.599 --> 00:36:42.319
it cool any other

00:36:43.839 --> 00:36:49.480
questions nice okay

00:36:46.880 --> 00:36:52.880
um

00:36:49.480 --> 00:36:55.119
so another question is how can we inject

00:36:52.880 --> 00:36:56.640
knowledge into language models um

00:36:55.119 --> 00:36:58.720
there's a bunch of different ways to do

00:36:56.640 --> 00:37:03.079
this um

00:36:58.720 --> 00:37:05.000
one very easy way is to somehow look up

00:37:03.079 --> 00:37:09.640
relevant knowledge in your knowledge

00:37:05.000 --> 00:37:09.640
graph and um oh

00:37:10.280 --> 00:37:15.440
sorry I was presenting on my own screen

00:37:13.040 --> 00:37:18.240
not the screen that everybody can see so

00:37:15.440 --> 00:37:22.000
um to look up all of the uh knowledge in

00:37:18.240 --> 00:37:24.000
a Knowledge Graph and um somehow provide

00:37:22.000 --> 00:37:26.800
it to the model one way you can provide

00:37:24.000 --> 00:37:28.720
it to the model is through prompting um

00:37:26.800 --> 00:37:32.400
but the problem with with prompting is

00:37:28.720 --> 00:37:33.920
that you're not necessarily going to uh

00:37:32.400 --> 00:37:37.319
be able

00:37:33.920 --> 00:37:41.359
to utilize knowledge that is kind of

00:37:37.319 --> 00:37:43.920
like minority knowledge because the

00:37:41.359 --> 00:37:47.560
embeddings of the entities that you're

00:37:43.920 --> 00:37:49.440
presenting may not be you know like well

00:37:47.560 --> 00:37:51.839
learned so

00:37:49.440 --> 00:37:53.200
you're requiring essentially the model

00:37:51.839 --> 00:37:55.359
to be able to generalize from the

00:37:53.200 --> 00:37:57.880
knowledge you provide in

00:37:55.359 --> 00:38:00.839
the prompt despite the fact that the

00:37:57.880 --> 00:38:02.240
prompt is like minor entities or other

00:38:00.839 --> 00:38:07.040
things like that that are not as well

00:38:02.240 --> 00:38:10.400
learned so is another um method to

00:38:07.040 --> 00:38:13.440
handle this um we previously proposed a

00:38:10.400 --> 00:38:15.599
method that allows you

00:38:13.440 --> 00:38:18.319
to essentially

00:38:15.599 --> 00:38:21.319
predict instead of predicting directly

00:38:18.319 --> 00:38:24.920
the words here you can predict a tag

00:38:21.319 --> 00:38:27.200
that says birth name or a given name or

00:38:24.920 --> 00:38:31.480
family name or something like that and

00:38:27.200 --> 00:38:32.839
then post talk the model will fill in uh

00:38:31.480 --> 00:38:36.720
that like birth

00:38:32.839 --> 00:38:39.400
name text based on a knowledge base so

00:38:36.720 --> 00:38:41.079
um you know if you have a a Wikipedia

00:38:39.400 --> 00:38:44.240
article about Barack Obama that you're

00:38:41.079 --> 00:38:48.680
trying to write it could predict um

00:38:44.240 --> 00:38:52.040
birth name born uh birth name comma born

00:38:48.680 --> 00:38:55.359
in birth date and that's like a very

00:38:52.040 --> 00:38:56.880
very common thing in Wikipedia right so

00:38:55.359 --> 00:39:00.960
because of that it can predict it very

00:38:56.880 --> 00:39:03.160
consistently very uh formulaically and

00:39:00.960 --> 00:39:04.599
that allows you to um you know with high

00:39:03.160 --> 00:39:06.079
confidence get something that makes

00:39:04.599 --> 00:39:08.599
sense and is factual and reduce

00:39:06.079 --> 00:39:11.400
hallucination and other stuff like that

00:39:08.599 --> 00:39:12.599
so um basically how could you inject

00:39:11.400 --> 00:39:14.280
this into language models there's

00:39:12.599 --> 00:39:16.240
multiple ways one is prompting that's

00:39:14.280 --> 00:39:18.160
maybe the easier way another way is

00:39:16.240 --> 00:39:21.520
through like templatic generation like

00:39:18.160 --> 00:39:23.200
this where you generate placeholders uh

00:39:21.520 --> 00:39:25.200
for all the information you want to add

00:39:23.200 --> 00:39:26.480
and then you add the information uh

00:39:25.200 --> 00:39:29.359
directly from the knowledge base through

00:39:26.480 --> 00:39:29.359
the placeholders like

00:39:30.680 --> 00:39:36.800
cool um there there's details about this

00:39:34.240 --> 00:39:38.920
in the paper like how we um formulate a

00:39:36.800 --> 00:39:41.319
training objective for something like

00:39:38.920 --> 00:39:43.480
this and the difficulty in formulating a

00:39:41.319 --> 00:39:46.400
training objective is that you need to

00:39:43.480 --> 00:39:48.280
figure out when you want to replace

00:39:46.400 --> 00:39:49.720
things so like you might not always want

00:39:48.280 --> 00:39:51.000
to replace with birth name you might

00:39:49.720 --> 00:39:53.920
want to replace with given name and

00:39:51.000 --> 00:39:55.839
family name and we demonstrate that you

00:39:53.920 --> 00:39:58.400
can figure out how to do this by

00:39:55.839 --> 00:40:00.960
essentially like Mar iing over the

00:39:58.400 --> 00:40:03.520
various ways of uh of doing this but

00:40:00.960 --> 00:40:05.880
that's kind of more complex detail

00:40:03.520 --> 00:40:05.880
that's in the

00:40:08.440 --> 00:40:15.480
paper another really interesting

00:40:11.000 --> 00:40:17.319
question um that uh we this is a also a

00:40:15.480 --> 00:40:19.440
paper that I was involved in from uh

00:40:17.319 --> 00:40:22.040
four years ago but I feel like this is

00:40:19.440 --> 00:40:25.040
not entirely solved even in like modern

00:40:22.040 --> 00:40:26.920
rag systems uh today is how can we

00:40:25.040 --> 00:40:28.880
reason over a lot of text that's

00:40:26.920 --> 00:40:32.440
included in a knowledge

00:40:28.880 --> 00:40:35.839
base um oh sorry reason over Text corpus

00:40:32.440 --> 00:40:40.480
like we reason over knowledge bases

00:40:35.839 --> 00:40:43.280
and basically uh what we did was we

00:40:40.480 --> 00:40:44.960
answered questions using text corpora as

00:40:43.280 --> 00:40:48.680
a traceable knowledge

00:40:44.960 --> 00:40:52.800
bases and we did relevance matching over

00:40:48.680 --> 00:40:54.920
mentions um and the way we did this is

00:40:52.800 --> 00:40:57.440
we created mentioned

00:40:54.920 --> 00:40:59.480
vectors and the mentioned vectors

00:40:57.440 --> 00:41:01.720
vectors of all of the mentions in the

00:40:59.480 --> 00:41:04.920
knowledge base of particular

00:41:01.720 --> 00:41:05.920
entities um and then we retrieved

00:41:04.920 --> 00:41:09.599
relevant

00:41:05.920 --> 00:41:13.440
mentions um from pre-trained Models uh

00:41:09.599 --> 00:41:15.040
so we we ran embeddings and generated uh

00:41:13.440 --> 00:41:16.000
embeddings for each of the mentions in

00:41:15.040 --> 00:41:20.440
the whole

00:41:16.000 --> 00:41:25.440
Corpus and based on this let let

00:41:20.440 --> 00:41:29.119
me find the place over here so based on

00:41:25.440 --> 00:41:32.720
this we basically um encoded all of

00:41:29.119 --> 00:41:35.040
these uh in here and then we had a dense

00:41:32.720 --> 00:41:37.359
query vector and the dense query Vector

00:41:35.040 --> 00:41:41.640
was specifically trained so that it

