What does the sea say to the shore? A BERT based DST style approach for speaker to dialogue attribution in novels

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5820 - 5829 May 22-27, 2022 c⃝2022 Association for Computational Linguistics What does the sea say to the shore? A BERT based DST style approach for speaker to dialogue attribution in novels Carolina Cuesta-Lazaro†∗ † Institute for Computational Cosmology Durham University, UK carolina.cuesta-lazaro@durham.ac.uk Animesh Prasad‡ Trevor Wood‡ ‡Alexa AI Amazon, UK {animpras, trevowoo}@amazon.com Abstract We present a complete pipeline to extract characters in a novel and link them to their direct-speech utterances. Our model is divided into three independent components: extracting direct-speech, compiling a list of characters, and attributing those characters to their utter- ances. Although we find that existing systems can perform the first two tasks accurately, at- tributing characters to direct speech is a chal- lenging problem due to the narrator’s lack of explicit character mentions, and the frequent use of nominal and pronominal coreference when such explicit mentions are made. We adapt the progress made on Dialogue State Tracking to tackle a new problem: attributing speakers to dialogues. This is the first appli- cation of deep learning to speaker attribution, and it shows that is possible to overcome the need for the hand-crafted features and rules used in the past. Our full pipeline improves the performance of state-of-the-art models by a relative 50% in F1-score. 1 Introduction Natural Language Processing has enabled a quanti- tative improvement in the humanities, by allowing for large-scale statistical measurements to be taken over hundreds of thousands of books compared to the order of tenths a human could analyse in a much longer time span. Some examples of large-scale literary analyses include studies on characters and their descriptions within the novel, mostly focused on gender differences (Underwood et al., 2018; Kraicer and Piper, 2019), and studies on charac- ter’s relations by extracting social networks from novels (Labatut and Bost, 2019; Jayannavar et al., 2015). Most of these studies demand special attention to dialogues, being a major part of character ex- pression and interaction with other characters. Di- alogues play an instrumental role in plot develop- ∗Work done during internship at Amazon ment, frequently encompassing focal plot moments, especially in fiction – which is also the focus for this study. Here we aim to identify direct-speech utterances that form part of dialogues and associate them with the speaking characters. Such informa- tion is not only useful to enable large-scale socio- temporal studies but also crucial to many down- stream challenging tasks like narrative understand- ing (Iyyer et al., 2016) and summarising (Ladhak et al., 2020). Further, the high-quality dialogue- character association is pertinent for generating engaging text-to-speech for novels with distinct voice profiles for characters. In the past, models that link direct speech to char- acters have been dominated by predefined rules (Muzny et al., 2017) or hand-crafted features (He et al., 2013). When evaluating these models the authors also presumed that a character list, together with the character’s aliases, has been precompiled and that direct-speech …

Content truncated to 500 words