combining relational and attributionalsimilarity for...
TRANSCRIPT
Combining Relational and Attributional Similarity
for Semantic Relation Classification
Preslav Nakov, National University of Singapore
Zornitsa Kozareva, University of Southern California
RANLP 2011
RANLP: Hissar, Bulgaria: September 14, 2011
2
What causes tumors to shrink?
The period of tumor shrinkage after radiation
therapy is often long and varied.
CAUSE-EFFECT
How do we build a system to classify the relation between two nouns?
RANLP: Hissar, Bulgaria: September 14, 2011
3Nakov & Kozareva: Combining Relational and Attributional Similarity ... 3
Relation Extraction (between Nouns)
• Given a pair of nouns, identify the semantic
relation(s) between them
malaria mosquito
effect-cause
content-container
orange basket
RANLP: Hissar, Bulgaria: September 14, 2011
4Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Semantic Relations: Applications
� Help real applications:
�information extraction
�document summarization
�machine translation
�construction of thesauri and semantic networks
� Facilitate auxiliary tasks:
�word sense disambiguation
�language modeling
�paraphrasing
�recognizing textual entailment
RANLP: Hissar, Bulgaria: September 14, 2011
5Nakov & Kozareva: Combining Relational and Attributional Similarity ...
� PART-WHOLE (Winston, Chaffin, and Hermann 1987)
� Component-Integral object
� cup-handle, kitchen-apartment, wing-bird
� Member-Collection
� soldier-army, professor-faculty, tree-forest
� Portion-Mass
� slice-pie, meter-kilometer
� Stuff-Object
� silk-dress, steel-car, and alcohol-wine
� Feature-Activity
� paying-shopping, chewing-eating
� Place-Area
� Geographic�oasis-desert, county-state, path-forest
� Geometric�the end (of a stick) is part of (that) stick
�the surface (of a lake) is part of (that) lake
�the side (of a building) is part of (that) building
Semantic Relations:Can Be Heterogeneous
RANLP: Hissar, Bulgaria: September 14, 2011
6Nakov & Kozareva: Combining Relational and Attributional Similarity ...
� Thus, often handled using instance-based classifiers
� We need to measure SIMILARITY8
Semantic Relations:Can Be Heterogeneous
RANLP: Hissar, Bulgaria: September 14, 2011
7Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Two Kinds of SimilarityTurney (2006)
Rel1(X1,Y1) vs. Rel2(X2,Y2)
� Attributional Similarity
�Correspondence between attributes
�X1::X2 (mason :: carpenter)
�Y1::Y2 (stone :: wood)
� Relational Similarity
�Correspondence between relations
�Rel1 :: Rel2 (mason:stone :: carpenter:wood)
mason:stone
(a) teacher:chalk
(b) carpenter:wood(c) soldier:gun
(d) photograph:camera
(e) book:word
RANLP: Hissar, Bulgaria: September 14, 2011
8Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Measuring Relational SimilarityTurney (2006)
� Attributional similarity can be used to measure relational
similarity, e.g.
Fine for near analogies:
mason:stone
(a) teacher:chalk
(b) carpenter:wood(c) soldier:gun
(d) photograph:camera
(e) book:word
Bad for far analogies:
traffic:street
(a) ship:gangplank
(b) crop:harvest
(c) car:garage
(d) pedestrians:feet
(e) water:riverbed
RANLP: Hissar, Bulgaria: September 14, 2011
9Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Semantic Relation Extraction
Going back to semantic relations8
� Similar split between two lines of research:
� relational: rely on patterns that can connect the arguments (Hearst,92;Turney,05; Turney&Littman,05; Turney,06; Kim&Baldwin,06; Pantel&Pennacchiotti,06; O’Seaghdha&Copestake,07; Davidov et al.07; Davidov&Rappoport,08; Nakov& Hearst,08; Katrenko et al.,10)
� attributional: generalize the arguments using a lexical hierarchy (Rosario&Hearst,01; Rosario et al.,02; Girju et al.,05; Kim&Baldwin,07; O’Seaghdha,09)
RANLP: Hissar, Bulgaria: September 14, 2011
10Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Pattern-based Relation Extraction
� Pattern learning (relational)
� for context-dependent, episodic relations?
� e.g., CAUSE-EFFECT: My Friday’s exam causes me anxiety.
