DiaASQ: Dialogue Sentiment Quadruple Extraction

• DiaASQ is a task that extracts sentiment quadruples (target, aspect, opinion, polarity) from dialogues by modeling cross-utterance dependencies and speaker interactions.
• It employs techniques such as grid tagging, graph-based neural networks, and multi-scale context aggregation to improve extraction accuracy.
• Benchmarks reveal a Micro-F1 boost from 33–35% to over 44% by integrating discourse-aware methods and advanced position encoding.

Dialogues Aspect-based Sentiment Quadruple Extraction (DiaASQ) refers to the task of end-to-end identification and linking of all sentiment quadruples—target, aspect, opinion, sentiment polarity—within multi-round, multi-participant dialogues. Unlike traditional aspect-based sentiment analysis (ABSA), DiaASQ must handle the diffuse and recursive nature of dialogue structure, including cross-utterance quadruples, speaker alternation, and intertwined reply threads. Recent advances in this domain have established benchmark datasets, novel neural architectures, and theoretically principled graph-based and partitioning frameworks that explicitly model these dialogic constraints (Li et al., 2022, Li et al., 3 May 2026, Peng et al., 7 Aug 2025, Li et al., 2023, Li et al., 2024).

1. Problem Definition and Task Formalization

DiaASQ formalizes input as a dialogue $D = \{u_1, \ldots, u_n\}$ with associated speaker sequence and reply links ($R = \{l_1,\ldots,l_n\}$, with $l_i \in \{1,\dots,i-1\}$ or $0/-1$ for roots). Output consists of a set of sentiment quadruples

$Q = \{q_k = (t_k, a_k, o_k, p_k)\},$

where $t_k, a_k, o_k$ are contiguous text spans from $D$ and $p_k$ is the sentiment polarity ($\mathrm{POS}$, $\mathrm{NEG}$, $R = \{l_1,\ldots,l_n\}$0).

To address the combinatorial space and cross-utterance dependencies, models typically cast the problem as “grid tagging” or token-pair classification over all possible tuples of tokens (or spans), decomposing it into subtasks: entity boundary detection, entity alignment (e.g., head-to-head (H2H), tail-to-tail (T2T)), and polarity detection (Li et al., 2022, Li et al., 3 May 2026).

A key challenge is that up to 22–25% of quadruples span multiple utterances, and correct extraction requires modeling complex discourse and interactional structure (Li et al., 2022, Li et al., 3 May 2026, Peng et al., 7 Aug 2025).

2. Benchmark Datasets and Evaluation

The principal DiaASQ benchmark, introduced by Li et al., comprises parallel Chinese and English corpora (1000 dialogues each, 7452/8000+ utterances, 5500+ quadruples, 22% cross-utterance), focusing on real-world mobile-device discussions (Li et al., 2022). Annotation is span-based, with trained annotators and explicit cross-lingual projection. Datasets are split 800/100/100 into train/dev/test.

Evaluation metrics include:

• Span F1: exact match for target/aspect/opinion identification
• Pair/Triplet/Quadruple F1: joint correctness for span linkage
• Quadruple Micro-F1: exact match for all four elements including polarity
• Identification F1: quadruple but ignoring polarity

Baseline and contemporary models report results on both overall and cross-utterance extraction.

Dataset	# Dialogues	# Quadruples	% Cross-Utt
DiaASQ (ZH/EN)	1,000	~5,500	~22%

Performance has advanced from 33–35% Micro-F1 (base models) to over 44% on ZH and 42% on EN with structural advances (Li et al., 2022, Li et al., 3 May 2026, Peng et al., 7 Aug 2025).

3. Structural Discourse Modeling Approaches

A central objective in DiaASQ is to explicitly model the non-sequential, tree-like reply and speaker structure of dialogues. Several frameworks have been proposed for this:

a. Thread-Constrained Directed Acyclic Graphs (TC-DAG)

The TCDA architecture encodes utterance-level dependencies as a thread-constrained DAG, where each utterance serves as a node and directed edges respect both reply-chain threading and speaker continuity. Edges are selectively constructed:

• Up to $R = \{l_1,\ldots,l_n\}$1 same-speaker predecessors per utterance within the same thread
• Different-speaker utterances within this window are also linked to preserve semantic flow
• Global “root anchoring” ensures all threads maintain global connectivity without arbitrary cross-thread edges

Graph message passing is performed by a relational GNN with attention and GRU-based dual updates, repeated for $R = \{l_1,\ldots,l_n\}$2 layers. This architecture enables strict enforcement of thread boundaries and chronological progression, while suppressing cross-thread noise—identified as a source of structural error in prior models (Li et al., 3 May 2026).

b. Graph- and Attention-based Methods

Other methods use combinations of intra-utterance dependency parsing (intra-GCN), dual graph attention networks (DualGATs) over speakers and dialogue structure, and masked attention mechanisms (e.g., speaker/reply/thread masks) to inject discourse signals into token representations (Li et al., 2024, Li et al., 2022).

c. Entropy-Based Partitioning (SEMDia)

Structural Entropy Minimization Partitioning divides a dialogue graph into semantically coherent sub-dialogues by partitioning an utterance similarity graph (via PLM [CLS] representations and cosine affinity) to minimize structural entropy:

$R = \{l_1,\ldots,l_n\}$3

where $R = \{l_1,\ldots,l_n\}$4 is the encoding tree. The result is a set of minimal, independent clusters (sub-dialogues), reducing irrelevant context and suppressing noise during quadruple extraction. This approach is empirically critical for high cross-utterance F1 (Peng et al., 7 Aug 2025).

