Semantic role labeling (SRL) is a task in natural language processing consisting of the detection of the semantic arguments associated with the predicate or verb of a sentence and their classification into their specific roles.[1] It is considered a shallow semantic parsing task.

A successful execution of SRL tranform a sentence into a set of propositions. One good way to start is to ask ourselves what kinds of propositions there are and what set of propositions are enough to transcribe human languages. Unfortunately, there isn't a definite answer for those questions although there are some candidates such as case theory and semantic frame(anything else?).

Most researches local identification and classification followed by global inference however integrated and incremental approaches have been developed.

Theories Edit

Two broad families exist:

  1. Syntax-based approach : explaining the varied expression of verb arguments within syntactic positions : Levin (1993) verb classes =⇒ VerbNet (Kipper et al., 2000) =⇒ PropBank (Palmer et al., 2005)[2]: Focused on verbs ( lately nominal bank (NomBank) is used together with PropBank in many semantic tasks )
  2. Situation-based approach (a word activates/invokes a frame of semantic knowledge that relates linguistic semantics to encyclopedic knowledge) : Frame semantics (Fillmore, 1976) =⇒ FrameNet (Fillmore et al., 2003)[3]: Words with other POS can invoke frames too (e.g., nouns, adjectives)

Local approaches Edit

Local stage in Björkelund (2009)[4]

Also uses the term "local": Toutanova et al. 2005

According to Zapirain et al. (2013)[5], this is mostly syntactic:

"... typically perform SRL in two sequential steps: argument identification and argument classification. Whereas the former is mostly a syntactic recognition task, the latter usually requires semantic knowledge to be taken into account"

Predicate identification Edit

Pruning Edit

Pruning: remove candidates that are clearly not argument of a given predicate to save training time and, more importantly, improve performance (Punyakanok et al, 2008)[6] (however, mate tools (Björkelund et al., 2009)[7] doesn't employ this step).

Pruning algorithm for constituent syntactic parse tree (Xue & Palmer, 2004)[8]:

  • Step 1: Designate the predicate as the current node and collect its sisters (constituents at- tached at the same level as the predicate) unless its sisters are coordinated with the predicate. If a sister is a PP, also collect its immediate children.
  • Step 2: Reset the current node to its parent and repeat Step 1 till it reaches the top level node.

Argument identification Edit

Argument classification Edit

Global (joint) scoring Edit

Reranking Edit

"The early work of Gildea and Jurafsky (2002)[9] produced a set of possible sequences of labels for the entire sentence by combining the most likely few labels for each constituent. The probabilities produced by the classifiers for individual constituents were combined with a probability for the (unordered) set of roles appearing in the entire sentence, conditioned on the predicate. This reranking step improves performance, but because of the use of frequency-based probabilities, the reranking suffers from the same inability to exploit larger numbers of features as the lattice backoff used for individual role classification."[10]

Toutanova et al. 2005: log-linear reranking model applied to top N solutions.

$ P(L|t,v) = \frac{e^{<\Phi(t,v,L), W>}}{\sum_{j=1}^N e^{<\Phi(t,v,L_j), W>}} $

Re-ranking of several candidate solutions (Toutanova et al., 2008) (+learning +dependencies –search)

Viterbi search Edit

Integer linear programming Edit

Combine local predictions through ILP to find the best solution according to structural and linguistic constraints (Koomen et al., 2005; Punyakanok et al., 2008) (–learning +dependencies +search)

Integrated (global) approaches Edit

An early research was done by Zhao et al. (2009)[11].

James Henderson and Ivan Titov's group put effort on joint, synchronized syntactic-semantic parsing (Henderson et al. 2008[12]; Titov et al. 2009[13]; Henderson et al. 2013[14])

Global search integrating joint scoring: Tree CRFs (Cohn & Blunsom, 2005) (+learning +/–dependencies +/–search)

CRF over tree structure (Cohn & Blunsom, 2005) [15], CRF over sequence (Marquez et al., 2005)[16]

Incremental approaches Edit

Choi and Palmer (2011)[17] devised an elegant transition-based model but didn't receive much attention.