� Argument generalization (attributional)
� for context-independent, permanent relations?
� e.g., PART-WHOLE: door-car
� can benefit from pre-existing resources like WordNet
� BUT limited coverage, need for WSD, etc.
RANLP: Hissar, Bulgaria: September 14, 2011
11Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Our Objectives
� Combine relational and attributional similarity
� and study the relative importance for different relations
� Use no pre-existing lexical resources
� in the real-world, these resources have limited coverage
� Human-readable, explicit semantic representation
� ideallyC
12Nakov & Kozareva: Combining Relational and Attributional Similarity ...
RANLP: Hissar, Bulgaria: September 14, 2011
Method
RANLP: Hissar, Bulgaria: September 14, 2011
13Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Our Web-based Approach at a Glance
• Learn from the Web
- paraphrasing verbs, prepositions, coordinating conjunctions
- hypernyms and co-hyponyms for each argument
• Compute vector similarity for kNN
malaria mosquito
carried by
transmitted via
spreadscauses
insectdisease fly
RANLP: Hissar, Bulgaria: September 14, 2011
14Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Learning Relational Patterns
� For “noun1 noun2”, query:
"noun1 THAT? * noun2"
"noun2 THAT? * noun1"
� Extract:
� V: verbs
� P: prepositions
� C: coordinating conjunctions
committee -- member
RANLP: Hissar, Bulgaria: September 14, 2011
15Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Learning Relational Patterns (cont.)
coffee -- guy
RANLP: Hissar, Bulgaria: September 14, 2011
16Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Sample of harvested patterns for
coffee -- guy
Learned Relational Patterns
RANLP: Hissar, Bulgaria: September 14, 2011
17Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Generalizing Arguments
• Doubly-anchored pattern has anchoring through
conjunctions or terms
- “ relation <seed> and * ”
- “ <seed> and * relation * ”
- “ * relation * and <seed> ”
• can mine both hypernyms and co-hyponyms
• achieves higher accuracy than (Etzioni,05; Pasca,04)
• easy to implement
“ making it vulnerable to predators such as jaguar and puma ”
“ big cats such as jaguar and puma. Similar animal
”“ car brands such as jaguar and land rover were known to be “
RANLP: Hissar, Bulgaria: September 14, 2011
� Instantiate the noun arguments of each sentence with the
doubly-anchored lexico-syntactic pattern
� Submit each pattern as a query to Yahoo! Boss
� Harvested hypernyms and co-hyponyms
� Build a hypernym, co-hyponym vector
18Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Argument Harvesting Procedure
coffee andsuch as
beveragedrink
food
teachocolate
cocoaproduct
RANLP: Hissar, Bulgaria: September 14, 2011
19Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Example: Learned (Co-)Hypernyms
“ * such as coffee and* ”
“* such as guy and * ”
20Nakov & Kozareva: Combining Relational and Attributional Similarity ...
RANLP: Hissar, Bulgaria: September 14, 2011
Dataset
RANLP: Hissar, Bulgaria: September 14, 2011
21Nakov & Kozareva: Combining Relational and Attributional Similarity ...
SemEval-1 task 4:
Classification of Semantic Relations between Nominals
Follow-up: SemEval-2 task 8
Multi-Way Classification of Semantic Relations Between Pairs of Nominals
Dataset
RANLP: Hissar, Bulgaria: September 14, 2011
Dataset: SemEval-1 Task#4
22Nakov & Kozareva: Combining Relational and Attributional Similarity ...
RANLP: Hissar, Bulgaria: September 14, 2011
23Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Experiments and Evaluation
RANLP: Hissar, Bulgaria: September 14, 2011
24Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Classifier, Weights, Similarity
� Classifier: 1NN
� Weighting functions
�Frequencies
�TF.IDF
�TF.IDF w/ add-one smoothing for IDF
� Similarity Measures
�cosine
�Dice
�Lin’s measure
RANLP: Hissar, Bulgaria: September 14, 2011
25Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Features
� Relational
� verbs
� prepositions
� coordinating conjunctions
� Attributional
� nouns (arguments, attributes)
� hypernyms
� co-hyponyms
RANLP: Hissar, Bulgaria: September 14, 2011
26Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Combining Relational and Attributional Similarity
� Linear combination for relational and attributional similarity
� Sm: argument 1
� Sh: argument 2
� Sr: the relation
� Weights tuned on the training set with cross-validation
Also tried:
(a) Whyp = 1
(b) Wcoh = 1
RANLP: Hissar, Bulgaria: September 14, 2011
Results: Overall Micro-averaged
� Combining attributional and relational similarity outperforms:
� all baselines by 0.5-19.5% absolute
(statistically significant in 15 out of 21 cases)
� the best system at SemEval-1 Task 4 (66.0%): by 5.3% absolute
� the state-of-the-art (Davidov&Rapport, 2008), who had 70.1%: by 1.2%
When using manually annotated WordNet senses was not allowed.
