Source-linked AI summary
TCDA: Thread-Constrained Discourse-Aware Modeling for Conversational Sentiment Quadruple Analysis
Xinran Li, Xinze Che, Yifan Lyu, Zhiqi Huang, Xiujuan Xu
TL;DR
DiaASQ requires fine-grained sentiment extraction from multi-round conversations despite structural noise and token-level Distance Dilution. TCDA combines thread-constrained graph modeling with discourse-aware positional encoding, and achieves state-of-the-art results on two benchmark datasets.
Problem
DiaASQ must model complex interrelationships in multi-round dialogues, but existing graph and RoPE methods face structural noise, temporal-order limitations, and Distance Dilution.
Method
TCDA combines TC-DAG, which constrains thread connections while preserving root-based connectivity, with D-RoPE, which separates token- and utterance-level semantics.
Results
TCDA achieves state-of-the-art results on two benchmark datasets.
Takeaways & Limitations
D-RoPE shows potential as a model-agnostic plugin for wider multi-turn dialogue tasks.
Abstract
from arXiv · showhide
Conversational Aspect-based Sentiment Quadruple Analysis (DiaASQ) needs to capture the complex interrelationships in multiple rounds of dialogues. Existing methods usually employ simple Graph Convolutional Networks (GCN), which introduce structural noise and fail to consider the temporal sequence of the dialogues, or use standard RoPE, which implicitly captures relative distances in a flat sequence but cannot clearly separate the token-level syntactic order from the utterance-level progression, and may suffer from the Distance Dilution problem. To address these issues, we propose a new framework that combines Thread-Constrained Directed Acyclic Graph (TC-DAG) and Discourse-Aware Rotary Position Embedding (D-RoPE). Specifically, TC-DAG filters out cross-thread noise based on thread constraints, maintains global connectivity through root anchoring, and incorporates the temporal sequence of the dialogues. D-RoPE aligns multi-layer semantics using dual-stream projection and multi-scale frequency signals, captures thread dependencies using tree-like distances, and alleviates the token-level Distance Dilution problem by incorporating utterance-level progressions. Experimental results on two benchmark datasets demonstrate that our framework achieves state-of-the-art performance.
1 Introduction
DiaASQ extracts sentiment quadruples from multi-round conversations, where fragmented information and complex dependencies challenge existing modeling approaches. TCDA addresses these challenges with thread-constrained structure modeling and discourse-aware positional encoding, achieving state-of-the-art performance.
- Motivation: DiaASQ extracts target, aspect, opinion, and sentiment quadruples from multi-round conversations.Targets, aspects, and opinions correspond to spans, while sentiment represents positive, negative, or neutral polarity.
- Motivation: Fragmented information and complex conversational dependencies make fine-grained sentiment understanding difficult.These challenges distinguish DiaASQ from traditional sentence-level sentiment analysis.
- Limitations of Existing Methods: Standard RoPE can suffer Distance Dilution when verbose utterances separate logically adjacent conversational turns.Expanded token distances may cause positional correlations to decay prematurely, weakening semantic connections.
- Proposed Framework: TCDA combines explicit topological structure with implicit positioning through TC-DAG and D-RoPE.TC-DAG constrains thread-level propagation, while D-RoPE integrates token- and utterance-level semantics in separate subspaces.
- Results: TCDA achieves state-of-the-art performance on the evaluated DiaASQ benchmarks.The paper also makes its implementation publicly available.
2 Related Work
DiaASQ extends aspect-based sentiment analysis from isolated sentences to multi-turn conversations with inter-utterance dependencies. Prior work explores attention, multi-scale mechanisms, thread-aware GCNs, DAGs, and RoPE, but existing structural and positional modeling remains limited.
- DiaASQ and ABSA: Traditional ABSA benchmarks focus mainly on sentence-level tasks, limiting applicability to multi-turn conversational scenarios.DiaASQ was introduced to address this gap by modeling complex inter-utterance correlations.
- Existing DiaASQ Methods: Prior DiaASQ methods use heterogeneous attention, ternary scoring, and multi-scale mechanisms to model quadruple cohesion and long-range dependencies.Examples include H2DT, DMCA, and ICMSR.
- Existing DiaASQ Methods: DMIN incorporates thread structure through GCN and multi-granularity integration to align token interactions with utterance-level discourse.It is described as the first method to use GCN in this way for DiaASQ.
- Structural and Positional Modeling: CA-DAGNet captures cross-utterance dependencies with a directed acyclic graph but ignores inherent thread-based topological constraints.Recent RoPE-based frameworks also typically encode only local token context or use fragmented strategies.
3 Methodology
TCDA combines TC-DAG and D-RoPE in a unified architecture. The framework therefore couples dialogue structure modeling with discourse-aware positional encoding.
- Overall Architecture: TCDA combines TC-DAG and D-RoPE as its overall architecture.The paper presents the architecture in Figure 2.
3.1 Problem Definition
DiaASQ represents each conversation as an ordered sequence of utterances with reply indices and speaker identities, then formulates quadruple extraction as unified relation tagging. The tags encode entity boundaries, entity alignment, and sentiment polarity.
- Conversation Representation: Each conversation is represented by utterances, reply indices, and a speaker sequence.Each utterance consists of a token sequence, and each reply index identifies the direct-response utterance.
- Unified Relation Tagging: Quadruple extraction is reformulated as a unified relation-tagging problem over word pairs in the flattened dialogue.The model identifies semantic connections between token pairs.
- Relation Types: Entity-alignment tags pair different entities through head-to-head and tail-to-tail relations.These relations associate entities such as a target with its corresponding aspect.
- Relation Types: Sentiment-polarity tags classify relations as positive, negative, or neutral, with other assigned when no specific relation exists.The polarity labels are POS, NEG, and NEU.
