1. Semantic Role Labeling Abdul-Lateef Yussiff 20-01-2011

2. Table of Contents Introduction & motivation Definition Semantic Role Application Areas Available corpus SRL Tasks

3. Introduction & Motivation Information extraction and dialogue understanding systems are usually based on domain-specific frame-and-slot templates For natural language understanding tasks to proceed beyond domain dependent system, semantic understanding systems is needed.

4. Introduction & Motivation Since 1999 CoNLL has organized several shared tasks exploring many NLP problems. This research proposal focuses on the problem studied in the CoNLL 2004 and 2005 shared tasks – Namely Semantic Role labeling (SRL).

5. Definition A semantic role is the underlying relationships that a participant has with the main verb in a clause. [www.sil.org] SRL is detecting basic event structures such as who did what to whom, when and where [IE view] SRL is identifying the arguments of a given predicate and assign them semantic labels describing the roles they play in the predicate [ Computational linguistic view ]

6. CoNLL Shared tasks SRL is the relationship that a syntactic constituent has with a predicate. –Core arguments: Agent, Patient, Instrument etc. –Adjunctive arguments: Locative, Temporal, Manner, Cause, etc. [http://www.lsi.uoc.edu/~srlconll]

7. Example Mary hit John John was hit by Mary (ROOT (S (NP (NNP John)) (VP (VBD was) (VP (VBN hit) (PP (IN by) (NP (NNP Mary))))) (..))) S S VP NP NNP VBD NNP Mary hit John (S (NP (NNP Mary)) (VP (VBD hit) (NP (NNP John))) (..)))

8. Fundamental tasks in NLP Semantic Role labeling is an essential step towards the final goal of natural language understanding. Several fundamental tasks in NLP are based on recognizing phrases or constituents. – Noun phrases, Chunking, clause identification, name entity identification Syntactic analysis, etc.

9. Semantic Role A variety of semantic role labels have been proposed, common ones are: – Agent: Actor of an action – Patient: Entity affected by the action – Instrument: Tool used in performing action. – Beneficiary: Entity for whom action is performed – Source: Origin of the affected entity – Destination: Destination of the affected entity

10. Application Areas Question Answering – “Who” questions usually use Agents – “What” question usually use Patients – “How” and “with what” questions usually use Instruments – “Where” questions frequently use Sources and Destinations. – “For whom” questions usually use Beneficiaries – “To whom” questions usually use Destinations Machine Translation Generation – Semantic roles are usually expressed using particular, distinct syntactic constructions in different languages.

11. Syntactic cues Semantic role is indicated by a particular syntactic position (e.g. object of a particular preposition). – Agent: subject – Patient: direct object – Instrument: object of “with” PP – Beneficiary: object of “for” PP – Source: object of “from” PP – Destination: object of “to” PP

12. SRL with Parse Trees Assume that a syntactic parse is available. For each predicate (verb), label each node in the parse tree as either not-a-role or one of the possible semantic roles S NP VP NP PP The Prep NP with the V NP bit a big dog girl boy Det A N ε Adj A ε Det A N ε

13. SRL Corpus FrameNet: Developed at University of California, Berkeley. Based on notion of frames VerbNet PropBank: Developed at University of Pennsylvania. Based on Treebank – Annotated over 1M words of WSJ text – 43, 594 sentences – 3,324 unique roles – 262,281 role assignments

14. And they believe the Big Board, under Mr. Phelan, has abandoned their interest. the Big Board under Mr. Phelan abandoned their interest John abandoned his pursuit of an Olympic gold medal as a waste of time. John abandoned his pursuit of an Olympic gold medal as a waste of time. And they believe the Big Board, under Mr. Phelan, has abandoned their interest. the Big Board under Mr. Phelan abandoned their interest John abandoned his pursuit of an Olympic gold medal as a waste of time. John abandoned his pursuit of an Olympic gold medal as a waste of time. Example of propBank frameset Roleset indicates different senses of the verb

15. Example of Corpus

16. SRL TASKS I am proposing a two tasks for SRL –Arguments identification: Parse the sentence into syntactic constituents Binary classification of predicates (verb) Use a binary classification to determine whether a constituents span a boundary or not –Argument classification into Semantic roles. Assign appropriate semantic role to the identified semantic argument This is a clearly a multi-class classification

17. System Architecture Raw Sentence Sentence labeled with Semantic Roles Parser POS & Tree Argument Identification Role Classification input Output

18. Baseline Features for SRL [ Gildea and Jurafsky, 2002 ] Phrase type: the syntactic label of the candidate role (e.g. NP) Parse tree path: the path in the parse tree between the predicate and the candidate role Position of the phrase type relative to the predicate voice of the predicate (active or passive) Head word of the phrase type

19. Challenges Bell, based in Los Angeles, makes and distributes electronic, computer and building products. [Bell,,, based, in, Los, Angeles,,] [Bell] [Los, Angeles] [electronic,,, computer, and, building, products] Iterator iterator = t.iterator(); while (iterator.hasNext()){ Tree node = iterator.next(); if(node.value().equalsIgnoreCase("NP")) System.out.println(node.getLeaves()); }//end of while iterator Iterator iterator = t.iterator(); while (iterator.hasNext()){ Tree node = iterator.next(); if(node.value().equalsIgnoreCase("NP")) System.out.println(node.getLeaves()); }//end of while iterator

20. (ROOT (S (NP (NP (NNP Bell)) (,,) (VP (VBN based) (PP (IN in) (NP (NNP Los) (NNP Angeles)))) (,,)) (VP (VBZ makes) (CC and) (VBZ distributes) (NP (UCP (JJ electronic) (,,) (NN computer) (CC and) (NN building)) (NNS products))) (..)))

21. Baseline Research Work Gildea and Jurafsky (2002) –Parse the training corpus using Collin’s parser –Match frame elements into constituents –Extract features from the parse tree –Train probabilistic model on the model –82% accuracy for presegmented constituents –65% precision and 61% recall for segmentation and identification

22. Machine learning approaches Maximum entropy SVM Enhanced perceptron Decision Trees Nearest neighbor CRF

23. Top Ten Systems in the CoNLL 2005

24. Schedules Argument Identification3 weeks Extracting Key features3 weeks Argument classification3 weeks

25. References Gildea, Daniel and Daniel Jurafsky. 2002. Automatic labeling of semantic roles. Computational Linguistics 28(3):245–288 Gildea, Daniel and Palmer, Martha. The necessity of parsing for predicate argument recognition. Proceeding of the 40 th Annual Meeting of ACL 2002. pp239.246 Hao Xiong, Haitao Mi, Yang Liu and Qun Liu. Forest-based Semantic Role Labeling. The Association of Advancement of AI,2010. Klenner Manfred, Extracting predicate structures from parse Trees Hacioglu Kadri. Semantic role labeling Using dependency Trees Marquez L, Carreras X., Litkowski K.C., Stevenson Suzanne. Semantic Role Labeling: An Introduction to special issue. Association for Computational Linguistics 2008. Surdeanu M., Marquez L., Carreras X. Comas P. Combination Strategies for Semantic Role Labeling. Journal of AI Research 29:105-151

26. Question ? THANK YOU FOR YOUR TIME