Parsing Paraphrases with Joint Inference

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1223–1233, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Parsing Paraphrases with Joint Inference Do Kook Choe∗ Brown University Providence, RI dc65@cs.brown.edu David McClosky IBM Research Yorktown Heights, NY dmcclosky@us.ibm.com Abstract Treebanks are key resources for develop- ing accurate statistical parsers. However, building treebanks is expensive and time- consuming for humans. For domains re- quiring deep subject matter expertise such as law and medicine, treebanking is even more difficult. To reduce annotation costs for these domains, we develop methods to improve cross-domain parsing inference using paraphrases. Paraphrases are eas- ier to obtain than full syntactic analyses as they do not require deep linguistic knowl- edge, only linguistic fluency. A sentence and its paraphrase may have similar syn- tactic structures, allowing their parses to mutually inform each other. We present several methods to incorporate paraphrase information by jointly parsing a sentence with its paraphrase. These methods are ap- plied to state-of-the-art constituency and dependency parsers and provide signif- icant improvements across multiple do- mains. 1 Introduction Parsing is the task of reconstructing the syntac- tic structure from surface text. Many natural lan- guage processing tasks use parse trees as a basis for deeper analysis. The most effective sources of supervision for training statistical parsers are treebanks. Unfortu- nately, treebanks are expensive, time-consuming to create, and not available for most domains. Compounding the problem, the accuracy of statis- tical parsers degrades as the domain shifts away from the supervised training corpora (Gildea, 2001; Bacchiani et al., 2006; McClosky et al., 2006b; Surdeanu et al., 2008). Furthermore, for ∗Work performed during an IBM internship. domains requiring subject matter experts, e.g., law and medicine, it may not be feasible to produce large scale treebanks since subject matter experts generally don’t have the necessary linguistic back- ground. It is natural to look for resources that are more easily obtained. In this work, we ex- plore using paraphrases. Unlike parse trees, para- phrases can be produced quickly by humans and don’t require extensive linguistic training. While paraphrases are not parse trees, a sentence and its paraphrase may have similar syntactic structures for portions where they can be aligned. We can improve parsers by jointly parsing a sentence with its paraphrase and encouraging cer- tain types of overlaps in their syntactic structures. As a simple example, consider replacing an un- known word in a sentence with a synonym found in the training data. This may help disambiguate the sentence without changing its parse tree. More disruptive forms of paraphrasing (e.g., topicaliza- tion) can also be handled by not requiring strict agreement between the parses. In this paper, we use paraphrases to improve parsing inference within and across domains. We develop methods using dual-decomposition (where the parses of both sentences from a depen- dency parser are encouraged to agree, Section 3.2) and pair-finding (which can be applied to any n- best parser, Section …

Content truncated to 500 words