Transcript slide presentation

Large-Scale Extraction and Use of
Knowledge from Text
Peter Clark and Phil Harrison
Boeing Research and Technology
Seattle, WA
Motivation and Goals
 World knowledge is required for many AI tasks
 especially natural language processing
“The man ate spaghetti with a fork”
 BUT: Not many knowledge resources are available
 This work: Create large knowledge resource
 called DART (Discovery& Aggregation of Relations in Text)
 based on Schubert’s conjecture
 publically available
Outline


Extracting Knowledge from Text
Using the Knowledge
1. Parsing
2. Recognizing Textual Entailment
3. Human Evaluation

Future Work
Extracting Knowledge from Text
Schubert’s Conjecture:
There is a largely untapped source of general
knowledge in texts, lying at a level beneath the explicit
assertional content, which can be harnessed.
“The camouflaged helicopter landed near the embassy.”
 helicopters can land
 helicopters can be camouflaged
Extracting Knowledge from Text
Newswire Article
HUTCHINSON SEES HIGHER
PAYOUT. HONG KONG. Mar 2.
Li said Hong Kong’s property
market remains strong while its
economy is performing better than
forecast. Hong Kong Electric
reorganized and will spin off its
non-electricity related activities.
Hongkong Electric shareholders
will receive one share in the new
subsidiary for every owned share
in the sold company. Li said the
decision to spin off …
Implicit, tacit knowledge
Shareholders may receive shares.
Shares may be owned.
Companies may be sold.
Extracting Knowledge from Text
Approach:
 Parse each sentence
 Extract “tuples” from selected nodes in the
parse tree
 Tuple = parse fragment retaining head words only
"The lazy men from the city walked to the fancy store"
(S "man" "walk" (PP "to" "store")) “Men can walk to stores”
(NPN "man" "from" "city")
“Men can be from cities”
(AN "lazy" "man")
“Men can be lazy”
(VPN "walk" "to" "store")
“Stores can be walked to”
(AN "fancy" "store")
“Stores can be fancy”
How Tuples are Extracted from the Parse Tree
Extracting Knowledge from Text
"The lazy men from the city walked to the fancy store"
(S "man" "walk" (PP "to" "store"))
(NPN "man" "from" "city")
(AN "lazy" "man")
(VPN "walk" "to" "store")
S
+-----------------+------------------+
(AN "fancy"
NP
VP
"store")
+--------+---------+
+------------+---+
NP
PP
VP
PP
+------+--+
+-----+--+
|
+---------+--+
DET
N
P
NP
V
P
NP
|
+--+--+
|
+--+---+
|
|
+------+---+
"the"
A
N
"from"
DET
N
"walked" "to" DET
N
|
|
|
|
|
+---+--+
"lazy" "men"
"the" "city"
"the"
A
N
|
|
"fancy" "store"
Extracting Knowledge from Text
Some details…
 Coordinate structures expanded
"The men and women ate beans and rice"
(S "man" "eat" "bean")
(S "woman" "eat" "bean")
(S "man" "eat" "rice")
(S "woman" "eat" "rice")
 Questions transformed
"Which book did John bring?"
(S (PN "John") "bring" "book")
 normalize the structures into 12 canonical forms:
(AN "small" "hotel")
"Hotels can be small."
(ANPN "subject" "agreement" "to" "approval")
"Agreements can be subject to approvals."
(NN "drug" "distributor")
"There can be drug distributors."
(NV "bus" "carry")
"Buses can carry [something/someone]
(NPN "sentence" "for" "offence")
"Sentences can be for offences."
(NVN "critic" "claim" "thing")
"Critics can claim things."
(NVPN "person" "go" "into" "room") "People can go into rooms."
(NVNPN "democrat" "win" "seat" "in" "election")
"Democrats can win seats in elections."
(QN "year" "contract")
"Contracts can be measured in years."
(VN "find" "spider")
"Spiders can be found."
(VPN "refer" "to" "business") "Referring can be to businesses."
(VNPN "educate" "person" "at" "college")
"People can be educated at colleges.”
Generating the DART Database
 Run over Reuters + British National Corpus
 2 months CPU
 Result: the DART database
(Discovery and Aggregation of Relations in Text)
 23 million propositions (110 million with duplicates)
Some Snapshots into the Database
 NV propositions about “planes”
294 (nv "plane" "carry")
262 (nv "plane" "land")
234 (nv "plane" "crash")
162 (nv "plane" "have")
159 (nv "plane" "take off")
100 (nv "plane" "leave")
85 (nv "plane" "bomb")
78 (nv "plane" "arrive")
…..
1 (nv "plane" "boast")
1 (nv "plane" "battle")
1 (nv "plane" "assist")
1 (nv "plane" "age")
1 (nv "plane" "advise")
; "planes can carry."
; "planes can land."
; "planes can crash."
; "planes can have."
; "planes can take off."
; "planes can leave."
; "planes can bomb."
; "planes can arrive.“
; "planes can boast."
; "planes can battle."
; "planes can assist."
; "planes can age."
; "planes can advise."
Some Snapshots into the Database
 VN propositions about “planes”
151 (vn "hijack" "plane")
; "planes can be hijacked."
149 (vn "buy" "plane")
; "planes can be bought."
142 (vn "use" "plane")
; "planes can be used."
108 (vn "board" "plane")
; "planes can be boarded."
101 (vn "send" "plane")
; "planes can be sent."
89 (vn "fly" "plane")
; "planes can be flown.”
87 (vn "have" "plane")
; "planes can be had."
78 (vn "sell" "plane")
; "planes can be sold."
74 (vn "operate" "plane")
; "planes can be operated."
70 (vn "build" "plane")
; "planes can be built."
…..
Some Snapshots into the Database
 NVNPN tuples about “planes”
9 (nvnpn "plane" "leave" "place" "for" "place")
; "planes can leave places for places."
8 (nvnpn "plane" "take" "person" "to" "place")
; "planes can take people to places."
7 (nvnpn "plane" "carry" "passenger" "on" "flight")
; "planes can carry passengers on flights."
7 (nvnpn "plane" "carry" "crew" "on" "flight")
; "planes can carry crews on flights."
6 (nvnpn "state" "sell" "plane" "to" "country")
; "states can sell planes to countries."
6 (nvnpn "plane" "have" "design" "with" "capability")
; "planes can have designs with capabilities."
6 (nvnpn "plane" "go" "thing" "to" "place")
; "planes can go things to places."
6 (nvnpn "person" "hijack" "plane" "to" "place")
; "people can hijack planes to places."
5 (nvnpn "plane" "police" "zone" "in" "place")
; "planes can police zones in places."
Some Snapshots into the Database
 What can be seen through?
(
(
(
(
(
(
(
22 vpn "see" "through" "window")
10 vpn "see" "through" "glass")
5 vpn "see" "through" "binocular")
3 vpn "see" "through" "porthole")
2 vpn "see" "through" "sky")
2 vpn "see" "through" "telescope")
2 vpn "see" "through" "keyhole")