4. Context Aggregation and Position Encoding

The effective capture of both micro- (within-utterance) and macro- (across-utterance, thread, or turn) position and context is essential.

• Dynamic Multi-scale Context Aggregation (DMCA): Generates exhaustive local-to-global utterance windows and aggregates token-pair logits via a dynamic hierarchical aggregation (DHA) module. Losses operate at window, thread, and dialogue scales to ensure consistency at all context levels (Li et al., 2023).
• Discourse-Aware Rotary Position Embedding (D-RoPE): Advances standard RoPE by simultaneously encoding token-level and utterance-level indices, and using thread-aware sign inversion to distinguish tokens from the same versus different threads. D-RoPE addresses “distance dilution”—where long utterances or threads degrade positional cues in vanilla RoPE—by aligning micro- and macro-rotary embeddings in a single dot product (Li et al., 3 May 2026).
• Masked Attention / Relative Positioning: Use of attention masks for speaker, reply, and thread, along with RoPE variants, further supports syntactic and discourse-level differentiation (Li et al., 2022).

These methodologies achieve robust fusion of local and global cues, which is quantitatively validated by persistence of cross-utterance F1 at distances $R = \{l_1,\ldots,l_n\}$53 utterances, where non-hierarchical methods collapse (Li et al., 2023, Li et al., 2022).

5. Extraction Pipelines and Training Objectives

Extraction frameworks typically consist of:

1. Span/Token Pair Identification: 2D grid-tagging (boundary, pair, polarity), or 2D span-relation tables over each utterance (SEMDia, DMCA) (Peng et al., 7 Aug 2025, Li et al., 2023).
2. Quadruple Matching/Decoding: Rule- or classifier-based matching of cross-type spans (target–aspect, aspect–opinion, etc.), with sentiment derived from aspect–opinion relations.
3. Multi-Stage Loss Optimization: Weighted cross-entropy losses are used for each grid (ent/pair/pol), possibly stratified at multiple context aggregation levels (window/thread/dialogue) (Li et al., 2023).

Utterance- and sub-dialogue-level partitioning (e.g., in SEMDia), as well as pooling/aggregation mechanisms (Top-K, MaxPool, dynamic hierarchical aggregation), are designed to cope with the volume and diversity of candidate quadruple elements while minimizing computational cost.

6. Empirical Results and Comparative Analysis

Recent models have pushed the quadruple Micro-F1 from $R = \{l_1,\ldots,l_n\}$634–35% (baseline) to over 44.6% (ZH) and 42.4% (EN), with further enhancements for cross-utterance performance.

Model	ZH Micro-F1	EN Micro-F1	Notable Techniques
DiaASQ (2022)	34.94	33.31	Grid tagging, multi-view attn, RoPE
Triple GNNs (2024)	42.87	38.32	Intra-GCN + DualGATs
DMCA (2023)	42.68	37.96	Multi-scale window, DHA
SEMDia (2025)	44.61	42.40	Entropy-based partition, 2-step match
TCDA (2026)	44.35	39.69	TC-DAG + D-RoPE

Ablation studies confirm that both discourse-level structure (TC-DAG, partitioning) and advanced position encoding (D-RoPE, hierarchical aggregation) are independently essential; suppressing either consistently reduces F1 by 0.6–1.0 points (Li et al., 3 May 2026, Peng et al., 7 Aug 2025).

SEMDia achieves state-of-the-art F1 with lower computational cost ($R = \{l_1,\ldots,l_n\}$7ms per dialogue), and entropy-based partitioning is crucial—removal leads to $R = \{l_1,\ldots,l_n\}$8 F1 drop (Peng et al., 7 Aug 2025).

7. Challenges, Limitations, and Future Directions

Persistent challenges in DiaASQ include:

• Extraction for deeply nested threads, multi-speaker co-reference, and highly overlapping quadruples (Li et al., 2022, Li et al., 3 May 2026)
• Noisy linguistic resources: e.g., dependency parse errors degrade intra-GCN performance (Li et al., 2024)
• Computational scaling for long or multi-party dialogues (SEMDia and TCDA suggest optimizations via partitioning and top-K aggregators)
• Explicit handling of coreference and multi-lingual transfer (Li et al., 2022)

Future directions identified include extending thread-constrained graph models to richer dialogue topologies, adaptive partitioning (e.g., dynamic windows), exploring plug-and-play positional encodings (D-RoPE) in related tasks, and development of models for multi-party or multi-domain settings (e.g., emotion recognition) (Li et al., 3 May 2026, Peng et al., 7 Aug 2025). Integration with explicit coreference models, and leveraging parallel ZH/EN annotations, are also open avenues (Li et al., 2022).