Abstract
Extracting information from multi-turn dialogues is significantly more complex than from formal documents due to fragmented information, frequent pronoun usage, and the need to understand speaker identities. This paper introduces TUCORE-GCN (TUrn COntext awaRE Graph Convolutional Network), a model designed to capture contextualized representations of turns. By integrating a surrounding turn mask with a heterogeneous dialogue graph, the model effectively extracts relationships and identifies emotions. TUCORE-GCN achieves state-of-the-art performance on major benchmarks, including DialogRE, MELD, and EmoryNLP.
Introduction
In a typical dialogue, information is rarely self-contained. For instance, identifying that "Frank" is the brother of "Speaker 2" requires a model to resolve the pronoun "he," link the name across turns, and understand the speaker's identity. Standard Relation Extraction (RE) models often fail in these scenarios due to low information density. Furthermore, dialogues often involve asymmetric relations (e.g., parent-child), which require the model to distinguish between the subject and object roles accurately.
Methods
The TUCORE-GCN architecture follows a four-stage pipeline designed to handle the nuances of conversational data.
1. Encoding and Turn Attention
The model utilizes a BERT-based backbone enhanced with speaker embeddings to track who is speaking. To resolve local context, the authors implement a surrounding turn mask. This sliding window allows a specific turn to attend primarily to its immediate neighbors, mimicking human cognitive processes for resolving ambiguity and pronouns.
2. Heterogeneous Dialogue Graph
The core of the model is a heterogeneous graph consisting of four node types: Dialogue, Turn, Subject, and Object. These are interconnected by:
- Speaker edges: Connecting turns from the same speaker.
- Dialogue edges: Linking turns to a global dialogue node.
- Argument edges: Connecting turns to the subjects or objects they mention.
3. Sequential GCN
To preserve the chronological order of the conversation—a feature often lost in standard GCNs—the turn nodes are processed through a BiLSTM before entering the Graph Convolutional Network. This ensures the model integrates temporal sequence reasoning with spatial graph reasoning.
4. Emotion Recognition as Relation Extraction
A key innovation is treating Emotion Recognition in Conversations (ERC) as a relation extraction task. By defining the speaker as the "Subject," the utterance as the "Object," and the emotion as the "Relation," the model can leverage its RE capabilities for emotion detection.
Evaluation and Results
The model was evaluated on DialogRE for relation extraction and MELD, EmoryNLP, and DailyDialog for emotion recognition.
- Relation Extraction: TUCORE-GCN achieved an F1 score of 73.1% on DialogRE, outperforming previous state-of-the-art models like GDPNet.
- Emotion Recognition: On MELD and EmoryNLP, the model surpassed COSMIC, despite COSMIC's use of massive external commonsense knowledge bases. This suggests that structural graph modeling can be more effective than simply increasing data volume.
Conclusion
TUCORE-GCN demonstrates that combining local turn context with global graph-based reasoning is highly effective for dialogue understanding. By addressing the "asymmetry problem" through distinct subject/object nodes and preserving temporal order via BiLSTMs, the model sets a new standard for extracting meaning from complex, multi-turn conversations. Future work will explore extending this architecture to tasks such as dialogue summarization and question answering.
Summary updated