When reading a novel, we rarely see every line of dialogue explicitly tagged with a name or a speech verb like "said Alice." To maintain narrative pacing and natural flow, authors frequently omit these markers, creating what are known as "floating" or "implicit" quotations. Human readers track who is speaking effortlessly by following the conversational rhythm, turn-taking patterns, and character relationships. For natural language processing models, however, this is a notorious blind spot. Traditional NLP pipelines rely heavily on local syntactic patterns and nearby name mentions; when these surface-level cues are missing, standard models fail.
The Scale of the Latent Speaker Problem
The challenge of quote attribution in literature stems from how frequently authors rely on implicit discourse. As shown by Jia-Xiang Chen et al. in A Chinese Dataset for Identifying Speakers in Novels, over 70% of quotes in literary dialogue can lack an explicit speaker mention or pronoun in the immediate sentence. This high frequency of latent speakers turns quote attribution from a local pattern-matching task into a long-range discourse analysis problem.
In Automatic Attribution of Quoted Speech in Literary Narrative, David K. Elson et al. noted that these floating quotes lack direct syntactic links to characters, requiring models to understand the pragmatics of storytelling rather than sentence-level syntax. When structural cues like "said" or "asked" are absent, model accuracy drops significantly. Benchmarks established by Wenjie Zhong et al. in Who Said What: Formalization and Benchmarks for the Task of Quote Attribution reveal that accuracy rates for implicit quotes are roughly 30% lower than for explicit ones, highlighting a fundamental limitation in how models track long-range discourse structure and turn-taking logic.
Modeling Conversational Flow and Alternation
To resolve latent speakers without explicit tags, models must capture the structural patterns inherent to multi-turn dialogue—specifically, the natural back-and-forth rhythm of speaker alternation.
Early machine learning approaches attempted to capture this by incorporating structural features directly into ranking frameworks. For example, Hua He et al. in Identification of Speakers in Novels combined speaker alternation patterns, vocative detection, and unsupervised actor-topic modeling to distinguish between potential speakers without direct textual labels.
As deep learning matured, researchers began treating dialogue as a continuous sequence rather than a set of isolated classification tasks. Yue Chen et al. demonstrated in A Neural-Network-Based Approach to Identifying Speakers in Novels that pairing a BERT-based scoring network with a global "revision" algorithm allows a system to leverage the natural back-and-forth rhythm of dialogue to correct local errors.
Taking this sequential perspective further, Carolina Cuesta-Lázaro et al. in What does the sea say to the shore? A BERT based DST style approach for speaker to dialogue attribution in novels adapted Dialogue State Tracking (DST) techniques from conversational AI. By combining Distil-BERT embeddings with a sequence-modeling Conditional Random Field (CRF), the authors modeled the conversation as an evolving state machine, achieving substantial performance gains by explicitly tracking speaker transitions across long narrative spans.
Forcing Models to Learn Context
A persistent issue in training speaker identification models is that neural networks are "lazy" learners; if a character's name happens to appear near a quote, the model will often rely on that local shortcut rather than learning the underlying narrative flow. To break this dependency on surface-level cues, researchers have introduced techniques that force models to look at the broader context.
One approach is strategic data augmentation. In A Simple-Yet-Effective Data Augmentation Method for Speaker Identification in Novels, Wenjie Zhong et al. generated synthetic dialogue and intentionally removed or replaced explicit speaker tags with ambiguous pronouns during training. This manipulation forces the model to ignore local linguistic crutches and instead infer identity from character interactions and the surrounding narrative.
Another mechanism involves abstracting character identities entirely during the encoding phase. The Symbolization, Prompt, and Classification (SPC) framework, introduced by Heng Lu et al. in Symbolization, Prompt, and Classification: A Framework for Implicit Speaker Identification in Novels, replaces explicit character names with abstract symbols. By masking the literal names, the framework prevents the model from relying on simple keyword matching and forces it to maintain a deeper, generalized understanding of conversation flow and character relationships.
The Generative Paradigm Shift
The transition toward Large Language Models (LLMs) has fundamentally shifted how implicit speaker attribution is handled, moving the field from discriminative classification to prompt-based generation.
Liyan Xu et al. explored this in SIG: Speaker Identification in Literature via Prompt-Based Generation, demonstrating that generative models using prompt-based templates achieve a substantial performance leap on implicit quotations. This approach succeeds because it leverages the model's pre-trained understanding of complex discourse patterns and social dynamics to fill in the blanks left by missing narrative tags.
More recently, Gaspard Michel et al. conducted a comprehensive evaluation in Evaluating LLMs for Quotation Attribution in Literary Texts: A Case Study of LLaMa3, showing that models like LLaMa-3 outperform traditional specialized pipelines by up to 19 percentage points. Crucially, the authors introduced metrics to verify that this performance is driven by genuine linguistic reasoning over long context windows rather than the model simply "reciting" classic books it memorized during pre-training.
Why It Matters: Downstream Applications
Resolving latent speakers is not just an academic exercise in computational linguistics; it is the foundational bottleneck for large-scale text analysis across several domains.
- Computational Humanities: In The Emotion Dynamics of Literary Novels, Adam Hammond et al. demonstrated that tracking the distinct emotional journeys of individual characters across large corpora is fundamentally limited by the difficulty of automated speaker attribution. Without accurate attribution, it is impossible to separate a narrator's voice from a character's direct speech to map character dynamics or stylistic shifts (Ali Al-Laith et al., Annotating and Classifying Direct Speech in Historical Danish and Norwegian Literary Texts).
- Scientific and Political Discourse: The underlying challenge of implicit attribution spans other text genres. In technical writing, the agent of an action is frequently omitted, making it difficult to determine if a claim represents the author's own work or a peer's research (Simone Teufel et al., What's Yours and What's Mine: Determining Intellectual Attribution in Scientific Text). Similarly, in parliamentary debates, speakers frequently embed quotes or reference the thoughts of others, creating nested layers of attribution that standard text analysis tools fail to disentangle (Annelen Brunner et al., Out of the Mouths of MPs: Speaker Attribution in Parliamentary Debates).
By moving away from local syntactic rules and toward models that capture long-range conversational flow, context-aware data augmentation, and generative reasoning, modern NLP systems are becoming increasingly capable of reading between the lines to discover exactly who is speaking.
Key Foundational Work
Automatic Attribution of Quoted Speech in Literary Narrative (2010) — contributed 1 concept elements to this direction.
What's Yours and What's Mine: Determining Intellectual Attribution in Scientific Text (2000) — contributed 1 concept elements to this direction.