00:41:37.359 --> 00:41:44.280
would be able to identify entity

00:41:41.640 --> 00:41:46.760
mentions that answered the problem so if

00:41:44.280 --> 00:41:50.240
we had like when was The Grateful Dead

00:41:46.760 --> 00:41:52.520
and uh Bob Dylan album released uh we

00:41:50.240 --> 00:41:54.760
would have Bob Dylan be one vector The

00:41:52.520 --> 00:41:56.560
Grateful Dead be another vector and the

00:41:54.760 --> 00:41:58.200
model would be specifically trained so

00:41:56.560 --> 00:42:00.040
that when you took took the entity

00:41:58.200 --> 00:42:03.319
embedding of this and matched it with an

00:42:00.040 --> 00:42:05.400
entity embedding in this big Corpus of

00:42:03.319 --> 00:42:07.920
encoded things here it would be most

00:42:05.400 --> 00:42:10.400
likely to return relevant information to

00:42:07.920 --> 00:42:13.160
answer these like entity relation

00:42:10.400 --> 00:42:14.680
questions so then the question is how do

00:42:13.160 --> 00:42:18.040
we train a model like this how do we

00:42:14.680 --> 00:42:20.280
train like a dense uh embedding model so

00:42:18.040 --> 00:42:21.520
that it gets relevant information for

00:42:20.280 --> 00:42:23.800
answering

00:42:21.520 --> 00:42:26.920
questions and basically the way we did

00:42:23.800 --> 00:42:29.280
this was through week supervision uh

00:42:26.920 --> 00:42:31.640
just like I talked about for relation

00:42:29.280 --> 00:42:33.599
extraction in relation extraction we can

00:42:31.640 --> 00:42:35.680
create weak supervision by taking a big

00:42:33.599 --> 00:42:37.960
existing knowledge base and identifying

00:42:35.680 --> 00:42:40.920
all of the sentences where the answer is

00:42:37.960 --> 00:42:43.319
included and so what we did is we took

00:42:40.920 --> 00:42:45.880
this big existing knowledge base and

00:42:43.319 --> 00:42:47.920
said okay what are some of the relations

00:42:45.880 --> 00:42:49.800
in the knowledge base one example of a

00:42:47.920 --> 00:42:51.559
relation in the knowledge base is Steven

00:42:49.800 --> 00:42:54.359
Spielberg is the director of Saving

00:42:51.559 --> 00:42:57.319
Private Ryan so we created questions

00:42:54.359 --> 00:42:59.119
that said um

00:42:57.319 --> 00:43:01.079
was the director of Saving Private Ryan

00:42:59.119 --> 00:43:03.920
we can create those with templates uh

00:43:01.079 --> 00:43:06.359
easily for many different relations and

00:43:03.920 --> 00:43:09.480
then we took the embedding for Saving

00:43:06.359 --> 00:43:10.760
Private Ryan in that question and we

00:43:09.480 --> 00:43:14.200
tried to

00:43:10.760 --> 00:43:17.119
upweight all of the Saving Private Ryan

00:43:14.200 --> 00:43:19.680
embeddings over all of Wikipedia where

00:43:17.119 --> 00:43:23.160
Steven Spielberg cooccurred in that

00:43:19.680 --> 00:43:25.640
sentence so that tries to match um you

00:43:23.160 --> 00:43:27.079
know artificially created questions with

00:43:25.640 --> 00:43:29.040
sentences that would be the answer

00:43:27.079 --> 00:43:31.040
answer to that question and so that

00:43:29.040 --> 00:43:32.480
gives you like supervision it gives you

00:43:31.040 --> 00:43:35.079
a lot of data to train over it gives you

00:43:32.480 --> 00:43:38.920
a good model so that that allowed us to

00:43:35.079 --> 00:43:41.319
learn this model well so um this is one

00:43:38.920 --> 00:43:43.160
example of how you can do like rag spe

00:43:41.319 --> 00:43:46.200
specifically like informed by knowledge

00:43:43.160 --> 00:43:46.200
bases and stuff like

00:43:47.280 --> 00:43:52.160
that um any any questions about this

00:43:53.480 --> 00:43:57.680
or

00:43:55.079 --> 00:44:00.079
okay so another thing that I I'd like to

00:43:57.680 --> 00:44:03.599
go into is uh something we call schema

00:44:00.079 --> 00:44:06.240
free extraction and so if I go back to

00:44:03.599 --> 00:44:09.960
the wiki Data

00:44:06.240 --> 00:44:10.760
Page um Wiki data has something we call

00:44:09.960 --> 00:44:13.599
a

00:44:10.760 --> 00:44:16.880
schema and the schema is basically like

00:44:13.599 --> 00:44:19.640
what are the relations that are included

00:44:16.880 --> 00:44:21.000
in the database so one of the relations

00:44:19.640 --> 00:44:25.079
that's included in the databas is

00:44:21.000 --> 00:44:25.079
instance of I guess also

00:44:25.200 --> 00:44:29.040
image lots of images

00:44:29.079 --> 00:44:33.880
um

00:44:30.440 --> 00:44:35.680
signature uh sex or gender country of

00:44:33.880 --> 00:44:38.319
citizenship and these relations are like

00:44:35.680 --> 00:44:41.079
decided a priori by the people who

00:44:38.319 --> 00:44:43.200
created Wiki data um and there's lots

00:44:41.079 --> 00:44:45.880
and lots of them but that doesn't

00:44:43.200 --> 00:44:48.880
necessarily mean

00:44:45.880 --> 00:44:50.400
that like similarly to the problem of

00:44:48.880 --> 00:44:51.839
not having all of the entities we can't

00:44:50.400 --> 00:44:55.119
have all of the relations and just to

00:44:51.839 --> 00:44:57.280
give one example I was um in preparation

00:44:55.119 --> 00:44:59.680
for our large language models lecture I

00:44:57.280 --> 00:45:02.640
actually created some structured data

00:44:59.680 --> 00:45:04.319
about large language models and some of

00:45:02.640 --> 00:45:06.119
the instru the structured data about

00:45:04.319 --> 00:45:09.319
large language models that I created was

00:45:06.119 --> 00:45:11.440
like what is the variety of positional

00:45:09.319 --> 00:45:13.079
embedding that they're using or

00:45:11.440 --> 00:45:15.800
positional embedding variety and

00:45:13.079 --> 00:45:18.720
positional embedding variety is not in

00:45:15.800 --> 00:45:20.359
Wiki data I think um I'd be surprised if

00:45:18.720 --> 00:45:23.200
it was in Wiki data but I think it's not

00:45:20.359 --> 00:45:25.760
in Wiki data um so like as you go down

00:45:23.200 --> 00:45:27.760
to like more esoteric Concepts or like

00:45:25.760 --> 00:45:29.599
specialized domains or stuff like that

00:45:27.760 --> 00:45:31.359
you're almost always guaranteed to not

00:45:29.599 --> 00:45:34.040
you know have all the entities you need

00:45:31.359 --> 00:45:36.680
or not have all the relations you need

00:45:34.040 --> 00:45:38.160
so that's the problem that schema free

00:45:36.680 --> 00:45:39.920
extraction is trying to solve it's

00:45:38.160 --> 00:45:41.680
trying to figure out how we can like

00:45:39.920 --> 00:45:45.920
jointly figure out the schema together

00:45:41.680 --> 00:45:45.920
with uh the information you want to

00:45:48.480 --> 00:45:54.040
extract and the um the most famous

00:45:52.319 --> 00:45:55.599
example of this is something called open

00:45:54.040 --> 00:45:57.200
information extraction in open

00:45:55.599 --> 00:46:01.160
information extraction basically what

00:45:57.200 --> 00:46:04.040
it's saying is um we don't need a schema

00:46:01.160 --> 00:46:06.359
uh there's no there's no schema um the

00:46:04.040 --> 00:46:08.720
