Using Mined Coreference Chains as a Resource for a Semantic Task

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1447–1452, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Using Mined Coreference Chains as a Resource for a Semantic Task Heike Adel and Hinrich Sch¨utze Center for Information and Language Processing University of Munich Germany heike.adel@cis.lmu.de Abstract We propose to use coreference chains ex- tracted from a large corpus as a resource for semantic tasks. We extract three mil- lion coreference chains and train word embeddings on them. Then, we com- pare these embeddings to word vectors de- rived from raw text data and show that coreference-based word embeddings im- prove F1 on the task of antonym classifi- cation by up to .09. 1 Introduction After more than a decade of work on coreference resolution, coreference resolution systems have reached a certain level of maturity (e.g., Recasens et al. (2010)). While accuracy is far from perfect and many phenomena such as bridging still pose difficult research problems, the quality of the out- put of these systems is high enough to be useful for many applications. In this paper, we propose to run coreference res- olution systems on large corpora, to collect the coreference chains found and to use them as a re- source for solving semantic tasks. This amounts to using mined coreference chains as an automat- ically compiled resource similar to the way cooc- currence statistics, dependency pairs and aligned parallel corpora are used in many applications in NLP. Coreference chains have interesting comple- mentary properties compared to these other re- sources. For example, it is difficult to distinguish true semantic similarity (e.g., “cows” – “cattle”) from mere associational relatedness (e.g., “cows” – “milk”) based on cooccurrence statistics. In con- trast, coreference chains should be able to make that distinction since only “cows” and “cattle” can occur in the same coreference chain, not “cows” and “milk”. As a proof of concept we compile a resource of mined coreference chains from the Gigaword corpus and apply it to the task of identifying antonyms. We induce distributed representations for words based on (i) cooccurrence statistics and (ii) mined coreference chains and show that a com- bination of both outperforms cooccurrence statis- tics on antonym identification. In summary, we make two contributions. First, we propose to use coreference chains mined from large corpora as a resource in NLP and publish the first such resource. Second, in a proof of concept study, we show that they can be used to solve a se- mantic task – antonym identification – better than is possible with existing resources. We focus on the task of finding antonyms in this paper since antonyms usually are distributionally similar but semantically dissimilar words. Hence, it is often not possible to distinguish them from synonyms with distributional models only. In con- trast, we expect that the coreference-based repre- sentations can provide useful complementary in- formation to this task. In general, coreference- based similarity can however be used as an addi- tional feature for any …

Content truncated to 500 words