Quotation Detection and Classification with a Corpus-Agnostic Model

Proceedings of Recent Advances in Natural Language Processing, pages 888–894, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_103 888 Quotation Detection and Classification with a Corpus-Agnostic Model Sean Papay & Sebastian Pad´o Institute for Natural Language Processing University of Stuttgart, Germany {sean.papay,pado}@ims.uni-stuttgart.de Abstract The detection of quotations (i.e., reported speech, thought, and writing) has established itself as an NLP analysis task. However, state-of-the-art models have been developed on the basis of specific corpora and incorpo- rate a high degree of corpus-specific assump- tions and knowledge, which leads to fragmen- tation. In the spirit of task-agnostic modeling, we present a corpus-agnostic neural model for quotation detection and evaluate it on three corpora that vary in language, text genre, and structural assumptions. The model (a) ap- proaches the state-of-the-art on the corpora when using established feature sets and (b) shows reasonable performance even when us- ing solely word forms, which makes it applica- ble for non-standard (i.e., historical) corpora. 1 Introduction Quotation is a general notion that covers different kinds of direct and indirect speech, thought, and writing in text (Semino and Short, 2004). Quo- tations are a prominent linguistic device used to express claims, assessments, or attitudes attributed to speakers. Consequently, the analysis of quota- tions is gaining traction in computational linguis- tics and digital humanities, providing evidence for speaker relationships (Elson et al., 2010; Agarwal et al., 2012), inter-speaker sentiment (Nalisnick and Baird, 2013), politeness (Faruqui and Pado, 2012), and narrative structure (Jannidis et al., 2018). As is often the case with semantic phenomena, manual annotation of quotations has shown to be slow and resource-intensive, in particular when undertaken in conjunction with the annotation of speakers and information quality (Brunner, 2013; Pareti, 2015). This provides the rationale for au- tomatic quotation recognition methods. After a first round of rule-based methods (Pouliquen et al., 2007; Brunner, 2013), recent supervised models use mostly sequence classifiers (Pareti et al., 2013; Almeida et al., 2014; Scheible et al., 2016). Not surprisingly, these corpora differ substan- tially across a number of relevant dimensions. in- cluding text genre, annotation scheme, and the- oretical assumptions. For example, Pareti et al. (2013) focus exclusively on newspaper text and focus on developing a uniform annotation schema that captures the shared properties of all kinds of annotations. Thus, even though this corpus con- tains direct, indirect, and mixed quotations, these not marked as instances of their specific subtypes. In addition, each quote is assumed to be introduced by a cue (markables are shown surrounded by red square brackets): (1) Hillary Clinton on Saturday [acknowledged]CUE [the state of the economy is good]QUOTE. This assumption is generally true for newspaper text, and simplifies the task of quotation detection. The situation is rather different in the literary texts considered by Semino and Short (2004). Cues are much more varied, and are sometimes omitted entirely, such as in this exchange from Dickens’ Christmas Carol: (2) [”Much!”]QUOTE – Marley’s voice, no doubt about it. [”Who are you?”]QUOTE [”Ask me who I was.”]QUOTE The study follows a generally …

Content truncated to 500 words