only schema that we have is the actual

00:46:06.359 --> 00:46:12.200
text in the sentences that we're

00:46:08.720 --> 00:46:14.520
referring to um the entities so if we

00:46:12.200 --> 00:46:16.040
have United United has a Hub in Chicago

00:46:14.520 --> 00:46:17.359
which is the headquarters of United

00:46:16.040 --> 00:46:21.200
Continental

00:46:17.359 --> 00:46:25.880
Holdings um the relation is literally

00:46:21.200 --> 00:46:29.359
has a Hub in um that that's the relation

00:46:25.880 --> 00:46:33.359
um and then for this we have Chicago is

00:46:29.359 --> 00:46:35.559
the headquarters of um but the problem

00:46:33.359 --> 00:46:37.520
with this uh is that this cannot

00:46:35.559 --> 00:46:40.359
abstract away so if we had another

00:46:37.520 --> 00:46:42.000
sentence that said Chicago or United

00:46:40.359 --> 00:46:44.319
Continental Holdings has its

00:46:42.000 --> 00:46:45.720
headquarters in Chicago that would be

00:46:44.319 --> 00:46:49.800
treated as completely different you

00:46:45.720 --> 00:46:49.800
wouldn't be able to like group those two

00:46:51.119 --> 00:46:57.720
together so um in open information

00:46:55.000 --> 00:47:00.079
extraction actually a lot of the methods

00:46:57.720 --> 00:47:02.800
this is one of the few things where

00:47:00.079 --> 00:47:05.480
people still use rule-based systems as

00:47:02.800 --> 00:47:07.640
kind of like uh you know almost

00:47:05.480 --> 00:47:09.319
state-of-the-art systems but basically

00:47:07.640 --> 00:47:11.559
the reason why you're able to do this is

00:47:09.319 --> 00:47:14.440
it's not actually that hard to extract

00:47:11.559 --> 00:47:16.839
kind of the relevant strings between uh

00:47:14.440 --> 00:47:19.599
two entities and so the both the

00:47:16.839 --> 00:47:21.359
Precision and recall are pretty high and

00:47:19.599 --> 00:47:24.079
another reason why people use rule-based

00:47:21.359 --> 00:47:25.760
systems is because they um like you want

00:47:24.079 --> 00:47:27.440
to run it over the whole web and running

00:47:25.760 --> 00:47:29.079
a neural model over the whole web is

00:47:27.440 --> 00:47:32.000
expensive so you can use a role-based

00:47:29.079 --> 00:47:35.319
model so some examples of this include

00:47:32.000 --> 00:47:37.640
text Runner and Reverb um the basic

00:47:35.319 --> 00:47:41.000
ideas behind them is that you use a

00:47:37.640 --> 00:47:43.720
parser to extract um to do a syntactic

00:47:41.000 --> 00:47:45.760
analysis of the sentence um in extract

00:47:43.720 --> 00:47:47.640
during according to rules so for example

00:47:45.760 --> 00:47:50.160
the relation must contain a

00:47:47.640 --> 00:47:52.720
predicate um the subject and object must

00:47:50.160 --> 00:47:56.040
be noun phrases other things like

00:47:52.720 --> 00:47:57.640
this um and then what they did later is

00:47:56.040 --> 00:47:59.240
what they did in this this paper

00:47:57.640 --> 00:48:00.800
arguably this is maybe no longer

00:47:59.240 --> 00:48:02.280
necessary with the compute power we have

00:48:00.800 --> 00:48:04.000
now but they trained an even faster

00:48:02.280 --> 00:48:06.960
model to extract over large amounts of

00:48:04.000 --> 00:48:08.720
data so they basically um use this as a

00:48:06.960 --> 00:48:10.599
su weak supervision and then train a

00:48:08.720 --> 00:48:12.160
model that could do it even faster with

00:48:10.599 --> 00:48:14.680
the sequence base

00:48:12.160 --> 00:48:18.119
model

00:48:14.680 --> 00:48:19.880
um another thing that they did was um

00:48:18.119 --> 00:48:22.280
they aggregated multiple pieces of

00:48:19.880 --> 00:48:24.480
evidence heris to find common and

00:48:22.280 --> 00:48:28.760
therefore potentially reliable

00:48:24.480 --> 00:48:28.760
extractions so like

00:48:29.800 --> 00:48:36.960
any piece of text on the internet like

00:48:31.559 --> 00:48:40.200
could be a lie right so um you know

00:48:36.960 --> 00:48:43.400
if I I might write on my blog United has

00:48:40.200 --> 00:48:45.119
a Hub in like Denver or on the other

00:48:43.400 --> 00:48:48.240
hand

00:48:45.119 --> 00:48:50.839
um wait a set

00:48:48.240 --> 00:48:52.680
right some something has a Hub in Denver

00:48:50.839 --> 00:48:54.960
but United has a Hub in Pittsburgh is

00:48:52.680 --> 00:48:58.040
definitely wrong so let's uh let's go

00:48:54.960 --> 00:49:00.000
with that um uh so somebody could write

00:48:58.040 --> 00:49:02.359
that on the internet and in fact because

00:49:00.000 --> 00:49:06.440
I just said it it's probably in YouTube

00:49:02.359 --> 00:49:09.119
comments somewhere but um uh

00:49:06.440 --> 00:49:10.760
like any any piece of information on the

00:49:09.119 --> 00:49:13.079
internet could be wrong so basically

00:49:10.760 --> 00:49:16.680
they had um heuristic methods to filter

00:49:13.079 --> 00:49:19.559
these out and usually these were

00:49:16.680 --> 00:49:21.559
frequency based so it's like um if both

00:49:19.559 --> 00:49:23.520
United and Pittsburgh are very common

00:49:21.559 --> 00:49:26.000
but it's very rare for somebody to says

00:49:23.520 --> 00:49:27.799
say United has a Hub in Pittsburgh then

00:49:26.000 --> 00:49:29.200
that means it's statistically unlikely

00:49:27.799 --> 00:49:30.799
for this to be correct because if it

00:49:29.200 --> 00:49:33.280
were correct we'd expect to see it much

00:49:30.799 --> 00:49:36.799
more frequently so um those were the

00:49:33.280 --> 00:49:36.799
kind of things that they they did

00:49:37.520 --> 00:49:44.440
here there's also some neural models for

00:49:40.400 --> 00:49:46.839
open IE um I I think these are uh used

00:49:44.440 --> 00:49:48.440
maybe a little bit less often um but

00:49:46.839 --> 00:49:52.559
basically heuristics are still not

00:49:48.440 --> 00:49:55.280
perfect and so what they did the problem

00:49:52.559 --> 00:49:56.720
with um like not relying on heuristics

00:49:55.280 --> 00:49:58.880
is you need to get training data from

00:49:56.720 --> 00:50:01.880
somewhere so there's a rather clever

00:49:58.880 --> 00:50:03.599
paper um and again if you're not

00:50:01.880 --> 00:50:05.119
interested in relation extraction in

00:50:03.599 --> 00:50:07.559
particular I think this is one thing

00:50:05.119 --> 00:50:10.000
that's still worth paying attention to

00:50:07.559 --> 00:50:12.680
um which is

00:50:10.000 --> 00:50:14.559
they demonstrated that it's possible to

00:50:12.680 --> 00:50:16.319
create relatively large data sets by

00:50:14.559 --> 00:50:18.160
asking people simple

00:50:16.319 --> 00:50:21.440
questions

00:50:18.160 --> 00:50:24.480
and in particular they wanted to

00:50:21.440 --> 00:50:27.119
get relation extraction data sets that

00:50:24.480 --> 00:50:30.799
are like um

00:50:27.119 --> 00:50:34.200
who finished something like UCD finished

00:50:30.799 --> 00:50:37.760
the two 2006 championships and if you

00:50:34.200 --> 00:50:40.720
ask people like okay select this span um

00:50:37.760 --> 00:50:44.559
select the entity span the relations

