Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Quotation Attribution with Fictional Character Embeddings

Gaspard Michel, Elena V. Epure, Romain Hennequin, Christophe Cerisara

TL;DR

This work tackles the challenge of quotation attribution in literary texts, focusing on non-explicit quotes where context alone is insufficient. It proposes enriching BookNLP with literary-domain fictional character embeddings derived from domain-specific Universal Authorship Representations (UAR), trained on a new DramaCV corpus. The authors introduce DramaCV and two UAR variants (UAR_Scene and UAR_Play), demonstrate AV improvements on DramaCV, and show that combining these embeddings with BookNLP yields state-of-the-art attribution of non-explicit quotes on the PDNC dataset. The results indicate that domain-tailored character embeddings substantially enhance attribution accuracy, especially for implicit quotes, and highlight the value of domain adaptation for literary text understanding in digital humanities applications.

Abstract

Humans naturally attribute utterances of direct speech to their speaker in literary works. When attributing quotes, we process contextual information but also access mental representations of characters that we build and revise throughout the narrative. Recent methods to automatically attribute such utterances have explored simulating human logic with deterministic rules or learning new implicit rules with neural networks when processing contextual information. However, these systems inherently lack \textit{character} representations, which often leads to errors in more challenging examples of attribution: anaphoric and implicit quotes. In this work, we propose to augment a popular quotation attribution system, BookNLP, with character embeddings that encode global stylistic information of characters derived from an off-the-shelf stylometric model, Universal Authorship Representation (UAR). We create DramaCV (Code and data can be found at https://github.com/deezer/character_embeddings_qa ), a corpus of English drama plays from the 15th to 20th century that we automatically annotate for Authorship Verification of fictional characters utterances, and release two versions of UAR trained on DramaCV, that are tailored for literary characters analysis. Then, through an extensive evaluation on 28 novels, we show that combining BookNLP's contextual information with our proposed global character embeddings improves the identification of speakers for anaphoric and implicit quotes, reaching state-of-the-art performance.

Improving Quotation Attribution with Fictional Character Embeddings

TL;DR

This work tackles the challenge of quotation attribution in literary texts, focusing on non-explicit quotes where context alone is insufficient. It proposes enriching BookNLP with literary-domain fictional character embeddings derived from domain-specific Universal Authorship Representations (UAR), trained on a new DramaCV corpus. The authors introduce DramaCV and two UAR variants (UAR_Scene and UAR_Play), demonstrate AV improvements on DramaCV, and show that combining these embeddings with BookNLP yields state-of-the-art attribution of non-explicit quotes on the PDNC dataset. The results indicate that domain-tailored character embeddings substantially enhance attribution accuracy, especially for implicit quotes, and highlight the value of domain adaptation for literary text understanding in digital humanities applications.

Abstract

Humans naturally attribute utterances of direct speech to their speaker in literary works. When attributing quotes, we process contextual information but also access mental representations of characters that we build and revise throughout the narrative. Recent methods to automatically attribute such utterances have explored simulating human logic with deterministic rules or learning new implicit rules with neural networks when processing contextual information. However, these systems inherently lack \textit{character} representations, which often leads to errors in more challenging examples of attribution: anaphoric and implicit quotes. In this work, we propose to augment a popular quotation attribution system, BookNLP, with character embeddings that encode global stylistic information of characters derived from an off-the-shelf stylometric model, Universal Authorship Representation (UAR). We create DramaCV (Code and data can be found at https://github.com/deezer/character_embeddings_qa ), a corpus of English drama plays from the 15th to 20th century that we automatically annotate for Authorship Verification of fictional characters utterances, and release two versions of UAR trained on DramaCV, that are tailored for literary characters analysis. Then, through an extensive evaluation on 28 novels, we show that combining BookNLP's contextual information with our proposed global character embeddings improves the identification of speakers for anaphoric and implicit quotes, reaching state-of-the-art performance.
Paper Structure (24 sections, 2 equations, 3 figures, 7 tables)

This paper contains 24 sections, 2 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Quotation Attribution.
  4. Fictional Character Embeddings.
  5. Representing Fictional Characters
  6. DramaCV
  7. Scene.
  8. Play.
  9. Literary UAR
  10. Authorship Verification Results
  11. Quotation attribution
  12. BookNLP
  13. Integrating Character Embeddings
  14. Experimental Setup
  15. Results
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Excerpt of Pride and Prejudice by Jane Austen (1813). Quotations are colored by quote type: explict, implicit and anaphoric. Speaker information given by the narrator are underlined.
  • Figure 2: Detailed process of BookNLP's quotation attribution system. We inject extra-information in the form of character embeddings computed from explicit quotes $v_{\text{entity}(m)}$ in the fourth step, and quote embeddings $u_q$ in the second step.
  • Figure 3: Accuracy results on non-explicit quotes averaged over subsets of novels that have a low percentage of unanswerable quotes, with standard deviations.