Source-linked AI summary
Abstractive Dialogue Summarization with Sentence-Gated Modeling Optimized by Dialogue Acts
Chih-Wen Goo, Yun-Nung Chen
TL;DR
Prior dialogue summarization emphasized salient content, while dialogue acts offer interactive signals and abstractive meeting summarization lacked suitable benchmark data. The paper builds an AMI benchmark and uses a sentence-gated model to connect dialogue acts with summaries. The model outperforms the state-of-the-art models reported for the task, supporting the usefulness of interactive cues in dialogue summarization.
Problem
Prior summarization focused mainly on single-speaker documents, and abstractive dialogue summarization lacked suitable benchmark data while interactive status was largely overlooked.
Source: §1 · p. 1Method
The paper benchmarks abstractive meeting summarization on AMI and introduces a sentence-gated model that explicitly relates dialogue acts to summaries.
Source: §1 · pp. 1–2, §3.4 · p. 4Results
The proposed approach outperforms all state-of-the-art models in the reported meeting-dataset experiments.
Source: §5 · p. 6Takeaways & Limitations
The experiments demonstrate the usefulness of interactive cues for dialogue summarization.
Source: §4.3 · p. 5, §5 · p. 6
Abstract
from arXiv · showhide
Neural abstractive summarization has been increasingly studied, where the prior work mainly focused on summarizing single-speaker documents (news, scientific publications, etc). In dialogues, there are different interactions between speakers, which are usually defined as dialogue acts. The interactive signals may provide informative cues for better summarizing dialogues. This paper proposes to explicitly leverage dialogue acts in a neural summarization model, where a sentence-gated mechanism is designed for modeling the relationship between dialogue acts and the summary. The experiments show that our proposed model significantly improves the abstractive summarization performance compared to the state-of-the-art baselines on AMI meeting corpus, demonstrating the usefulness of the interactive signal provided by dialogue acts.
1. INTRODUCTION
Prior dialogue summarization emphasized salient content, while interactive behavior remained underexplored; abstractive meeting summarization also lacked suitable benchmark data. This work benchmarks the task on AMI and introduces a sentence-gated model that explicitly connects dialogue acts with summaries.
1. INTRODUCTION
- Prior speech summarization largely used salient content rather than interactive status, despite the importance of inter-speaker interaction for meeting summaries.Earlier dialogue summarization used prosody or speaker information for extractive salience, while abstractive meeting summarization lacked suitable benchmark data.Source: §1 · p. 1
- AMI topic descriptions serve as short, high-level abstractive summaries that may contain no words directly spoken in the meeting.This makes the benchmark task challenging because the summaries do not necessarily overlap lexically with the dialogue.Source: §1 · p. 1
- Dialogue acts and summarization were usually treated independently, and prior discourse-aware summarization modeled between-utterance cues implicitly in structured publications.This paper instead focuses on explicit integration of dialogue acts as interactive signals in neural dialogue summarization.Source: §1 · p. 2
- The proposed model achieves state-of-the-art dialogue summarization performance and uses dialogue acts as explicit interactive signals.The paper also presents a meeting-domain benchmark whose summaries describe high-level meeting goals.Source: §1 · p. 2
2. DIALOGUE SUMMARIZATION DATASET
The paper constructs an abstractive dialogue-summarization benchmark from AMI because the corpus lacks summary annotations. It uses transcripts as input, annotated topic descriptions as summaries, and only samples with available topic and dialogue-act annotations.
2. DIALOGUE SUMMARIZATION DATASET
- AMI transcript samples use annotated topic descriptions as summaries because the corpus has no summary annotations; samples lacking topic or dialogue-act annotations are excluded.The resulting summaries describe the meeting's high-level goal.Source: §2 · p. 2
- The dataset contains 7,824 samples from 36 recordings, split into 7,024 training, 400 development, and 400 test samples.The training set contains 15 dialogue-act labels.Source: §2 · p. 2
- Meetings are divided into samples using a 50-word sliding window, with boundaries adjusted to avoid cutting utterances mid-sentence.When topics change within a window, their descriptions are concatenated in order.Source: §2 · p. 2
3. PROPOSED APPROACH
The approach combines dialogue-history encoding, dialogue-act labeling, and attentional summary generation. A sentence gate integrates dialogue-act and summary information, with full-attention and summary-attention variants trained jointly.