00:50:40.720 --> 00:50:46.160
span and the um in the second entity the

00:50:44.559 --> 00:50:49.079
head entity the relation and the tail

00:50:46.160 --> 00:50:51.839
entity select it on this interface and

00:50:49.079 --> 00:50:54.200
then uh tell me is it this relation or

00:50:51.839 --> 00:50:55.640
this relation or this relation that's

00:50:54.200 --> 00:50:58.160
actually pretty hard and getting like

00:50:55.640 --> 00:51:01.280
crowd workers to start learning how to

00:50:58.160 --> 00:51:03.280
do that task is a bit tricky and it

00:51:01.280 --> 00:51:06.400
takes some you know it takes some time

00:51:03.280 --> 00:51:07.799
to get them onboarded basically um but

00:51:06.400 --> 00:51:09.760
basically what they said is instead

00:51:07.799 --> 00:51:11.359
we'll just ask them questions where the

00:51:09.760 --> 00:51:14.240
answer to the question basically gives

00:51:11.359 --> 00:51:17.160
us the answer to what the relation is so

00:51:14.240 --> 00:51:20.319
they ask like who finished something and

00:51:17.160 --> 00:51:23.680
the answer is like UCD and um what did

00:51:20.319 --> 00:51:25.359
someone finish the 2006 Championship

00:51:23.680 --> 00:51:28.920
what did someone fish some finish

00:51:25.359 --> 00:51:31.760
something as and basically um in doing

00:51:28.920 --> 00:51:33.319
this they created uh something called

00:51:31.760 --> 00:51:34.359
semantic roles which we're actually

00:51:33.319 --> 00:51:35.960
probably going to talk about a little

00:51:34.359 --> 00:51:37.559
bit later but you can take the semantic

00:51:35.960 --> 00:51:41.200
roles and then you can use them to

00:51:37.559 --> 00:51:43.920
annotate uh relation extraction data and

00:51:41.200 --> 00:51:46.720
then they trained a supervised neural

00:51:43.920 --> 00:51:46.720
tager for

00:51:48.799 --> 00:51:53.480
this

00:51:50.480 --> 00:51:56.040
cool um so another thing I'd like to

00:51:53.480 --> 00:51:57.880
talk about is I talked about learning um

00:51:56.040 --> 00:51:59.920
information about entities from entity

00:51:57.880 --> 00:52:02.079
embeddings but you can actually learn

00:51:59.920 --> 00:52:04.520
information about relations from

00:52:02.079 --> 00:52:07.680
relation information about other

00:52:04.520 --> 00:52:12.359
relations and this can help solve the

00:52:07.680 --> 00:52:16.119
problem um of like essentially the fact

00:52:12.359 --> 00:52:18.760
that open IE is not able to abstract and

00:52:16.119 --> 00:52:20.680
generalize so word embeddings or entity

00:52:18.760 --> 00:52:23.079
embeddings give information of the word

00:52:20.680 --> 00:52:26.920
in context um which can be indicative

00:52:23.079 --> 00:52:29.640
for knowledge uh knowledge bases

00:52:26.920 --> 00:52:32.640
but other relations or combinations

00:52:29.640 --> 00:52:34.960
thereof are also indicative of them and

00:52:32.640 --> 00:52:36.960
um if anybody is familiar with graphs or

00:52:34.960 --> 00:52:39.520
graph processing there's the whole idea

00:52:36.960 --> 00:52:41.400
of um link prediction where you're given

00:52:39.520 --> 00:52:42.680
like a a small number of links in a

00:52:41.400 --> 00:52:45.760
graph and you want to predict what other

00:52:42.680 --> 00:52:50.559
links are likely to uh

00:52:45.760 --> 00:52:52.920
exist and like as I said um a lot of uh

00:52:50.559 --> 00:52:54.839
you know very prominent AI researchers

00:52:52.920 --> 00:52:57.440
got their start in uh relation

00:52:54.839 --> 00:53:01.480
extraction and uh it sker is another one

00:52:57.440 --> 00:53:04.319
of them actually um and uh basically

00:53:01.480 --> 00:53:07.880
this 2009 paper proposed to use tensor

00:53:04.319 --> 00:53:09.400
de composition to do uh induction of

00:53:07.880 --> 00:53:13.520
relations

00:53:09.400 --> 00:53:15.319
and the way it worked is um you model

00:53:13.520 --> 00:53:18.400
relations by decomposing a tensor

00:53:15.319 --> 00:53:21.599
containing entity relation entity tles

00:53:18.400 --> 00:53:24.000
so you have the left entity the right

00:53:21.599 --> 00:53:27.160
entity and whether the relation exists

00:53:24.000 --> 00:53:31.319
is this big um uh big tensor in the

00:53:27.160 --> 00:53:33.160
Middle where these are embeddings of the

00:53:31.319 --> 00:53:35.760
left entity these are embeddings of the

00:53:33.160 --> 00:53:38.839
right entity and then the the depth of

00:53:35.760 --> 00:53:40.680
the tensor is like which relations exist

00:53:38.839 --> 00:53:43.760
and so we know that some exist so we

00:53:40.680 --> 00:53:46.640
give them a one we know others exist um

00:53:43.760 --> 00:53:48.680
don't exist so we give them a zero um

00:53:46.640 --> 00:53:51.040
and then we do a low rank approximation

00:53:48.680 --> 00:53:52.559
of this tensor and if we do a low rank

00:53:51.040 --> 00:53:55.720
approximation of the tensor we have

00:53:52.559 --> 00:53:57.280
reconstruction ER basically so when we

00:53:55.720 --> 00:53:59.960
reconstruct the are some things that

00:53:57.280 --> 00:54:01.960
were previously zero become one and so

00:53:59.960 --> 00:54:04.760
the things that were previously zero and

00:54:01.960 --> 00:54:07.880
then become close to one are the ones

00:54:04.760 --> 00:54:10.559
that we think like actually might exist

00:54:07.880 --> 00:54:12.000
they might be real um they might be real

00:54:10.559 --> 00:54:13.640
relations that we were just missing

00:54:12.000 --> 00:54:16.599
because our previous knowledge base was

00:54:13.640 --> 00:54:16.599
complete uh

00:54:18.640 --> 00:54:26.880
incomplete and um one thing that takes

00:54:21.799 --> 00:54:28.559
us a step further is uh what if if we

00:54:26.880 --> 00:54:30.079
actually do have a knowledge basee or

00:54:28.559 --> 00:54:31.839
what if we even have multiple knowledge

00:54:30.079 --> 00:54:35.520
bases like what if we have Wiki data and

00:54:31.839 --> 00:54:36.640
we have wordnet and we have um uh other

00:54:35.520 --> 00:54:38.920
things like

00:54:36.640 --> 00:54:40.680
this and in addition to that we also

00:54:38.920 --> 00:54:43.400
have open IE

00:54:40.680 --> 00:54:45.960
extractions so there's an idea of

00:54:43.400 --> 00:54:47.880
something called Universal schema and

00:54:45.960 --> 00:54:50.200
what Universal schema do is they embed

00:54:47.880 --> 00:54:55.119
relations from multiple schema or

00:54:50.200 --> 00:54:56.960
schemata in the same space and based on

00:54:55.119 --> 00:54:59.559
this they then

00:54:56.960 --> 00:55:01.359
predict which ones exist are likely to

00:54:59.559 --> 00:55:04.400
exist or which ones are not likely to

00:55:01.359 --> 00:55:06.680
exist so here we might have a free base

00:55:04.400 --> 00:55:08.640
or Wiki data we might have another uh

00:55:06.680 --> 00:55:11.559
kind of relation extraction data set

00:55:08.640 --> 00:55:15.480
called Tac and then on the training data

00:55:11.559 --> 00:55:17.040
set we have um like all of these uh

00:55:15.480 --> 00:55:20.240
