Incremental Semantic Role Labeling with Tree Adjoining Grammar

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 301–312, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Incremental Semantic Role Labeling with Tree Adjoining Grammar Ioannis Konstas⋆, Frank Keller⋆, Vera Demberg† and Mirella Lapata⋆ ⋆: Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh {ikonstas,keller,mlap}@inf.ed.ac.uk †: Cluster of Excellence Multimodal Computing and Interaction, Saarland University vera@coli.uni-saarland.de Abstract We introduce the task of incremental se- mantic role labeling (iSRL), in which se- mantic roles are assigned to incomplete input (sentence prefixes). iSRL is the semantic equivalent of incremental pars- ing, and is useful for language model- ing, sentence completion, machine trans- lation, and psycholinguistic modeling. We propose an iSRL system that combines an incremental TAG parser with a seman- tically enriched lexicon, a role propaga- tion algorithm, and a cascade of classi- fiers. Our approach achieves an SRL F- score of 78.38% on the standard CoNLL 2009 dataset. It substantially outper- forms a strong baseline that combines gold-standard syntactic dependencies with heuristic role assignment, as well as a baseline based on Nivre’s incremental de- pendency parser. 1 Introduction Humans are able to assign semantic roles such as agent, patient, and theme to an incoming sentence before it is complete, i.e., they incrementally build up a partial semantic representation of a sentence prefix. As an example, consider: (1) The athlete realized [her goals]PATIENT/THEME were out of reach. When reaching the noun phrase her goals, the hu- man language processor is faced with a semantic role ambiguity: her goals can either be the PA- TIENT of the verb realize, or it can be the THEME of a subsequent verb that has not been encoun- tered yet. Experimental evidence shows that the human language processor initially prefers the PA- TIENT role, but switches its preference to the theme role when it reaches the subordinate verb were. Such semantic garden paths occur because human language processing occurs word-by-word, and are well attested in the psycholinguistic litera- ture (e.g., Pickering et al., 2000). Computational systems for performing seman- tic role labeling (SRL), on the other hand, proceed non-incrementally. They require the whole sen- tence (typically together with its complete syntac- tic structure) as input and assign all semantic roles at once. The reason for this is that most features used by current SRL systems are defined globally, and cannot be computed on sentence prefixes. In this paper, we propose incremental SRL (iSRL) as a new computational task that mimics human semantic role assignment. The aim of an iSRL system is to determine semantic roles while the input unfolds: given a sentence prefix and its partial syntactic structure (typically generated by an incremental parser), we need to (a) identify which words in the input participate in the seman- tic roles as arguments and predicates (the task of role identification), and (b) assign correct seman- tic labels to these predicate/argument pairs (the task of role labeling). Performing these two tasks incrementally is substantially harder than doing it …

Content truncated to 500 words