27Nakov & Kozareva: Combining Relational and Attributional Similarity ...
RANLP: Hissar, Bulgaria: September 14, 2011
Results: Relational Similarity
28Nakov & Kozareva: Combining Relational and Attributional Similarity ...
� Instrument-Agency (laser--printer) and Product-Producer(honey--bee) are better characterized by patterns
� For Instrument-Agency
� the head (argument 2) is also important
� hypernyms are more important than co-hyponyms
RANLP: Hissar, Bulgaria: September 14, 2011
Results: Attributional Similarity
29Nakov & Kozareva: Combining Relational and Attributional Similarity ...
� Theme-Tool (copyright--law) and Origin-Entity (olive--oil) are
best characterized by the properties of the arguments
� For Theme-Tool
� the head (argument 2) is most critical
� co-hyponyms are more important than hypernyms
RANLP: Hissar, Bulgaria: September 14, 2011
Results: Both Similarities
30Nakov & Kozareva: Combining Relational and Attributional Similarity ...
� For Cause-Effect (growth -- hormone) both the modifier and the
relation are important
� For Cause-Effect
� the modifier (argument 1) is most critical
� again, co-hyponyms are more important than hypernyms
RANLP: Hissar, Bulgaria: September 14, 2011
Results: Comparing to WordNet
� For some relations, the results outpeform those at SemEval-1
task 4, even when manual WordNet senses are allowed
� Origin-Entity: 77.8% vs. 72.8% (stat. significant)
� Theme-Tool: 74.7% vs. 74.6%
31Nakov & Kozareva: Combining Relational and Attributional Similarity ...
RANLP: Hissar, Bulgaria: September 14, 2011
32Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Conclusion
RANLP: Hissar, Bulgaria: September 14, 2011
33Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Conclusion
� Presented:
� Web-based approach to relation extraction
� Uses no pre-existing lexical resources
� Based on human-readable, explicit semantic representation
� paraphrases: verbs, prepositions, coordinating conjunctions
� hypernyms and co-hyponyms
� Studied the combination of relational and attributional similarity
� for various relations
� Achieved
� sizable improvements over a strong baseline
� small improvements over the state-of-the-art
RANLP: Hissar, Bulgaria: September 14, 2011
34Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Future Work
� We plan to
� use other relation inventories, e.g., SemEval-2 task 8
� model sentence context
� jointly generalize arguments and paraphrases
� use bootstrapping to mine more examples
� try explicit paraphrases beyond verbs/preps/conj.
RANLP: Hissar, Bulgaria: September 14, 2011
35Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Paraphrasing Semantics: Beyond Verbs
“onion tears”
tears from onions
tears due to cutting onion
tears induced when cutting onions
tears that onions induce
tears that come from chopping onions
tears that sometimes flow when onions are chopped
tears that raw onions give you
SemEval-2013 task 24C. Butnariu, I. Hendrickx, S. N. Kim, Z. Kozareva, P. Nakov, D. Ó Séaghdha, S. Szpakowicz, T. Veale
RANLP: Hissar, Bulgaria: September 14, 2011
36Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Future Work
� We plan to
� use other relation inventories, e.g., SemEval-2 task 8
� model sentence context
� jointly generalize arguments and paraphrases
� try explicit paraphrases beyond verbs/preps/conj.
� use bootstrapping to mine more examples
Thank you!
RANLP: Hissar, Bulgaria: September 14, 2011
37Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Detailed Results (1)
RANLP: Hissar, Bulgaria: September 14, 2011
38Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Detailed Results (2)
RANLP: Hissar, Bulgaria: September 14, 2011
39Nakov & Kozareva: Combining Relational and Attributional Similarity ...
Detailed Results (3)