things that are like positive or

00:55:17.040 --> 00:55:23.960
negative or something like this and then

00:55:20.240 --> 00:55:26.960
on the heldout data set we have only

00:55:23.960 --> 00:55:29.480
information about like open

00:55:26.960 --> 00:55:30.920
for example so um for all of the

00:55:29.480 --> 00:55:33.079
entities that exist in the knowledge

00:55:30.920 --> 00:55:34.839
base we know you know whether the

00:55:33.079 --> 00:55:36.039
relations exist for but for all the

00:55:34.839 --> 00:55:39.640
entities that don't exist in the

00:55:36.039 --> 00:55:41.760
database we don't know and so uh then

00:55:39.640 --> 00:55:43.839
just from the existence of open IE

00:55:41.760 --> 00:55:45.480
relations or non-existence of open IE

00:55:43.839 --> 00:55:47.920
relations we can predict that other

00:55:45.480 --> 00:55:49.359
relations might exist for example so

00:55:47.920 --> 00:55:51.079
this is a great way to combine the two

00:55:49.359 --> 00:55:53.920
together like open IE you can run it

00:55:51.079 --> 00:55:55.880
over you know very large data sets um

00:55:53.920 --> 00:55:58.000
but it doesn't have a good schema free

00:55:55.880 --> 00:56:00.400
uh Wiki data has a good schema but you

00:55:58.000 --> 00:56:02.960
can't you know it's all manually created

00:56:00.400 --> 00:56:04.720
so you can suggest other ones and one

00:56:02.960 --> 00:56:07.960
other like interesting thing is you can

00:56:04.720 --> 00:56:09.640
suggest other um things that might exist

00:56:07.960 --> 00:56:13.039
in Wiki data but you could also track

00:56:09.640 --> 00:56:15.039
that back to the original text that

00:56:13.039 --> 00:56:17.000
indicated that it might exist in Wiki

00:56:15.039 --> 00:56:18.720
data so then you could have a human go

00:56:17.000 --> 00:56:20.520
back and check it to make sure that

00:56:18.720 --> 00:56:24.200
that's actually true and trustworthy and

00:56:20.520 --> 00:56:24.200
other things like that

00:56:26.400 --> 00:56:31.400
cool um so if you like uh you like

00:56:29.400 --> 00:56:33.160
tensors or you like linear algebra or

00:56:31.400 --> 00:56:34.720
things like this this is maybe something

00:56:33.160 --> 00:56:37.880
that you could take a look at and think

00:56:34.720 --> 00:56:40.240
a little bit more about um any any

00:56:37.880 --> 00:56:40.240
questions

00:56:42.799 --> 00:56:46.240
here okay

00:56:46.880 --> 00:56:53.680
cool um so another thing I'd like to

00:56:50.640 --> 00:56:56.920
talk about is uh modeling relation paths

00:56:53.680 --> 00:57:00.359
so this is a really nice uh idea

00:56:56.920 --> 00:57:00.359
which is you

00:57:00.440 --> 00:57:05.000
can make inferences across multiple hops

00:57:04.240 --> 00:57:08.400
of

00:57:05.000 --> 00:57:12.280
relations um based on uh particular

00:57:08.400 --> 00:57:14.200
relations existing and so um multi-step

00:57:12.280 --> 00:57:17.280
passs can be informative for indicating

00:57:14.200 --> 00:57:20.000
whether individual relations exist so um

00:57:17.280 --> 00:57:24.400
for example uh given a word given a

00:57:20.000 --> 00:57:27.960
particular word in a paper title

00:57:24.400 --> 00:57:29.880
recommend a venue in which to the paper

00:57:27.960 --> 00:57:32.559
and so this is the the problem that they

00:57:29.880 --> 00:57:36.079
were trying to solve and then basically

00:57:32.559 --> 00:57:38.440
you have a word um you

00:57:36.079 --> 00:57:41.119
find if you have that word in your paper

00:57:38.440 --> 00:57:42.920
title you then find other papers that

00:57:41.119 --> 00:57:45.280
have that title uh that have that word

00:57:42.920 --> 00:57:48.359
in their title and those papers are in a

00:57:45.280 --> 00:57:52.039
journal and that gets a high weight with

00:57:48.359 --> 00:57:54.119
respect to like that your paper being

00:57:52.039 --> 00:57:56.839
you know relevant to that particular

00:57:54.119 --> 00:57:59.880
Journal you can also say

00:57:56.839 --> 00:58:01.000
okay I have a a word find papers with

00:57:59.880 --> 00:58:03.240
that word in the

00:58:01.000 --> 00:58:07.240
title find the first author of that

00:58:03.240 --> 00:58:09.280
paper find another paper uh that had

00:58:07.240 --> 00:58:11.599
that author as a first author and then

00:58:09.280 --> 00:58:13.240
find the Journal of it and they

00:58:11.599 --> 00:58:15.839
demonstrate a way where you can like

00:58:13.240 --> 00:58:18.280
expand these paths and feed them into a

00:58:15.839 --> 00:58:22.400
prediction model and use that to predict

00:58:18.280 --> 00:58:25.480
um you know additional relations so

00:58:22.400 --> 00:58:26.680
unlike this method here this method was

00:58:25.480 --> 00:58:29.240
saying like

00:58:26.680 --> 00:58:30.920
other single relations are indicative of

00:58:29.240 --> 00:58:34.160
a particular relation

00:58:30.920 --> 00:58:36.880
existing this paper is saying not just

00:58:34.160 --> 00:58:38.720
individual relations are indicative of

00:58:36.880 --> 00:58:40.640
another relation existing but actually

00:58:38.720 --> 00:58:43.839
relation paths are indicative of a

00:58:40.640 --> 00:58:46.400
relation existing so this is more um

00:58:43.839 --> 00:58:46.400
expressive

00:58:47.520 --> 00:58:55.359
basically um and this followup paper

00:58:52.640 --> 00:58:57.480
uh using differentiable logic rules

00:58:55.359 --> 00:59:00.799
actually made this endtoend

00:58:57.480 --> 00:59:03.079
trainable so this allows you to consider

00:59:00.799 --> 00:59:07.599
whole paths in a differentiable

00:59:03.079 --> 00:59:09.960
framework and so the way they did this

00:59:07.599 --> 00:59:13.359
is like if you have you know City in

00:59:09.960 --> 00:59:16.440
country and has office in country um

00:59:13.359 --> 00:59:18.920
that or sorry City and Country and has

00:59:16.440 --> 00:59:22.200
office in city that indicates has office

00:59:18.920 --> 00:59:24.160
in country and I I'm sure you know many

00:59:22.200 --> 00:59:26.760
people here have thought like learned

00:59:24.160 --> 00:59:29.520
about logic and you know and induction

00:59:26.760 --> 00:59:32.720
from or deduction from uh logic rules

00:59:29.520 --> 00:59:34.359
and stuff like this but the problem is

00:59:32.720 --> 00:59:37.079
deduction from logic rules is very

00:59:34.359 --> 00:59:39.039
fragile like there are cases where there

00:59:37.079 --> 00:59:41.119
are counter examples so if you say that

00:59:39.039 --> 00:59:43.280
something is always true deductively

00:59:41.119 --> 00:59:45.839
then um that can cause problems so in

00:59:43.280 --> 00:59:47.839
reality it's like if you have two pieces

00:59:45.839 --> 00:59:52.400
of information something can become much

00:59:47.839 --> 00:59:56.920
much more likely um and so you know just

00:59:52.400 --> 00:59:59.880
to give an example um somebody studying

00:59:56.920 --> 01:00:01.280
studying at CMU makes it very likely

00:59:59.880 --> 01:00:03.799
much more likely that they're studying

01:00:01.280 --> 01:00:06.359
