Dynamic Knowledge-Base Alignment for Coreference Resolution

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 153–162, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Dynamic Knowledge-Base Alignment for Coreference Resolution Jiaping Zheng Luke Vilnis Sameer Singh Jinho D. Choi Andrew McCallum School of Computer Science University of Massachusetts Amherst MA 01003 {jzheng,luke,sameer,jdchoi,mccallum}@cs.umass.edu Abstract Coreference resolution systems can benefit greatly from inclusion of global context, and a number of recent approaches have demonstrated improvements when precom- puting an alignment to external knowledge sources. However, since alignment itself is a challenging task and is often noisy, ex- isting systems either align conservatively, resulting in very few links, or combine the attributes of multiple candidates, leading to a conflation of entities. Our approach instead performs joint inference between within-document coreference and entity linking, maintaining ranked lists of candi- date entities that are dynamically merged and reranked during inference. Further, we incorporate a large set of surface string vari- ations for each entity by using anchor texts from the web that link to the entity. These forms of global context enables our system to improve classifier-based coreference by 1.09 B3 F1 points, and improve over the previous state-of-art by 0.41 points, thus introducing a new state-of-art result on the ACE 2004 data. 1 Introduction Coreference resolution is the task of identifying sets of noun phrase mentions from a document that refer to the same real-world entities. For ex- ample, in the following excerpt: “The Chicago suburb of Arlington Heights is the first stop for ⟨George W. Bush⟩1 today. ⟨The Texas governor⟩2 stops in ⟨Gore’s home state⟩3 of ⟨Tennessee⟩4 this afternoon. . . ”, (m1, m2) and (m3, m4) define the coreferent pairs. Coreference resolution forms an important component for natural language process- ing and information extraction pipelines due to its utility in relation extraction, cross-document coref- erence, text summarization, and question answer- ing. The task of coreference is challenging for automated systems as the local information con- tained in the document is often not enough to accu- rately disambiguate mentions, for example, corefer- encing (m1, m2) requires identifying that George W. Bush (m1) is the governor of Texas (m2), and similarly for (m3, m4). External knowledge-bases such as FrameNet (Baker et al., 1998), Wikipedia, Yago (Suchanek et al., 2007), and Freebase (Bol- lacker et al., 2008), can be used to provide global context, and there is a strong need for coreference resolution systems to accurately use such sources for disambiguation. Incorporating external knowledge bases into coreference has been the subject of active recent research. Ponzetto and Strube (2006) and Ratinov and Roth (2012) precompute a fixed alignment of the mentions to the knowledge base entities. The attributes of these entities are used during corefer- ence by incorporating them in the mention features. Since alignment of mentions to the external enti- ties is itself a difficult task, these systems favor high-precision linking. Unfortunately, this results in fewer alignments, and improvements are only shown on mentions that are easier to align and core- fer (such as the non-transcript …

Content truncated to 500 words