What a Parser Can Learn from a Semantic Role Labeler and Vice Versa

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 736–744, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics What a Parser can Learn from a Semantic Role Labeler and Vice Versa Stephen A. Boxwell, Dennis N. Mehay, Chris Brew The Ohio State University {boxwell, mehay, cbrew}@ling.ohio-state.edu Abstract In many NLP systems, there is a unidirectional flow of information in which a parser supplies input to a semantic role labeler. In this paper, we build a sys- tem that allows information to flow in both direc- tions. We make use of semantic role predictions in choosing a single-best parse. This process relies on an averaged perceptron model to distinguish likely semantic roles from erroneous ones. Our system pe- nalizes parses that give rise to low-scoring semantic roles. To explore the consequences of this we per- form two experiments. First, we use a baseline gen- erative model to produce n-best parses, which are then re-ordered by our semantic model. Second, we use a modified version of our semantic role labeler to predict semantic roles at parse time. The perfor- mance of this modified labeler is weaker than that of our best full SRL, because it is restricted to fea- tures that can be computed directly from the parser’s packed chart. For both experiments, the resulting se- mantic predictions are then used to select parses. Fi- nally, we feed the selected parses produced by each experiment to the full version of our semantic role labeler. We find that SRL performance can be im- proved over this baseline by selecting parses with likely semantic roles. 1 Introduction In the semantic role labeling task, words or groups of words are described in terms of their relations to a pred- icate. For example, the sentence Robin admires Leslie has two semantic role-bearing words: Robin is the agent or experiencer of the admire predicate, and Leslie is the patient. These semantic relations are distinct from syntactic relations like subject and object – the proper nouns in the sentence Leslie is admired by Robin have the same semantic relationships as Robin admires Leslie, even though the syntax differs. Although syntax and semantics do not always align with each other, they are correlated. Almost all automatic se- mantic role labeling systems take a syntactic representa- tion of a sentence (taken from an automatic parser or a human annotator), and use the syntactic information to predict semantic roles. When a semantic role labeler pre- dicts an incorrect role, it is often due to an error in the parse tree. Consider the erroneously annotated sentence from the Penn Treebank corpus shown in Figure 1. If a semantic role labeling system relies heavily upon syntac- tic attachment decisions, then it will likely predict that in 1956 describes the time that asbestos was used, rather than when it ceased to be used. Errors of this kind are common in treebanks and in au- tomatic parses. It is telling, though, that while the hand- annotated Penn Treebank …

Content truncated to 500 words