computer science and much less likely

01:00:03.799 --> 01:00:08.000
that they're studying medicine or

01:00:06.359 --> 01:00:09.520
something like that but that doesn't

01:00:08.000 --> 01:00:11.720
mean that it like

01:00:09.520 --> 01:00:13.559
entirely the first one is definitely not

01:00:11.720 --> 01:00:15.480
entirely implied and I'm sure there's

01:00:13.559 --> 01:00:16.760
like a few people at CMU who are somehow

01:00:15.480 --> 01:00:18.440
studying medicine through a joint

01:00:16.760 --> 01:00:21.480
program with pit or something like that

01:00:18.440 --> 01:00:24.400
so you know like very it's very rare

01:00:21.480 --> 01:00:26.799
that logic rules are hard and fast and

01:00:24.400 --> 01:00:28.480
so basically what they do is they treat

01:00:26.799 --> 01:00:30.559
each path as a sequence of Matrix

01:00:28.480 --> 01:00:34.839
multiplies it where they have a rule

01:00:30.559 --> 01:00:36.599
weight um like this and um in the end

01:00:34.839 --> 01:00:38.359
that allows you to make a a prediction

01:00:36.599 --> 01:00:40.839
about whether a predic logic rule is

01:00:38.359 --> 01:00:40.839
correct or

01:00:40.880 --> 01:00:49.319
not um so this is uh i' I've been

01:00:46.880 --> 01:00:51.119
working mostly in like structured

01:00:49.319 --> 01:00:54.480
knowledge space structured knowledge

01:00:51.119 --> 01:00:56.599
graphs other uh other things like this

01:00:54.480 --> 01:00:59.760
um I I don't

01:00:56.599 --> 01:01:02.720
think there's a whole lot of work that

01:00:59.760 --> 01:01:05.640
directly applies this to language models

01:01:02.720 --> 01:01:07.319
um like differentiable logic rules and

01:01:05.640 --> 01:01:10.079
language models or things like that just

01:01:07.319 --> 01:01:12.440
because it's less clean it's you know uh

01:01:10.079 --> 01:01:13.839
harder um there there's a little bit of

01:01:12.440 --> 01:01:16.079
work which I'm going to talk about now

01:01:13.839 --> 01:01:18.599
but I think like this kind of work is

01:01:16.079 --> 01:01:21.440
interesting because a lot of models are

01:01:18.599 --> 01:01:23.119
not super great at reasoning and how to

01:01:21.440 --> 01:01:25.119
like allow them to be better at

01:01:23.119 --> 01:01:26.559
reasoning is kind of an open problem so

01:01:25.119 --> 01:01:28.039
learning from these old older works that

01:01:26.559 --> 01:01:30.200
did it in a more structured space and

01:01:28.039 --> 01:01:32.160
trying to figure out how to apply them

01:01:30.200 --> 01:01:34.400
to less structured spaces is still

01:01:32.160 --> 01:01:36.240
interesting I think

01:01:34.400 --> 01:01:39.160
so

01:01:36.240 --> 01:01:40.720
cool um then the final talk topic I want

01:01:39.160 --> 01:01:42.920
to talk about is probing knowledge in

01:01:40.720 --> 01:01:44.920
LMS and so we have these knowledge bases

01:01:42.920 --> 01:01:47.319
that encode you know tons and tons of

01:01:44.920 --> 01:01:49.880
knowledge um which allows us to figure

01:01:47.319 --> 01:01:52.200
out you know oh well how well do uh

01:01:49.880 --> 01:01:56.200
language models know about these

01:01:52.200 --> 01:01:59.079
things and so

01:01:56.200 --> 01:02:02.760
traditional um kind of QA machine

01:01:59.079 --> 01:02:04.799
reading comprehension rag models um

01:02:02.760 --> 01:02:06.359
usually referred to external resources

01:02:04.799 --> 01:02:10.039
to answer questions like Wikipedia

01:02:06.359 --> 01:02:14.359
articles um or things like this but then

01:02:10.039 --> 01:02:16.119
the question is without doing rag can we

01:02:14.359 --> 01:02:18.160
you know answer questions like what

01:02:16.119 --> 01:02:20.920
knowledge is

01:02:18.160 --> 01:02:24.079
encoded and so the first paper that kind

01:02:20.920 --> 01:02:26.520
of handled this sort of problem uh is

01:02:24.079 --> 01:02:29.200
this paper which actually was also

01:02:26.520 --> 01:02:33.359
called uh

01:02:29.200 --> 01:02:35.960
wama surprisingly um or released a

01:02:33.359 --> 01:02:41.000
resource called llama except it was l m

01:02:35.960 --> 01:02:44.880
a um but what they did is they

01:02:41.000 --> 01:02:46.960
uh used they in contrast to using

01:02:44.880 --> 01:02:50.000
structural queries like SQL or or

01:02:46.960 --> 01:02:52.119
Sparkle two query KBS they tried to use

01:02:50.000 --> 01:02:54.240
natural language prompts to query LM so

01:02:52.119 --> 01:02:58.160
this was actually one of the the first

01:02:54.240 --> 01:03:02.359
uh kind of paper on prompts uh prompting

01:02:58.160 --> 01:03:05.079
for uh language models in a way and the

01:03:02.359 --> 01:03:08.359
way they did this is they had um they

01:03:05.079 --> 01:03:10.039
did like Dante was born in mask and then

01:03:08.359 --> 01:03:13.279
they tried to fill in the mask using a

01:03:10.039 --> 01:03:15.839
mask language model and uh and output

01:03:13.279 --> 01:03:18.559
Florence so

01:03:15.839 --> 01:03:19.960
um when they did this work now now we

01:03:18.559 --> 01:03:21.359
don't do this quite as much but when

01:03:19.960 --> 01:03:23.520
they did this work they basically used

01:03:21.359 --> 01:03:25.440
the knowledge base as the ground truth

01:03:23.520 --> 01:03:28.880
and tried to probe whether the knowledge

01:03:25.440 --> 01:03:31.520
in in um in the knowledge base was also

01:03:28.880 --> 01:03:34.880
uh recoverable from the neural

01:03:31.520 --> 01:03:37.720
map um and they proposed the Llama

01:03:34.880 --> 01:03:39.760
Benchmark um basically it was manual

01:03:37.720 --> 01:03:42.480
prompts for 41 relations they created

01:03:39.760 --> 01:03:44.839
the prompts manually uh so like X was

01:03:42.480 --> 01:03:46.480
founded in y The Prompt template and

01:03:44.839 --> 01:03:49.400
they filled in the subjects and had the

01:03:46.480 --> 01:03:52.160
LMS uh for such as Bert predict the

01:03:49.400 --> 01:03:55.839
objects uh like blueberg LP was founded

01:03:52.160 --> 01:03:59.000
in mask and they demonstrated that like

01:03:55.839 --> 01:04:02.440
basically Elmo uh Transformer XL and

01:03:59.000 --> 01:04:04.960
Bert base got uh you know up to 31%

01:04:02.440 --> 01:04:06.480
accuracy now I'm sure uh the modern

01:04:04.960 --> 01:04:09.200
language models would have much higher

01:04:06.480 --> 01:04:11.279
accuracy than

01:04:09.200 --> 01:04:13.920
that

01:04:11.279 --> 01:04:17.839
um this is a a follow-up paper that we

01:04:13.920 --> 01:04:21.160
did to this um where we tried to do this

01:04:17.839 --> 01:04:23.400
multilingually um I I think this is

01:04:21.160 --> 01:04:25.680
really let

01:04:23.400 --> 01:04:29.520
me I think one thing that's interesting

01:04:25.680 --> 01:04:31.960
interesting about this paper is um even

01:04:29.520 --> 01:04:37.240
if you're not interested in multilingual

01:04:31.960 --> 01:04:38.920
stuff per se there is an interesting

01:04:37.240 --> 01:04:40.760
dichotomy about like what knowledge is

01:04:38.920 --> 01:04:43.079
included in LMS and whether we can

