A Joint Model for Quotation Attribution and Coreference Resolution

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 39–48, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Joint Model for Quotation Attribution and Coreference Resolution Mariana S. C. Almeida∗† Miguel B. Almeida∗† Andr´e F. T. Martins∗† ∗Priberam Labs, Alameda D. Afonso Henriques, 41, 2o, 1000-123 Lisboa, Portugal †Instituto de Telecomunicac¸˜oes, Instituto Superior T´ecnico, 1049-001 Lisboa, Portugal {mla,mba,atm}@priberam.pt Abstract We address the problem of automatically attributing quotations to speakers, which has great relevance in text mining and me- dia monitoring applications. While cur- rent systems report high accuracies for this task, they either work at mention- level (getting credit for detecting uninfor- mative mentions such as pronouns), or as- sume the coreferent mentions have been detected beforehand; the inaccuracies in this preprocessing step may lead to error propagation. In this paper, we introduce a joint model for entity-level quotation attri- bution and coreference resolution, exploit- ing correlations between the two tasks. We design an evaluation metric for attribu- tion that captures all speakers’ mentions. We present results showing that both tasks benefit from being treated jointly. 1 Introduction Quotations are a crucial part of news stories, giv- ing the perspectives of the participants in the nar- rated event, and making the news sound objective. The ability of extracting and organizing these quo- tations is highly relevant for text mining applica- tions, as it may aid journalists in fact-checking, help users browse news threads, and reduce human intervention in media monitoring. This involves assigning the correct speaker to each quote—a problem called quotation attribution (§2). There is significant literature devoted to this task, both for narrative genres (Mamede and Chaleira, 2004; Elson and McKeown, 2010) and newswire domains (Pouliquen et al., 2007; Sar- mento et al., 2009; Schneider et al., 2010). While the earliest works focused on devising lexical and syntactic rules and hand-crafting grammars, there has been a recent shift toward machine learning approaches (Fernandes et al., 2011; O’Keefe et al., 2012; Pareti et al., 2013), with latest works re- porting high accuracies for speaker identification in newswire (in the range 80–95% for direct and mixed quotes, according to O’Keefe et al. (2012)). Despite these encouraging results, quotation min- ing systems are not yet fully satisfactory, even when only direct quotes are considered. Part of the problem, as we next describe, has to do with inaccuracies in coreference resolution (§3). The “easiest” instances of quotation attribution problems arise when the speaker and the quote are semantically connected, e.g., through a reported speech verb like said. However, in newswire text, the subject of this verb is commonly a pronoun or another uninformative anaphoric mention. While the speaker thus determined may well be correct— being in most cases consistent with human annota- tion choices (Pareti, 2012)—from a practical per- spective, it will be of little use without a corefer- ence system that correctly resolves the anaphora. Since the current state of the art in coreference res- olution is far from perfect, errors …

Content truncated to 500 words