…
(
(
(
(
….
37 vpn "see" "through" "person")
9 vpn "see" "through" "end")
4 vpn "see" "through" "hype")
1 vpn "see" "through" "dress")
?
Some Snapshots into the Database
 What color are sheep?
76 (AN "black" "sheep") Sheep can be black.
8 (AN "white" "sheep") Sheep can be white.
4 (AN "brown" "sheep") Sheep can be brown.
2 (AN “green" "sheep") Sheep can be green.
2 (AN “blue" "sheep") Sheep can be blue.
Some Snapshots into the Database
 What color are sheep?
76 (AN "black" "sheep") Sheep can be black.
8 (AN "white" "sheep") Sheep can be white.
4 (AN "brown" "sheep") Sheep can be brown.
2 (AN “green" "sheep") Sheep can be green.
2 (AN “blue" "sheep") Sheep can be blue.
?
Some Snapshots into the Database
 What color are sheep?
76 (AN "black" "sheep") Sheep can be black.
8 (AN "white" "sheep") Sheep can be white.
4 (AN "brown" "sheep") Sheep can be brown.
2 (AN “green" "sheep") Sheep can be green.
2 (AN “blue" "sheep") Sheep can be blue.
An Irishman said, for no apparent reason, "I wear a green
pullover because in Ireland we have green sheep."
?
Some Snapshots into the Database
 What color are sheep?
76 (AN "black" "sheep") Sheep can be black.
8 (AN "white" "sheep") Sheep can be white.
4 (AN "brown" "sheep") Sheep can be brown.
2 (AN “green" "sheep") Sheep can be green.
2 (AN “blue" "sheep") Sheep can be blue.
?
Sources of Error
1. Misparses
2. Bad part of speech assignment