01:04:40.760 --> 01:04:46.000
retrieve it and the reason why I'm

01:04:43.079 --> 01:04:48.359
saying this is because in this paper

01:04:46.000 --> 01:04:51.200
we created

01:04:48.359 --> 01:04:52.599
queries from a knowledge base and

01:04:51.200 --> 01:04:54.160
because we created queries from a

01:04:52.599 --> 01:04:55.760
knowledge base and knowledge bases are

01:04:54.160 --> 01:04:57.240
multilingual we can also create

01:04:55.760 --> 01:05:00.039
multilingual queries from knowledge

01:04:57.240 --> 01:05:01.720
bases right so we can use exactly the

01:05:00.039 --> 01:05:03.359
same entities but just ask the same

01:05:01.720 --> 01:05:05.920
question in different languages and so

01:05:03.359 --> 01:05:07.480
we had a bunch of people manually uh

01:05:05.920 --> 01:05:10.119
create prompts for all of these

01:05:07.480 --> 01:05:13.000
languages here and you can see that in

01:05:10.119 --> 01:05:15.960
English it's much better at responding

01:05:13.000 --> 01:05:19.000
uh to these queries than it is in any

01:05:15.960 --> 01:05:21.039
other language and in particular like

01:05:19.000 --> 01:05:22.880
lower resource languages or languages

01:05:21.039 --> 01:05:26.400
that are less similar to English it did

01:05:22.880 --> 01:05:29.079
much worse and notably we we counted the

01:05:26.400 --> 01:05:32.160
answer correct if it got it

01:05:29.079 --> 01:05:34.279
um we we had two settings one setting is

01:05:32.160 --> 01:05:35.799
we counted the answer correct if it only

01:05:34.279 --> 01:05:38.359
if it answered in the language we

01:05:35.799 --> 01:05:39.680
queried it in but we in other setting we

01:05:38.359 --> 01:05:42.640
also counted the answer correct if it

01:05:39.680 --> 01:05:44.200
answered in any language so we um it

01:05:42.640 --> 01:05:46.640
didn't necessarily have to even know the

01:05:44.200 --> 01:05:48.200
name of the entity in that uh language

01:05:46.640 --> 01:05:50.520
and we would still count it

01:05:48.200 --> 01:05:54.720
correct and so what I mean by there's a

01:05:50.520 --> 01:05:56.440
dichotomy between the information that

01:05:54.720 --> 01:05:59.240
language models have

01:05:56.440 --> 01:06:02.480
encoded and whether they're able to

01:05:59.240 --> 01:06:02.480
retrieve it

01:06:02.680 --> 01:06:07.640
is in English it's able to answer the

01:06:06.000 --> 01:06:10.799
models we tested were able to answer

01:06:07.640 --> 01:06:13.000
like 177% of queries

01:06:10.799 --> 01:06:14.359
but if the fact that they're able to

01:06:13.000 --> 01:06:16.160
answer in English means that the

01:06:14.359 --> 01:06:18.520
language model quote unquote knows the

01:06:16.160 --> 01:06:20.200
answer right like it knows the answer in

01:06:18.520 --> 01:06:22.680
English we're asking exactly the same

01:06:20.200 --> 01:06:24.400
question in all the other languages so

01:06:22.680 --> 01:06:26.079
you know it should know the answer in

01:06:24.400 --> 01:06:27.680
the other languages too

01:06:26.079 --> 01:06:30.000
but it's not able to retrieve the answer

01:06:27.680 --> 01:06:33.079
because we asked in another language

01:06:30.000 --> 01:06:35.920
so um that brings up some interesting

01:06:33.079 --> 01:06:38.079
questions about how we can make models

01:06:35.920 --> 01:06:39.680
better at retrieving the the knowledge

01:06:38.079 --> 01:06:43.559
that they already know in English when

01:06:39.680 --> 01:06:45.520
you query them in other languages or um

01:06:43.559 --> 01:06:48.119
and there was another paper recently I

01:06:45.520 --> 01:06:52.720
don't know if I'd be able to find it um

01:06:48.119 --> 01:06:56.119
exactly which is um they

01:06:52.720 --> 01:07:01.799
prompted models with personas and so

01:06:56.119 --> 01:07:04.599
they said I um you know I am a old man I

01:07:01.799 --> 01:07:07.160
am an old woman I am a young man I am

01:07:04.599 --> 01:07:10.039
young woman I am a child or something

01:07:07.160 --> 01:07:12.799
like that um or they also talked about

01:07:10.039 --> 01:07:15.640
things like uh physical disabilities and

01:07:12.799 --> 01:07:17.200
things and they said um please answer

01:07:15.640 --> 01:07:19.640
this question after they prompted with a

01:07:17.200 --> 01:07:22.680
Persona and just having that Persona

01:07:19.640 --> 01:07:24.839
greatly changed the ability of the model

01:07:22.680 --> 01:07:26.400
to answer questions so it's this very

01:07:24.839 --> 01:07:28.200
weird thing which which is like the

01:07:26.400 --> 01:07:29.799
models are actually capable of answering

01:07:28.200 --> 01:07:31.520
the questions but based on how you probe

01:07:29.799 --> 01:07:32.880
them whether it's in like different

01:07:31.520 --> 01:07:34.599
languages or if you give them a

01:07:32.880 --> 01:07:36.839
different Persona they manage to answer

01:07:34.599 --> 01:07:39.000
things differently and so on the plus

01:07:36.839 --> 01:07:42.920
side like you can create you can make

01:07:39.000 --> 01:07:44.799
ways to reduce the language models

01:07:42.920 --> 01:07:45.920
performance by giving it like a Persona

01:07:44.799 --> 01:07:49.839
that shouldn't be good at answering

01:07:45.920 --> 01:07:53.279
questions or something like that um

01:07:49.839 --> 01:07:54.839
but on the plus side um like when you're

01:07:53.279 --> 01:07:57.279
doing code generation there was this

01:07:54.839 --> 01:07:58.960
magic prompt which is like um I have

01:07:57.279 --> 01:08:01.319
checked this carefully in all the unit

01:07:58.960 --> 01:08:03.240
tests pass and that would improve your

01:08:01.319 --> 01:08:05.760
code generation accuracy by like five

01:08:03.240 --> 01:08:07.559
five points or something like that so um

01:08:05.760 --> 01:08:09.240
you just get the the model in the right

01:08:07.559 --> 01:08:11.359
mood to answer the question accurately

01:08:09.240 --> 01:08:13.319
and it does a better job at doing it so

01:08:11.359 --> 01:08:15.960
it's kind of uh it goes in both

01:08:13.319 --> 01:08:15.960
directions I

01:08:16.679 --> 01:08:27.080
guess cool um yeah uh any any questions

01:08:23.679 --> 01:08:30.120
here um another thing that you can do uh

01:08:27.080 --> 01:08:31.000
is fine-tune models specifically so

01:08:30.120 --> 01:08:34.080
they're good at answering

01:08:31.000 --> 01:08:35.560
knowledge-based questions so um uh this

01:08:34.080 --> 01:08:38.080
paper demonstrated that you could find

01:08:35.560 --> 01:08:39.480
tune models uh on synthetically created

01:08:38.080 --> 01:08:41.159
knowledge based questions and that would

01:08:39.480 --> 01:08:42.920
improve the ability of the model to

01:08:41.159 --> 01:08:47.679
answer questions about knowledge

01:08:42.920 --> 01:08:47.679
bases um it's

01:08:49.120 --> 01:08:57.440
uh yeah um it's pretty straightforward

01:08:53.199 --> 01:08:57.440
so uh there's that

01:08:57.799 --> 01:09:03.120
um yeah we already talked about this in

01:09:00.000 --> 01:09:07.560
the rag class so I think I might skip

