Model Architectures for Quotation Detection

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1736–1745, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Model Architectures for Quotation Detection Christian Scheible, Roman Klinger and Sebastian Pad´o Institut f¨ur Maschinelle Sprachverarbeitung Universit¨at Stuttgart {scheibcn,klinger,pado}@ims.uni-stuttgart.de Abstract Quotation detection is the task of locating spans of quoted speech in text. The state of the art treats this problem as a sequence labeling task and employs linear-chain con- ditional random fields. We question the ef- ficacy of this choice: The Markov assump- tion in the model prohibits it from making joint decisions about the begin, end, and internal context of a quotation. We per- form an extensive analysis with two new model architectures. We find that (a), sim- ple boundary classification combined with a greedy prediction strategy is competitive with the state of the art; (b), a semi-Markov model significantly outperforms all others, by relaxing the Markov assumption. 1 Introduction Quotations are occurrences of reported speech, thought, and writing in text. They play an impor- tant role in computational linguistics and digital humanities, providing evidence for, e.g., speaker relationships (Elson et al., 2010), inter-speaker sen- timent (Nalisnick and Baird, 2013) or politeness (Faruqui and Pado, 2012). Due to a lack of general- purpose automatic systems, such information is often obtained through manual annotation (e.g., Agarwal et al. (2012)), which is labor-intensive and costly. Thus, models for automatic quotation detec- tion form a growing research area (e.g., Pouliquen et al. (2007); Pareti et al. (2013)). Quotation detection looks deceptively simple, but is challenging, as the following example shows: [The pipeline], the company said, [would be built by a proposed joint venture ..., and Trunkline . . . will “build and operate” the system . . . ].1 1Penn Attributions Relation Corpus (PARC), wsj 0260 Note that quotations can (i) be signalled by lexi- cal cues (e.g., communication verbs) without quota- tion marks, (ii) contain misleading quotation marks; (iii) be discontinuous, and (iv) be arbitrarily long. Early approaches to quotation detection use hand-crafted rules based on syntactic mark- ers (Pouliquen et al., 2007; Krestel et al., 2008). While yielding high precision, they suffered from low recall. The state of the art (Pareti et al., 2013; Pareti, 2015) treats the task as a sequence classifi- cation problem and uses a linear-chain conditional random field (CRF). This approach works well for the prediction of the approximate location of quo- tations, but yields a lower performance detecting their exact span. In this paper, we show that linear-chain sequence models are a sub-optimal choice for this task. The main reason is their length, as remarked above: Most sequence labeling tasks in NLP (such as most cases of named entity recognition) deal with spans of a few tokens. In contrast, the median quotation length on the Penn Attributions Relation Corpus (PARC, Pareti et al. (2013)) is 16 tokens and the longest span has over 100 tokens. As a result of the strong Markov assumptions that linear-chain CRFs make to ensure tractability, …

Content truncated to 500 words