Learning and Evaluating Character Representations in Novels

Findings of the Association for Computational Linguistics: ACL 2022, pages 1008 - 1019 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Learning and Evaluating Character Representations in Novels Naoya Inoue, Charuta Pethe, Allen Kim, Steven Skiena Stony Brook University {ninoue,cpethe,allekim,skiena}@cs.stonybrook.edu Abstract We address the problem of learning fixed- length vector representations of characters in novels. Recent advances in word embeddings have proven successful in learning entity rep- resentations from short texts, but fall short on longer documents because they do not cap- ture full book-level information. To overcome the weakness of such text-based embeddings, we propose two novel methods for represent- ing characters: (i) graph neural network-based embeddings from a full corpus-based charac- ter network; and (ii) low-dimensional embed- dings constructed from the occurrence pattern of characters in each novel. We test the qual- ity of these character embeddings using a new benchmark suite to evaluate character repre- sentations, encompassing 12 different tasks. We show that our representation techniques combined with text-based embeddings lead to the best character representations, outperform- ing text-based embeddings in four tasks. Our dataset is made publicly available to stimulate additional work in this area. 1 Introduction High-quality distributed representations of char- acters (henceforth, character embeddings) play an important role for the computational analysis of nar- rative texts (Iyyer et al., 2016; Xanthos et al., 2016; Skorinkin, 2017; Azab et al., 2019; Labatut and Bost, 2019; Kubis, 2021; Brahman et al., 2021). Ideally, characters who share similar properties such as job, gender and a relationship to other characters, should possess similar character em- beddings even if they are in different stories (e.g. Cinderella and Juliet, both young women in for- bidden romance situations). This paper aims for learning such fixed-length, distributed representa- tions from novels. The core problem of learning character embed- dings is how to aggregate and embed the contex- tual information of characters into distributed rep- o ( x ( . ( ^ ( s ( (Th v ( Irv * ( D ( < ( (Ha > ( Frank Merriwell Farmer Green Motor Matt Tom Gray Peter Rabbit Johnnie Green Tom Rover Tom Swift Dick Prescott Dave Darrin Figure 1: t-SNE visualization of our character embed- dings for ten characters. Each character is sampled from more than 24 different books. The proposed method assigns similar representations to each charac- ter even though they exist in different books. The pro- posed method uses no surface form matching. resentations. Conventionally, this has been exten- sively studied in word embeddings, including static word embeddings such as word2vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014), and in contextualized word embeddings such as ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019). All these methods follow the Distributional Hypothesis: “words that occur in the same context tend to have similar meanings” (Harris, 1954). One limitation of these approaches is that they represent word embeddings by local context: they split documents into individual sentences or small chunks, ignoring the document information of each input. …

Content truncated to 500 words