01:09:03.120 --> 01:09:10.239
that um a final paper that I'd like to

01:09:07.560 --> 01:09:12.600
talk about this is also a paper uh done

01:09:10.239 --> 01:09:13.759
by my student Jung B Jong and this is

01:09:12.600 --> 01:09:16.080
interesting from the point of view of

01:09:13.759 --> 01:09:18.000
multihop reasoning and so I talked a

01:09:16.080 --> 01:09:19.679
little bit about like multihop reasoning

01:09:18.000 --> 01:09:23.239
along reasoning

01:09:19.679 --> 01:09:26.159
chains um in knowledge bases and this is

01:09:23.239 --> 01:09:28.520
one example of multihop reasoning

01:09:26.159 --> 01:09:30.080
among along reasoning chains within the

01:09:28.520 --> 01:09:33.400
parameters of the model so testing

01:09:30.080 --> 01:09:36.759
whether models can answer

01:09:33.400 --> 01:09:38.480
um Can it answer multihop questions and

01:09:36.759 --> 01:09:40.839
basically what we did here is we took a

01:09:38.480 --> 01:09:42.679
knowledge base and a knowledge base can

01:09:40.839 --> 01:09:44.279
have

01:09:42.679 --> 01:09:49.480
um

01:09:44.279 --> 01:09:49.480
like uh country country is

01:09:49.600 --> 01:09:52.600
US

01:09:53.480 --> 01:09:58.600
president um and then a

01:10:00.880 --> 01:10:06.560
birthday um and so we can create these

01:10:04.280 --> 01:10:08.640
multihop questions right uh and just

01:10:06.560 --> 01:10:10.280
follow the relation links and then we

01:10:08.640 --> 01:10:11.440
know the answer to the multihop question

01:10:10.280 --> 01:10:13.560
by following the link and we can

01:10:11.440 --> 01:10:18.159
generate you know the question given a

01:10:13.560 --> 01:10:19.800
template um so we did this and had like

01:10:18.159 --> 01:10:22.800
question one which is return the artist

01:10:19.800 --> 01:10:25.719
who recorded party a over um and then

01:10:22.800 --> 01:10:28.159
where in Georgia does uh Usher live and

01:10:25.719 --> 01:10:29.920
then we can turn this into a question

01:10:28.159 --> 01:10:31.679
which part of Georgia in which part of

01:10:29.920 --> 01:10:34.239
Georgia does the artist that recorded

01:10:31.679 --> 01:10:37.560
the party8 overlive and so we now have a

01:10:34.239 --> 01:10:45.000
multi multihop question and what we did

01:10:37.560 --> 01:10:47.440
is we measured whether um the model was

01:10:45.000 --> 01:10:49.760
able to answer the first question the

01:10:47.440 --> 01:10:53.320
second question and the comp like

01:10:49.760 --> 01:10:56.120
compound question and what we found is

01:10:53.320 --> 01:10:59.440
like what we would expect

01:10:56.120 --> 01:11:01.719
if models were like perfect knowledge

01:10:59.440 --> 01:11:04.360
processors right

01:11:01.719 --> 01:11:08.120
is we have

01:11:04.360 --> 01:11:10.800
like yes on the first question

01:11:08.120 --> 01:11:14.000
no

01:11:10.800 --> 01:11:16.560
yes um yes on the first question and no

01:11:14.000 --> 01:11:16.560
on the first

01:11:17.199 --> 01:11:24.760
question and we would expect that

01:11:21.920 --> 01:11:26.080
basically if it knew both of the answers

01:11:24.760 --> 01:11:27.239
to the first question and the second

01:11:26.080 --> 01:11:30.600
question it would get the compound

01:11:27.239 --> 01:11:31.800
question right and if it got uh like

01:11:30.600 --> 01:11:34.800
either of them wrong it would get it

01:11:31.800 --> 01:11:37.120
wrong right um you know in the in the

01:11:34.800 --> 01:11:39.400
ideal world where the knowledge of the

01:11:37.120 --> 01:11:41.280
two sub questions is necessary to answer

01:11:39.400 --> 01:11:43.880
the comp composite question and the

01:11:41.280 --> 01:11:45.840
model is a perfect knowledge processor

01:11:43.880 --> 01:11:47.120
and basically what we found we tried a

01:11:45.840 --> 01:11:49.280
whole bunch of different types of

01:11:47.120 --> 01:11:51.199
questions and what we found is this is

01:11:49.280 --> 01:11:55.960
totally not the case like it's not the

01:11:51.199 --> 01:11:58.520
case at all um and what we found in said

01:11:55.960 --> 01:12:01.560
is if it's able to answer the second

01:11:58.520 --> 01:12:04.120
question correctly it was much more

01:12:01.560 --> 01:12:07.480
likely to be able to answer the

01:12:04.120 --> 01:12:08.840
composite question um even if it can

01:12:07.480 --> 01:12:11.000
answer the first question that has

01:12:08.840 --> 01:12:13.120
almost no relation with whether it could

01:12:11.000 --> 01:12:15.520
answer the composite question at all so

01:12:13.120 --> 01:12:17.679
it's more like somehow from the answer

01:12:15.520 --> 01:12:19.320
to the second question it was able to to

01:12:17.679 --> 01:12:22.280
get the answer right and it kind of

01:12:19.320 --> 01:12:24.040
makes sense actually because like um

01:12:22.280 --> 01:12:26.320
let's say the answer to the second

01:12:24.040 --> 01:12:27.920
question is some like really long list

01:12:26.320 --> 01:12:30.719
like who are all the presidents of the

01:12:27.920 --> 01:12:33.320
United States um or something like that

01:12:30.719 --> 01:12:35.639
that's just hard to answer um so if I

01:12:33.320 --> 01:12:38.000
said who are all the presidents of the

01:12:35.639 --> 01:12:40.800
country where Washington DC is located

01:12:38.000 --> 01:12:42.679
in um you know like the second question

01:12:40.800 --> 01:12:44.040
is really hard so that's hard to get but

01:12:42.679 --> 01:12:46.120
if I say

01:12:44.040 --> 01:12:49.920
um

01:12:46.120 --> 01:12:53.520
uh what what is the

01:12:49.920 --> 01:12:57.120
capital what is the capital of the

01:12:53.520 --> 01:12:57.120
country uh

01:12:57.400 --> 01:13:02.440
what is what is the capital of the

01:12:58.840 --> 01:13:05.400
country where the most

01:13:02.440 --> 01:13:06.800
um people live or something like that

01:13:05.400 --> 01:13:08.679
even if you weren't sure about the

01:13:06.800 --> 01:13:10.880
country where the most people live you

01:13:08.679 --> 01:13:13.040
could pick a random capital and get it

01:13:10.880 --> 01:13:16.199
right some of the time or something like

01:13:13.040 --> 01:13:18.239
that so um that's what we found in this

01:13:16.199 --> 01:13:19.800
paper and I I think like another nice

01:13:18.239 --> 01:13:22.360
thing about knowledge bases is they

01:13:19.800 --> 01:13:24.880
allow you to ask like really interesting

01:13:22.360 --> 01:13:26.400
questions like this about what language

01:13:24.880 --> 01:13:29.120
model know or what language models don't

01:13:26.400 --> 01:13:31.040
know in a structured way so um I think

01:13:29.120 --> 01:13:32.280
if you're interested in probing language

01:13:31.040 --> 01:13:35.320
models and what they know and what they

01:13:32.280 --> 01:13:38.639
can infer what logic they can do that's

01:13:35.320 --> 01:13:42.320
good um cool yeah that's all I have for

01:13:38.639 --> 01:13:44.920
today um are there any questions or

01:13:42.320 --> 01:13:48.679
discussion or things like that or happy

01:13:44.920 --> 01:13:48.679
to talk up here too