3.2 Textual Feature Extraction
The framework divides each conversation into sequential threads rooted at a shared node, representing utterances with speaker-aware token features for thread-level encoding.
- Each conversation is divided into multiple threads that start from a common root node.
- Threads are arranged sequentially and intersect only at the root node.
- Each utterance is formatted with a classification token, the utterance, and speaker information.
- Thread-level encoding uses the resulting utterance representations and token features.
3.3 Dual-scale Contextual Encoding
The dual-scale encoder combines thread-level knowledge enhancement with conversation-level discourse modeling, then applies TC-DAG to utterance representations in temporal and reply order.
- Token-level Knowledge Encoding: The framework performs knowledge enhancement within each individual thread to balance global and local interactions.
- Token-level Knowledge Encoding: Thread-level syntactic and semantic adjacency matrices provide complementary knowledge features for reconstructing global token features.
- Token-level Knowledge Encoding: The enhanced token representation is formed by adding original and knowledge features before layer normalization.
- Utterance-level Discourse Modeling: A Top-K aggregator abstracts global token features into utterance representations that capture conversational flow.
- Utterance-level Discourse Modeling: TC-DAG processes utterance representations according to the conversation’s temporal order and replying topology.
3.4 Thread-Constrained DAG
TC-DAG is a chronological, speaker-aware graph that constrains edges within threads while using root anchoring to preserve global connectivity; relational propagation then updates utterance states sequentially.
- TC-DAG represents utterances as nodes with directed edges only from earlier to later utterances.
- The relation set distinguishes same-speaker from inter-speaker dependencies during graph propagation.
- Edges are restricted to a local thread and connect each node to prior utterances until the window reaches the thread boundary.
- When the thread boundary is reached before the window is filled, TC-DAG connects the node to the root to preserve global connectivity.
- Nodes are updated chronologically so predecessor hidden states are available during relation-aware aggregation.
- A dual gated update mechanism uses parallel GRUs to capture complementary context and node-history information flows.
3.5 Global-Local Interaction and Discourse-Aware Position Encoding
The framework reconnects global utterance structure with token features, then uses D-RoPE to encode discourse topology at both token and utterance scales. This dual-scale design preserves semantic and positional consistency while addressing distance dilution.
- Global-Local Interaction: Cross-attention uses token representations as queries and global utterance representations as keys and values to produce Hfinal.This reintegrates TC-DAG-derived discourse context into token-level features.
- Discourse-Aware Position Encoding: D-RoPE decomposes Hfinal into parallel token and utterance streams projected onto separate subspaces.The projections separate local syntactic cues from global discourse semantics before interaction.
- Discourse-Aware Position Encoding: Micro-RoPE encodes token positions using cumulative topological distance from the global root and adjusts coordinates for divergent threads.This represents path lengths across threads while preserving linear relative distances within a thread.
- Discourse-Aware Position Encoding: Macro-RoPE uses utterance-level indices with a reduced base frequency to counteract distance dilution caused by verbose utterances.The design preserves attention to logical dependencies and provides constant turn-level distances as a discourse anchor.
- Discourse-Aware Position Encoding: The two rotated subspaces are concatenated and scored through a dot product, ensuring dual-scale semantic and positional consistency.The fusion creates a unified feature vector before calculating the topology-adaptive score.
3.6 Quadruple Decoding and Learning
The decoding stage applies D-RoPE to task-specific spaces and derives topology-adaptive probabilities, training the model with weighted cross-entropy loss.
- Quadruple Decoding: The integrated representation is projected into sentiment, relation, and polarity spaces before D-RoPE derives topology-adaptive probabilities.These spaces isolate semantic influence for the three decoding tasks.
- Learning: The model minimizes weighted cross-entropy loss during learning.The category weight is denoted by α_ij.
4 Experiments and Analysis
Experiments on Chinese and English DiaASQ benchmarks show that TCDA is competitive or state of the art, while component and structural analyses support the roles of TC-DAG and D-RoPE.
- 4.3 Main Results: TCDA achieves SOTA or competitive performance across all benchmarks.
- 4.4 Ablation Study: Removing TC-DAG, D-RoPE, or both degrades performance, with the sharpest decline when both modules are absent.The ablation results indicate complementary benefits from the two components.
- 4.5 Further Analysis: Best performance occurs at L = 2, while increasing the TC-DAG depth decreases performance because of over-smoothing.
- 4.5 Further Analysis: Dense speaker connections with w ≥2 outperform sparse connections, while performance saturates at w ≥2 because windows often exceed actual thread lengths.Dense connections facilitate direct transmission from the root sentence to subsequent nodes in long-distance dialogue contexts.
- 4.5 Further Analysis: D-RoPE improves MVQPN micro F1 by 1.84% on ZH and 0.80% on EN, and consistently improves performance across architectures and languages.The authors characterize D-RoPE as a model-independent plugin for alleviating Distance Dilution.
- 4.5 Further Analysis: Without thread isolation, unrelated-thread interference lowers standard DAG performance below Reply-GCN on EN, whereas TC-DAG achieves superior results across all metrics.TC-DAG combines chronological order with strict topology.
5 Conclusion
TCDA addresses structural noise and scale mismatch in DiaASQ through TC-DAG and D-RoPE, achieves state-of-the-art results on two benchmarks, and shows potential as a model-agnostic plugin for multi-turn dialogue tasks.
- TCDA addresses structural noise and scale mismatch in DiaASQ using TC-DAG and D-RoPE.
- TCDA achieves SOTA results on two benchmarks.
- D-RoPE generalizes across models and may serve as a model-agnostic plugin for wider multi-turn dialogue tasks.