Konstas et al. (2014) thinks that incremental SRL is intrinsically harder and should be viewed as a separate task. They rely on an intricate syntactic parser and build a complicated SRL system... Their evaluation is not compatible with standard evaluation.

Uncategorized Edit

  • Back-off lattice-based relative frequency models ([Gildea&Jurafsky 02], [Gildea& Palmer 02])
  • Decision trees ([Surdeanu et al. 03])
  • Support Vector Machines ([Pradhan et al. 04] [Moschitti et al. 07])
  • Log-linear models ([Xue&Palmer 04][Toutanova et al. 05])
  • SNoW ([Punyakanok et al. 04,05])
  • AdaBoost, TBL, IBL

Features Edit

Main article: List of features for semantic role labeling

Various features were proposed for SRL which can be divided into broad categories:

  • Lexical ~: word form, lemma
  • Morphosyntactical ~: part-of-speech
  • Positional ~: distance
  • Syntactic ~: dependency label, valency, constituent/dependency paths
  • Semantic ~: role, frame

Evaluation Edit

Metrics Edit

Some papers report P, R, F1 on argument identification and argument classification (but not predicate identification and disambiguation).[18] In CoNLL-2005, "for an argument to be correctly recognized, the words spanning the argument as well as its semantic role have to be correct." (Carreras & Màrques 2005)[19]

"F1 score on the SemEval 2007 task of collectively identifying frame-evoking targets, a disambiguated frame for each target, and the set of role-labeled arguments for each frame."[20]

See also: Dependency-based SRL evaluation

Data and evaluation campaigns Edit

Constituent-based Edit

  1. CONLL 2004 and 2005

Dependency-based Edit

  1. CONLL 2008
  2. CONLL 2009

Coverage gap Edit

Available lexical resources represent only a small portion of English. Palmer et al. (2010)[21] showed that the accuracy of a straight supervised system has an upper bound of approximately 46.8% on full texts. Semi-supervised, unsupervised and crosslingual approaches have been proposed to ease this problem.[22]

Applications Edit

"Shallow semantic analysis based on FrameNet data has been recently utilized across various natural language processing applications with success. These include the generation of meeting summaries (Kleinbauer, 2012), the prediction of stock price movement using (Xie et al., 2013), inducing slots for domain-specific dialog systems (Chen et al., 2013), stance classification in debates (Hasan and Ng, 2013), modeling the clarity of student essays (Persing and Ng, 2013) to name a few. There is strong potential in using frame-semantic structures in other applications such as question answering and machine translation, as demonstrated by prior work using PropBank-style SRL annotations (Shen and Lapata, 2007; Liu and Gildea, 2010)."[20]