“coal mine” → “coals can mine”
3. Textual statements that don’t reflect reality

jokes, metaphor, hypotheticals
4. Negation that gets lost

“I don’t believe in green sheep” → “sheep can be green”
5. Rhetorical vocabulary taken literally

“The cost of certain foods rose” → “foods can be certain”
6. Overgeneralization to something meaningless

“Things can be of things”
Interpreting the frequencies
 Can’t rely on raw frequency counts
 common words → high frequency counts
 Need to normalize for prior probabilities
 Mutual Information:
 how much the observation goes beyond chance
MI = log
observed frequency of w1 and w2 together
random chance that w1 and w2 will be together
p(NN w1 w2 )
MI  log
p(NN w1 *). p(NN * w2 )
Outline


Extracting Knowledge from Text
Using the Knowledge
1. Parsing
2. Recognizing Textual Entailment
3. Human Evaluation

Future Work
Using the DART Database: 1. Improving Parsing
“The man ate spaghetti with a fork”
 Ambiguous!
→ “ate using a fork”?
→ “eating spaghetti with a fork inside it”?
DART propositions can help!
(VPN “eat” “with” “fork”)
(NPN “spaghetti” “with” ‘fork”)
4 examples in DART
0 examples in DART
So: bias the parser to prefer “eat with fork”
Using the DART Database: 1. Improving Parsing
Count = 4, MI = 0.74
Count = 0, MI = 0
(VPN “eat” “with” “fork”)?
OR
eat spaghetti with
(NPN “spaghetti” “with” ‘fork”)?
?
eat
spaghetti
with
a
fork
a
fork
Using the DART Database: 1. Improving Parsing
Count = 4, MI = 0.74
(VPN “eat” “with” “fork”)?
Count = 0, MI = 0
(NPN “spaghetti” “with” ‘fork”)?
?
eat
spaghetti
eat spaghetti with
with
a

fork
a
fork
Using the DART Database: 1. Improving Parsing
Evaluation
 Use the Brown corpus
 1 million words, with “gold” parses
 Measure parser performance without/with DART
Mean Relative
Precision (%):
Without tuples: 46.3
With tuples: 47.1
Mean Relative
Recall (%):
76.9
77.6
 Often cases with zero-frequency tuple counts
 Tradeoff between tuple bias vs. built-in parser bias
Outline


Extracting Knowledge from Text
Using the Knowledge
1. Parsing
2. Recognizing Textual Entailment
3. Human Evaluation

Future Work
Uses: 2. Recognizing Textual Entailment (RTE)
 The task: Given T, does H “reasonably” follow?
T: The president visited Iraq.
H: The president traveled to Iraq. Answer: Yes
 Annual competition
 Common approach = use paraphrase rules, e.g.,
IF X visits Y THEN X travels to Y
Uses: 2. Recognizing Textual Entailment (RTE)
 BUT: rules can be overgeneral:
IF X shoots Y THEN X injures Y
 “Joe shot Mary” → “Joe injured Mary” (plausible)
 “Joe shot the gun” → “Joe injured the gun” (!!!)
 Some work on type restrictions on X, Y
 BUT: limited success (Y = animate#n1 or artifact#n1)
 DART propositions can help!
Uses: 2. Recognizing Textual Entailment (RTE)
Rule:
IF X shoots Y THEN X injures Y
Is this instantiation plausible?
No!
IF Joe shoots a gun THEN Joe injures a gun
?
No
injured
guns
Uses: 2. Recognizing Textual Entailment (RTE)
None:
Average Precision
of rule-based
entailments for:
RTE3 RTE4
0.592
0.650
Original rule confidence:
0.634
0.712
Tuple (DART) derived confidence
0.641
0.728
Confidence Assignment Method:
Outline


Extracting Knowledge from Text
Using the Knowledge
1. Parsing
2. Recognizing Textual Entailment
3. Human Evaluation

Future Work
Assessing the DART Database: Human Evaluation
How much “junk” does DART contain?
 Subjects rated random propositions in DART
 true / partially true / unsure / mainly false / false
Tuple in DART
Verbalized proposition
Score
(1-5)
(nvn "person" "enter" "place") "People can enter places."

