Span-Level Dialogue Summarization via Distant Supervision and Machine Reading Comprehension (DSMRC‑S)

1. Dialogue Summarization Background

With the rapid growth of textual data on the Internet, information overload has become a serious problem. Summarization is an essential technique to reduce the dimensionality of text. Dialogue summarization, a special case of text summarization, focuses on extracting key information from multi‑turn conversations such as customer‑service calls, meetings, or chats.

2. Classic Text and Dialogue Summarization Models

2.1 Extractive Models

Lead‑3 : selects the first three sentences of a document as the summary.

TextRank : builds a graph of sentences and applies a PageRank‑like algorithm to rank them.

Clustering : encodes sentences (e.g., Skip‑Thought, Paragram) and clusters them with K‑means or Mean‑Shift, then picks the centroid‑closest sentences.

Neural Extractive Models : include sequence‑labeling approaches (e.g., SummaRuNNer) and sentence‑ranking approaches (e.g., NeuSUM). SummaRuNNer uses a Bi‑GRU to obtain word‑ and sentence‑level representations and predicts a binary label for each sentence. NeuSUM adds a sentence‑benefit score and a unidirectional GRU to model inter‑sentence dependencies.

2.2 Generative Models

Seq2Seq models have become the backbone of abstractive summarization. Recent work augments Seq2Seq with pre‑trained language models such as BERT.

Pointer‑Generator : combines attention‑based generation with a copy mechanism and coverage loss to alleviate OOV and repetition problems.

Leader‑Writer : a hierarchical Transformer where the Leader predicts a sequence of key points (e.g., background, conclusion) and the Writer generates the final summary conditioned on those points.

3. Span‑Level Extractive Summarization (DSMRC‑S)

3.1 Motivation

In Meituan’s customer‑service scenario, agents manually write call summaries, which is time‑consuming. Existing models either suffer from instability (generative) or require sentence‑level labels (extractive). DSMRC‑S converts dialogue summarization into a machine‑reading‑comprehension (MRC) task, requiring no extra annotation.

3.2 Method

Two‑stage pipeline:

**Stage 1 – Token‑level supervision**: Each token in the dialogue is automatically labeled 1 if it appears in the human‑written summary, otherwise 0. A BERT‑based encoder predicts the probability of each token being part of the answer. The loss is a binary cross‑entropy over the token predictions.

**Stage 2 – Density‑based span selection**: For any candidate span, density = (sum of token probabilities) / (span length). The algorithm enumerates all spans that do not cross speaker boundaries and selects the one with the highest density as the answer for each predefined question (e.g., user background, user request, solution).

3.3 Experiments

Dataset: 400 k real Meituan dialogues, each annotated with four key elements (background, request, solution, etc.).

Metrics: BLEU, ROUGE‑L (F1), and Distinct‑1.

Results: DSMRC‑S outperforms all baselines (including S2S+Att, Pointer‑Generator, Leader‑Writer, TDS‑SATM) by ~3 % on BLEU and ROUGE‑L and by 3.9 % on Distinct‑1. It also shows consistent gains across different key elements, dialogue turns, and summary lengths.

4. Conclusion and Future Directions

The paper first reviews classic extractive and generative summarization methods, then introduces a distance‑supervised, span‑level extractive approach that achieves state‑of‑the‑art performance on real‑world dialogue data. Future work includes multi‑span answer extraction, prompt‑based generative dialogue summarization, and deeper modeling of dialogue structure.