Learning to Jointly Predict Ellipsis and Comparison Structures

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 62–74, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Learning to Jointly Predict Ellipsis and Comparison Structures Omid Bakhshandeh1, Alexis Wellwood2, James Allen1,3 1 University of Rochester, 2 Northwestern University, 3 Florida Institute for Human and Machine Cognition {omidb,james}@cs.rochester.edu, wellwood@northwestern.edu Abstract Domain-independent meaning representa- tion of text has received a renewed in- terest in the NLP community. Compar- ison plays a crucial role in shaping ob- jective and subjective opinion and mea- surement in natural language, and is of- ten expressed in complex constructions in- cluding ellipsis. In this paper, we intro- duce a novel framework for jointly captur- ing the semantic structure of comparison and ellipsis constructions. Our framework models ellipsis and comparison as inter- connected predicate-argument structures, which enables automatic ellipsis resolu- tion. We show that a structured prediction model trained on our dataset of 2,800 gold annotated review sentences yields promis- ing results. Together with this paper we release the dataset and an annotation tool which enables two-stage expert annotation on top of tree structures. 1 Introduction Representing the underlying meaning of text has been a long-standing topic of interest in com- putational linguistics. Recently there has been a renewed interest in representation of meaning for various tasks such as semantic parsing, where the task is to map a natural language sentence into its corresponding formal meaning representa- tion (Zelle and Mooney, 1996; Berant and Liang, 2014). Open-domain and broad-coverage seman- tic representation of text (Banarescu et al., 2013; Bos, 2008; Allen et al., 2008) is crucial for many language understanding tasks such as read- ing comprehension tests and question answering. With the rise of continuous-space models there is even more interest in capturing deeper generic se- mantics of text as opposed to surface word repre- sentations. One of the most common ways for expressing evaluative sentiment towards different entities is using comparison. A simple natural language ex- ample of comparison is Their pizza is the best. Capturing the underlying meaning of comparison structures, as opposed to their surface wording, is required for accurate evaluation of qualities and quantities. For instance, given a more complex comparison example, The pizza was great, but it was not as awesome as the sandwich, the state-of- the-art sentiment analysis system (Manning et al., 2014) assigns an overall ‘neutral’ sentiment value, which clearly lacks deeper understanding of the comparison happening in the sentence. Consider the generic meaning representation depicted in in Figure 1 according to frame seman- tic parsing 1 (Das et al., 2014) for the following sentence: (1) My Mazda drove faster than his Hyundai. It is evident that this meaning representation does not fully capture how the semantics of the adjec- tive fast relates to the driving event, and what it actually means for a car to drive faster than an- other car. More importantly, there is an ellipsis in this sentence, the resolution of which results in complete understood reading of My Mazda drove …

Content truncated to 500 words