Abstract
Identifying speakers in novels aims at determining who says a quote in a given context by text analysis. This task is important for speech synthesis systems to assign appropriate voices to the quotes when producing audio books. Several English datasets have been constructed for this task. However, the difference between English and Chinese impedes processing Chinese novels using the models built on English datasets directly. Therefore, this paper presents a Chinese dataset, which contains 2,548 quotes from World of Plainness, a famous Chinese novel, with manually labelled speaker identities. Furthermore, two baseline speaker identification methods, i.e., a rule-based one and a classifier-based one, are designed and experimented using this Chinese dataset. These two methods achieve accuracies of 53.77% and 58.66% respectively on the test set.
Introduction
In the realm of automated audiobook production, the ability of an AI to distinguish between different speakers in a narrative is crucial for providing an immersive listening experience. This task, known as speaker identification or quote attribution, involves determining which character is responsible for a specific piece of dialogue within a given context. While significant progress has been made in English NLP using corpora like the Columbia Quoted Speech Corpus, Chinese literature presents unique linguistic and structural challenges. Naming conventions, such as the frequent use of titles or relational terms (e.g., "Director Tian" or "Sister Runye"), and the lack of explicit speaker mentions in many sentences necessitate the development of language-specific resources. This paper introduces a new Chinese dataset and evaluates baseline models to establish a benchmark for this task.
Dataset Construction
The researchers developed a benchmark dataset using the renowned Chinese novel World of Plainness. The construction process followed a systematic five-step pipeline to ensure high-quality annotations.
The dataset comprises 2,548 manually labeled quotes. The extraction process identified 126 distinct speaking characters. To provide sufficient context for identification, each quote is accompanied by a "context window" of 21 sentences (10 preceding and 10 following). The final dataset achieved a high manual annotation agreement rate of 94.44%, reflecting the reliability of the labels.
Methodology
Two primary methods were evaluated to establish performance baselines:
- Rule-Based Method: This approach utilizes a salience-based logic, scoring character candidates based on their proximity to the quote, the presence of specific speech verbs (e.g., "said," "replied"), and established conversation patterns. It identifies four key candidates for each quote: the Actor (highest salience score), Best (most frequent mentions), Next (second most frequent), and Last (the speaker of the preceding quote).
- Classifier-Based Method: A 2-layer Multi-Layer Perceptron (MLP) was trained using 16 linguistic features. These features include mention counts inside and outside the quote, syntactic roles (subject/object) determined via dependency parsing, gender-specific pronouns, and punctuation patterns (such as colons or periods) that typically signal a speaker.
Results and Analysis
The experimental results demonstrate that the data-driven classifier outperforms the rule-based system, though both significantly exceed a random baseline.
A deeper analysis reveals that performance varies significantly by quote category. The models excel at "Explicit" quotes, where the speaker is named directly in the sentence, achieving up to 92.16% accuracy. However, performance drops sharply for "Implicit" quotes (where pronouns are used) and "Latent" quotes (where no speaker is mentioned in the immediate vicinity). Given that over 70% of the dataset consists of "Latent" quotes, this category represents the primary bottleneck for current systems. Furthermore, the distribution of quotes follows a power law, with the top 30 characters accounting for nearly 80% of all dialogue, creating an unbalanced distribution that models must navigate.
Conclusion and Future Work
This work provides the first dedicated Chinese dataset for speaker identification in novels, highlighting the specific difficulties posed by Chinese literary text. While the MLP classifier provides a solid baseline, the high prevalence of latent speakers suggests that more advanced context-aware models are needed. Future research will focus on expanding the dataset to include multiple novels and genres, implementing transformer-based architectures like BERT to better capture narrative flow, and improving coreference resolution to better link pronouns to their respective characters.
Summary updated