Source-linked AI summary
Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs
Jiaao Chen, Diyi Yang
TL;DR
Conversation summaries can omit, repeat, or misstate content because human interactions are unstructured and complex, with interruptions and multiple speakers. The paper represents utterance relations and speaker actions with graphs and combines those representations through a multi-granularity decoder. Its models outperform prior methods on SAMSum and show zero-shot generalization on ADCS.
Problem
Conversation summaries can be insufficient, redundant, or incorrect, while interruptions and multiple speakers make salient-content selection and correct action attribution difficult.
Source: §1 · p. 1Method
The model encodes discourse relations between utterances and WHO-DOING-WHAT action triples as graphs, then uses a multi-granularity decoder to combine information levels for summary generation.
Source: §1 · p. 2, §3.1 · p. 3Results
The combined discourse and action graph model outperforms prior state-of-the-art methods on SAMSum and improves zero-shot ADCS results; on ADCS, ROUGE-1 increases 66.2% over BART.
Source: §4.5 · p. 7, §4.4 · p. 7, §5 · p. 9Takeaways & Limitations
Results on SAMSum and zero-shot ADCS support the models’ effectiveness and generalizability for conversation summarization.
Source: §5 · pp. 8–9
Abstract
from arXiv · showhide
Abstractive conversation summarization has received much attention recently. However, these generated summaries often suffer from insufficient, redundant, or incorrect content, largely due to the unstructured and complex characteristics of human-human interactions. To this end, we propose to explicitly model the rich structures in conversations for more precise and accurate conversation summarization, by first incorporating discourse relations between utterances and action triples ("who-doing-what") in utterances through structured graphs to better encode conversations, and then designing a multi-granularity decoder to generate summaries by combining all levels of information. Experiments show that our proposed models outperform state-of-the-art methods and generalize well in other domains in terms of both automatic evaluations and human judgments. We have publicly released our code at https://github.com/GT-SALT/Structure-Aware-BART.
1 Introduction
Conversation summarization must organize complex, unstructured exchanges into concise, informative text, but interruptions and multiple speakers make salient-content selection and correct speaker–action attribution difficult. The paper proposes discourse-relation and action graphs to represent these structures and combine them for summarization.
1 Introduction
- Conversation interruptions and dispersed key information complicate salience selection, while multiple speakers and varied language styles complicate correct action attribution.These conversational properties make it difficult to generate summaries that are both informative and factually grounded.Source: §1 · pp. 1–2
- Prior approaches still face succinctness and faithfulness challenges, with limited explicit modeling of utterance dependencies and structured speaker–action associations.The paper argues these structures matter for identifying salient conversation portions and generating factual summaries.Source: §1 · p. 2
- Discourse-relation graphs link utterances, while action graphs represent WHO-DOING-WHAT triples, giving the model structure for identifying key content and speaker–action associations.The paper describes 16 discourse relation types between utterances and uses action triples to make references such as who will fetch an item explicit.Source: §1 · p. 2
- The proposed structure-aware models combine discourse and action graphs with a multi-granularity decoder, and experiments report effectiveness on SAMSum and generalization to debate summarization.The introduction identifies SAMSum as the evaluation dataset and debate summarization as a new domain.Source: §1 · p. 2
2 Related Work
Prior work on document summarization includes sequence-to-sequence, pointer-generator, reinforcement-learning, and pretrained-model approaches, while conversation summarization has explored hierarchical models, commonsense, and conversational structures. This paper’s related-work discussion identifies gaps in explicitly modeling dependencies between utterances and speaker actions.
2 Related Work
- Graph-based document summarization has used entity types, knowledge graphs, and fact-correction modules; this paper draws on that line of work to construct action graphs for more factual conversation summaries.The cited examples include entity-type extraction, knowledge-graph use, and extra fact-correction modules.Source: §2 · p. 2
- Conversation summarization research includes hierarchical models, commonsense knowledge, dialogue acts, key-point sequences, topic segments, and stage developments.Prior work also used discourse relations as classifier features to detect important conversation content.Source: §2 · pp. 2–3
- The paper identifies limited explicit use of dependencies between utterances and of speaker actions and their interactions, linking these gaps to difficulty leveraging salient utterances and risks of incorrect references or reasoning.Its proposed response is to model actions within utterances and relations between utterances using structured graphs and graph encoders with multi-granularity decoding.Source: §2 · p. 3
3 Methods
The method models conversations with discourse-relation and action graphs, encodes those structures alongside the conversation, and combines their representations through a multi-granularity decoder.