See also Edit

External links Edit

References Edit

  1. Wikipedia contributors, "Semantic role labeling," Wikipedia, The Free Encyclopedia, (accessed September 10, 2014).
  2. Palmer, M., Gildea, D., & Kingsbury, P. (2005). The proposition bank: An annotated corpus of semantic roles. Computational Linguistics, 31(1), 71–106.
  3. Fillmore, C. J., Johnson, C. R., & Petruck, M. R. L. (2003). Background to framenet. International Journal of Lexicography, 16(3), 235–250.
  4. Björkelund, A., Hafdell, L., & Nugues, P. (2009). Multilingual Semantic Role Labeling. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL 2009): Shared Task (pp. 43–48). Association for Computational Linguistics.
  5. Zapirain, B., Agirre, E., Màrquez, L., & Surdeanu, M. (2013). Selectional preferences for semantic role classification. Computational Linguistics, 39(3), 631-663.
  6. Punyakanok, V., Roth, D., & Yih, W. (2008). The Importance of Syntactic Parsing and Inference in Semantic Role Labeling. Computational Linguistics, 34(2), 257–287. doi:10.1162/coli.2008.34.2.257
  7. Björkelund, A., Hafdell, L., & Nugues, P. (2009). Multilingual Semantic Role Labeling. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL 2009): Shared Task (pp. 43–48). Association for Computational Linguistics.
  8. Xue, N., & Palmer, M. (2004). Calibrating Features for Semantic Role Labeling. Emnlp, 88–94. Retrieved from
  9. Gildea, D., & Jurafsky, D. (2002). Automatic labeling of semantic roles. Computational linguistics, 28(3), 245-288.
  10. Palmer, M., Gildea, D., & Xue, N. (2010). Semantic Role Labeling. Synthesis Lectures on Human Language Technologies, 3(1), page 44. doi:10.2200/S00239ED1V01Y200912HLT006
  11. Zhao, H., Chen, W., & Kit, C. (2009, August). Semantic dependency parsing of NomBank and PropBank: An efficient integrated approach via a large-scale feature selection. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1-Volume 1 (pp. 30-39). Association for Computational Linguistics.
  12. Henderson, J., Merlo, P., Musillo, G., & Titov, I. (2008, August). A latent variable model of synchronous parsing for syntactic and semantic dependencies. In Proceedings of the Twelfth Conference on Computational Natural Language Learning (pp. 178-182). Association for Computational Linguistics.
  13. Titov, I., Henderson, J., Merlo, P., & Musillo, G. (2009, July). Online Graph Planarisation for Synchronous Parsing of Semantic and Syntactic Dependencies. In IJCAI (pp. 1562-1567).
  14. Henderson, J., Merlo, P., Titov, I., & Musillo, G. (2013). Multilingual joint parsing of syntactic and semantic dependencies with a latent variable model. Computational Linguistics, 39(4), 949-998.
  15. Cohn, T., & Blunsom, P. (2005). Semantic Role Labelling with Tree Conditional Random Fields. In Proceedings of the Ninth Conference on Computational Natural Language Learning (CoNLL-2005) (pp. 169–172). Association for Computational Linguistics.
  16. Màrquez, L., Comas, P., Giménez, J., & Català, N. (2005). Semantic Role Labeling as Sequential Tagging. In Proceedings of the Ninth Conference on Computational Natural Language Learning (CoNLL-2005) (pp. 193–196). Association for Computational Linguistics.
  17. Choi, J. D., & Palmer, M. (2011, June). Transition-based semantic role labeling using predicate argument clustering. In Proceedings of the ACL 2011 Workshop on Relational Models of Semantics (pp. 37-45). Association for Computational Linguistics.
  18. Choi, J. D., & Palmer, M. (2011). Transition-based Semantic Role Labeling Using Predicate Argument Clustering. In Proceedings of the ACL 2011 Workshop on Relational Models of Semantics (pp. 37–45). Stroudsburg, PA, USA: Association for Computational Linguistics.
  19. Carreras, X., & Màrques, L. (2005). Introduction to the CoNLL-2005 Shared Task: Semantic Role Labeling. Proceedings of CoNLL-2005, 152–164.
  20. 20.0 20.1 Das, D. (2014). Statistical Models for Frame-Semantic Parsing. Proceedings of Frame Semantics in NLP: A Workshop in Honor of Chuck Fillmore (1929-2014), (2007), 26–29. Retrieved from
  21. Alexis Palmer and Caroline Sporleder. 2010. Evaluating FrameNet-style semantic parsing: the role of coverage gaps in FrameNet. In Proceedings of the 23rd International Conference on Computational Linguistics: Posters (COLING '10). Association for Computational Linguistics, Stroudsburg, PA, USA, 928-936.
  22. Ivan Titov. Semantic Role Labeling Tutorial: Part 3 - Semi- , unsupervised and cross-lingual approaches. NAACL 2013
Community content is available under CC-BY-SA unless otherwise noted.