(npn "sanction" "against" "place") “Sanctions can be against places"

(npn “people” “love” “wall”) “People can love walls.”

(an "french" "counterpart")
“Counterparts can be french."
(nv “coal” “mine")
“Coals can mine"
Assessing the DART Database: Human Evaluation
 Divide propositions into 4 “buckets”
 Frequencies: 1, 2-10, 11-100, >100
 12 judges rated ~300 propositions each (~75 per bucket)
 ~3600 total
0.9
Fraction true/partially true
0.8
Fraction
true/
partially
true
~70% of
propositions seen
>10 times were
ALL
judged
true/partially true
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
2 to 10
11 to 100
>100
Figure 2: Frequency
of propositionininDART
DART
Frequency
of proposition
database
Outline


Extracting Knowledge from Text
Using the Knowledge
1. Parsing
2. Recognizing Textual Entailment
3. Human Evaluation

Discussion and Future Work
Other Resources
 KNEXT – inspiration for this work
 we’ve expanded size, types of proposition, and evaluation
 TextRunner
 larger, less structured database of triples (only), e.g.,
(“iodine” “will kill” “the lactic bacteria”)
(“my friends” “had” “a car”)
 ConceptNet
 by Web volunteers, ~20 relations, less structured, e.g.
(CapableOf “a car” “crash”)
(UsedFor “a car” “getting to work”)
Future Directions: Generalizing DART propositions
 Biggest limitation:
 Still often near-zero tuples for “good” propositions
 Can we generalize the propositions in DART?
 thus cover the zero/near-zero cases
Future Directions: Generalizing DART propositions
(NPN “giraffe” “in” “zoo”)
frequency in DART = 0
but…
(NPN “lion” “in” “zoo”)
(NPN “monkey” “in” “zoo”)
(NPN “elephant” “in” “zoo”)
frequency in DART > 0
(NPN “animal” “in” “zoo”)
?
(NPN animal#n#1 “in” “zoo”)
(NPN lion#n#1 “in” “zoo”)
(NPN monkey#n#1 “in” “zoo”)
(NPN elephant#n#1 “in” “zoo”)
?
?

Future Directions: Finding inter-proposition relationships
 Use part of DART to disambiguate other parts, eg
(NN “car” “engine”)
“There can be car engines”
 what is the relation between “car” and “engine”?
 other tuples can tell us what it is! e.g.:
(VNPN “car” “power”
“Cars can be powered by engines”
“by” “engine”)
(NPN “engine” “in” “car”) “Engines can be in cars”
Future Directions: Natural Language Understanding
 Use DART to fill in implicit knowledge for NLP
“John had studied hard, but he was still worried”
worried about what?
DART: (ANPN “worried” “person” “about” ?X)
→ ?X = survival, sellof, plan, export, exam, delay, …
studied for what?
DART: (VPN “study” “for” ?X)
→ ?X = ..., exam, project, …
Future Directions: Natural Language Understanding
 Use DART to fill in implicit knowledge for NLP
“John had studied hard, but he was still worried”
worried about what?
DART: (ANPN “worried” “person” “about” ?X)
→ ?X = survival, sellof, plan, export, exam, delay, …
studied for what?
DART: (VPN “study” “for” ?X)
→ ?X = ..., exam, …
Summary
 DART (Discovery and Aggregation of Relations in Text)
 23 million unique, general propositions
 acquired through text mining
 Appears to be useful
 Improves parsing
 Improves recognizing textual entailment
 Judges say ~50% to 70% is true / partially true
 Many other ways it can be used!
 Give it a try! (Google “DART database”)