3.1 Structured Graph Construction
- Each utterance is treated as an EDU, and a parser predicts discourse links and relation types to form a graph over utterances.The parser is pretrained on a human-annotated multiparty dialogue corpus, then used to predict relations in SAMSum conversations.Source: §3.1 · p. 4
- First- and second-person utterances are rewritten using speaker names, then OpenIE extracts WHO-DOING-WHAT triples whose arguments become connected action-graph nodes.The method also replaces third-person pronouns using conversation coreference clusters.Source: §3.1 · p. 4
3.2.1 Utterance Encoder
- The utterance encoder starts from BART-base and maps each utterance’s tokens, including a leading special token representing the utterance, into hidden representations.The special token is added at the beginning of each utterance.Source: §3.2.1 · p. 4
3.2.2 Graph Encoder
- The discourse graph initializes each node from its utterance representation, while action-graph argument nodes are initialized by averaging their token embeddings; graph attention then encodes the nodes.Discourse relations are encoded with a one-hot embedding layer.Source: §3.2.2 · pp. 4–5
3.3 Multi-Granularity Decoder
- Each decoder layer attends to utterance tokens and, in parallel, to encoded discourse- and action-graph nodes, then combines the graph-attended vectors into a structure-aware representation.The decoder adds discourse and action cross-attentions to the pretrained BART decoder.Source: §3.3 · p. 5
- A trainable ReZero parameter scales graph-attention residual updates to mitigate their early negative impact on pretrained BART; training uses cross-entropy with teacher forcing.The authors state that ReZero also accelerates learning of newly introduced modules.Source: §3.3 · p. 5
4 Experiments
The experiments evaluate S-BART on daily conversations and debate summaries, comparing discourse and action graphs with existing models and testing graph design choices. S-BART improves automatic and human evaluation results, including in zero-shot transfer, while gains vary with graph structure and conversation complexity.
4.1 Datasets
- SAMSum covers messenger-style conversations about daily topics, while ADSC provides debate summaries for testing generalization.The ADSC model was evaluated zero-shot after training on SAMSum.Source: §4.1 · p. 5, §4.5 · p. 7
4.4 Results on In-Domain Corpus
- 2.0% on ROUGE-1, 4.3% on ROUGE-2, and 1.2% on ROUGE-L over BART were the gains from combining discourse and action graphs on SAMSum.The combined model also outperformed prior state-of-the-art methods; Table 2 reports scores averaged over three runs and a significant improvement over BART (p < 0.05).Source: Table 2 · p. 6, §4.4 · p. 7
4.2 Baselines; 4.3 Implementation Details
- Across all experiments, the sequence-to-sequence model used original BART encoder/decoder default settings; graph encoders had 768 hidden dimensions and two layers, and ReZero residual weights α were initialized to 1.Source: §4.3 · p. 6
4.4 Results on In-Domain Corpus
- S-BART variants received significantly higher human ratings than BART for factualness, succinctness, and informativeness.For factualness, S-BART with action graphs improved by 0.27 over BART.Source: Table 4 · p. 6, §4.4 · p. 7
4.5 Results on Out-Of-Domain Corpus
- 66.2% on ROUGE-1, 373.4% on ROUGE-2, and 82.2% on ROUGE-L over BART were the gains for combined graphs in zero-shot ADSC evaluation.Single-graph models also improved ROUGE over BART; action graphs yielded larger gains than discourse graphs in this domain shift.Source: Table 3 · p. 6, §4.5 · p. 7
4.6 Ablation Studies
- Constructed discourse graphs outperformed random graphs with the same number of edges, and parallel graph attention outperformed sequential combination.Sequential strategies did not improve over single-graph S-BART models.Source: Table 5 · p. 7, Table 6 · p. 7, §4.6 · p. 8
- Action graphs received higher trained ReZero weights than discourse graphs, while using both graphs yielded still higher weights.Initializing α at 1 produced larger final weights and more performance gains than initializing it at 0.Source: §4.6 · p. 8, Figure 3 · p. 8
4.7 Error Analyses
- S-BART matched BART on simpler conversations and outperformed it as discourse and action structures grew more complex, but both struggled with very high interaction counts.The analysis grouped conversations by relative ROUGE performance and marked challenging examples where both systems had low scores.Source: §4.7 · p. 8
5 Conclusion
The paper concludes that modeling discourse relations and speaker-action connections helps abstractive conversation summarization. Experiments and ablations on SAMSum and zero-shot results on ADSC support the effectiveness and generalizability of the approach.
5 Conclusion
- The proposed sequence-to-sequence model incorporates discourse relations between utterances and connections between speakers and actions.The conclusion describes support for these structures through quantitative and qualitative evaluation.Source: §5 · p. 8
- Experiments and ablations on SAMSum support the effectiveness of the structured graphs for conversation summarization.The conclusion refers to both quantitative and qualitative evaluation metrics.Source: §5 · p. 8
- Zero-shot results on ADSC further support generalization to debate summarization.The paper identifies emails, debates, podcasts, longer utterances, and more participants as future application settings.Source: §5 · p. 9
A Discourse Relation Distributions
The appendix describes discourse-relation labeling and the differing relation distributions in daily conversations and debates.
A Discourse Relation Distributions
- A deep sequential model pretrained on STAC was used for link prediction and relation classification in SAMSum and ADSC.STAC contains 1,062 dialogues.Source: A Discourse Relation Distributions · p. 12
- Daily conversations mainly contain Comment, Clarification, and QA-pair relations, while debates also prominently contain Contrast.The listed major debate relations are Comment, Contrast, Clarification, and QA pairs.Source: A Discourse Relation Distributions · p. 12
B Impact of Different ReZero Weight Initializations
The appendix reports testing structure-aware BART with two ReZero weight initializations.
B Impact of Different ReZero Weight Initializations
- The experiments tested structure-aware BART (S-BART) with discourse/action structure under two ReZero settings.Source: B Impact of Different ReZero Weight… · p. 12