Identifying Causal Relations Using Parallel Wikipedia Articles

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1424–1433, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Identifying Causal Relations Using Parallel Wikipedia Articles Christopher Hidey Department of Computer Science Columbia University New York, NY 10027 chidey@cs.columbia.edu Kathleen McKeown Department of Computer Science Columbia University New York, NY 10027 kathy@cs.columbia.edu Abstract The automatic detection of causal relation- ships in text is important for natural lan- guage understanding. This task has proven to be difficult, however, due to the need for world knowledge and inference. We fo- cus on a sub-task of this problem where an open class set of linguistic markers can provide clues towards understanding causality. Unlike the explicit markers, a closed class, these markers vary signifi- cantly in their linguistic forms. We lever- age parallel Wikipedia corpora to identify new markers that are variations on known causal phrases, creating a training set via distant supervision. We also train a causal classifier using features from the open class markers and semantic features pro- viding contextual information. The results show that our features provide an 11.05 point absolute increase over the baseline on the task of identifying causality in text. 1 Introduction The automatic detection of causal relationships in text is an important but difficult problem. The identification of causality is useful for the under- standing and description of events. Causal in- ference may also aid upstream applications such as question answering and text summarization. Knowledge of causal relationships can improve performance in question answering for “why” questions. Summarization of event descriptions can be improved by selecting causally motivated sentences. However, causality is frequently ex- pressed implicitly, which requires world knowl- edge and inference. Even when causality is ex- plicit, there is a wide variety in how it is expressed. Causality is one type of relation in the Penn Dis- course Tree Bank (PDTB) (Prasad et al, 2008). In general, discourse relations indicate how two text spans are logically connected. In PDTB the- ory, these discourse relations can be marked ex- plicitly or conveyed implicitly. In the PDTB, there are 102 known explicit discourse markers such as “and”, “but”, “after”, “in contrast”, or “in addi- tion”. Of these, 28 explicitly mark causal relations (e.g., “because”, “as a result”, “consequently”). In addition to explicit markers, PDTB re- searchers recognize the existence of an open class of markers, which they call AltLex. There is a tremendous amount of variation in how AltLexes are expressed and so the set of AltLexes is ar- guably infinite in size. In the PDTB, non-causal AltLexes include “That compares with” and “In any event.” Causal AltLexes include “This may help explain why” and “This activity produced.” Discourse relations with explicit discourse markers can be identified with high precision (Pitler and Nenkova, 2009) but they are also rela- tively rare. Implicit relations are much more com- mon but very difficult to identify. AltLexes fall in the middle; their linguistic variety makes them difficult to identify but their presence improves the identification …

Content truncated to 500 words