3.1. Dialogue History Encoder
- The dialogue history encoder represents each utterance with an average of its word embeddings and encodes the sentence sequence with a bidirectional LSTM.The resulting hidden states encode the source dialogue.Source: §3.1 · p. 3
3.2. Dialogue Act Labeler
- The dialogue-act labeler predicts an act for each utterance, using attention-weighted hidden states to form dialogue-act context vectors.Those hidden states and context vectors are optimized for dialogue-act modeling.Source: §3.2 · p. 3
3.3. Attentional Summary Decoder
- An attentional decoder generates the summary as a word sequence, initialized with the dialogue encoder’s final hidden state.Generation stops at an end-of-string token or the predefined maximum length.Source: §3.3 · pp. 3–4
3.4. Sentence-Gated Mechanism
- The sentence-gated mechanism uses summary context to model relationships between dialogue acts and summaries, with full-attention and summary-attention variants.The gate is intended to improve both the dialogue-act labeler and the summary decoder.Source: §3.4 · p. 4
3.4.1. Full Attention
- Full attention combines dialogue-act and averaged summary context to compute a gate that weights the dialogue-act context against the current hidden state.The paper interprets a larger gate value as indicating more similar attention to the input and a stronger relationship between the contexts.Source: §3.4.1 · p. 4
3.4.2. Summary Attention
- Summary attention removes dialogue-act attention and shares the attention mechanism between dialogue acts and summaries.The authors describe this as allowing both types of information to be mutually improved more directly than in full attention.Source: §3.4.2 · p. 4
3.5. Joint End-to-End Training
- The model uses a joint conditional objective for dialogue acts and the summary, allowing end-to-end training from the input dialogue.The supplied Table 2 caption identifies AMI performance and significance markers but gives no result values.Source: §3.5 · p. 4, Table 2 · p. 5
4. EXPERIMENTS
Experiments on AMI compare the proposed sentence-gated approach with dialogue-act and summarization baselines. The summary-attention model significantly outperforms all baselines on almost all measurements, while attention analysis highlights Inform utterances.
4.1. Setup
- Experiments use AMI; models are similarly sized, trained with Adam for up to 30 epochs with early stopping, and assessed by dialogue-act accuracy and four ROUGE scores.Reported results are averaged over 20 runs, and significance is tested with a single-tailed t-test at p < 0.05.Source: §4 · p. 5, §4.1 · p. 5
4.2. Baselines
- Baselines span dialogue-act labeling and summarization; only the discourse-aware hierarchical model implicitly uses interaction cues, while the proposed model jointly optimizes summaries with dialogue acts.The compared systems include a bidirectional LSTM labeler, attentional seq2seq, pointer-generator, and discourse-aware hierarchical attentional seq2seq.Source: §4.2 · p. 5
4.3. Results
- Discourse-aware hierarchical seq2seq beats the other summarization baselines; pointer-generator performs worst with 1.2% input-summary overlap, which authors link to copy and coverage noise.The authors interpret the comparison as suggesting that dialogue summarization focuses more on interaction goals than mentioned content.Source: §4.3 · p. 5
- The summary-attention sentence-gated model significantly outperforms all baselines on almost all measurements and is 12% smaller than the best baseline combination.The authors report that the result demonstrates the usefulness of interactive signals and the mechanism’s ability to model their relationship with summaries.Source: §4.3 · p. 5
4.4. Attention Analysis
- In Fig. 4, darker weights indicate greater importance when generating the underlined target word; the analysis finds that Inform utterances usually receive more attention.The authors attribute the learned key-sentence attention to the sentence gate conditioning dialogue acts on the summary.Source: §4.4 · p. 6, Figure 4 · p. 6
5. CONCLUSION
The paper addresses abstractive dialogue summarization by modeling interactive behaviors with a sentence-gate that conditions dialogue-act signals on the learned summary. On a meeting dataset, the approach outperforms state-of-the-art models.
5. CONCLUSION
- The proposed sentence-gated model conditions dialogue-act signals on the learned summary and outperforms state-of-the-art models in experiments on a meeting dataset.The conclusion presents the result as evidence for the importance of interactive cues in dialogue summarization.Source: §5 · p. 6