💬 NLP

Discourse Coherence and Response-Guided Context Rewriting for Multi-Party Dialogue Generation

This paper introduces DRCR, a novel framework that enhances multi-party dialogue generation by employing discourse coherence and response-guided context rewriting, supported by a dynamic self-evolution learning method to iteratively improve both context understanding and response quality.

Original authors: Zhiyu Cao, Peifeng Li, Qiaoming Zhu

Published 2026-04-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhiyu Cao, Peifeng Li, Qiaoming Zhu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are sitting in a crowded coffee shop with three friends, all talking at once. One friend mentions "the blue one," another says "do that," and a third jumps in with a half-finished sentence. It's a chaotic mess. If you tried to write a summary of this conversation or reply to a specific question, you'd probably get confused. You might not know who "the blue one" refers to, or you might miss the fact that someone already suggested a solution five minutes ago.

This is exactly the problem computers face when trying to generate responses in multi-party dialogues (conversations with more than two people). The paper you shared, DRCR, proposes a clever new way to fix this.

Here is the paper explained in simple terms, using some everyday analogies.

The Problem: The "Broken Telephone" of Group Chats

In a normal two-person chat, it's easy to follow the thread. But in a group chat, people use slang, skip words, and refer to things vaguely ("that," "it," "the thing").

The Issue: When a computer tries to understand a messy group chat, it gets lost. It might think Person A is talking to Person B when they are actually talking to Person C. Because the computer is confused about the structure, its replies are often nonsensical or repetitive.
The Old Way: Previous methods tried to force the computer to memorize the "structure" of the chat (who talked to whom) like a rigid flowchart. But if the chat is messy, the flowchart breaks.

The Solution: The "Editor and the Actor"

The authors propose a system called DRCR. Think of it as a two-step process involving two AI agents working together:

The Editor (The Rewriter): Before the computer tries to answer a question, this AI acts like a strict editor. It looks at the messy, colloquial group chat and rewrites it into a clear, logical story.
- Analogy: Imagine a messy transcript of a meeting. The Editor goes in and changes "Hey, do that thing with the blue one" to "Speaker 2 is asking Speaker 1 to backup the Windows partition using the second hard drive."
- Goal: Make the context so clear that anyone (or any AI) can instantly understand who is talking to whom and what is being discussed.
The Actor (The Responder): Once the Editor has cleaned up the script, the Actor reads this clear version and generates the perfect response.
- Goal: Because the script is now clear, the Actor doesn't have to guess. It can give a helpful, coherent answer.

The Secret Sauce: "Mutual Self-Evolution"

Here is the really cool part. How do you teach these two AIs to get better without hiring a human to grade every single answer?

The authors created a training loop where the Editor and the Actor teach each other.

The Analogy: Imagine a Drama Rehearsal.
- The Editor rewrites a messy scene to make it clearer.
- The Actor performs the scene based on that rewrite.
- They both look at the result: "Did the rewrite help the actor perform better? Did the actor's performance show that the rewrite was good?"
- If the rewrite helped the actor give a great line, they both get a "thumbs up" and learn to do it again. If the rewrite was confusing and the actor stumbled, they get a "thumbs down" and learn to try a different approach.
- They keep doing this over and over, getting better and better at each other's jobs without needing a human teacher. This is called Mutual Self-Evolution.

The Two "Feedback Signals"

To know if they are doing a good job, the system uses two specific "grades":

Discourse Coherence (The "Logic" Grade): Does the rewritten chat make sense? Can we easily tell who is talking to whom? The system checks this by seeing if it can correctly guess who the next speaker is addressing.
Response Quality (The "Helpfulness" Grade): Did the rewritten chat lead to a better answer? If the answer is good, the rewrite was successful.

The system balances these two grades. Sometimes a chat is very logical but the answer is boring; sometimes the answer is great but the chat is confusing. The system learns to find the perfect balance.

Why This Matters

The paper tested this on four different datasets (like real group chats from Ubuntu forums and TV shows). The results showed that:

It works: The new method generated much better responses than previous methods.
It's flexible: It works with different types of AI models.
It's smart: By cleaning up the conversation first, the AI stops making silly mistakes like ignoring what someone said five minutes ago.

The Catch (Limitations)

The authors admit two small downsides:

It needs a strong teacher to start: To begin the "rehearsal," they need a very smart AI to create the first set of examples.
It takes time: Rewriting the chat before answering takes a little extra computing power, though it's worth it for the quality.

Summary

DRCR is like hiring a professional Editor to clean up a messy group conversation before asking an Actor to reply. By having the Editor and Actor practice together in a loop, they learn to understand complex, messy group chats perfectly, resulting in responses that are logical, helpful, and actually make sense in the context of the whole group.

1. Problem Statement

Multi-Party Dialogue Generation (MDG) aims to generate responses in conversations involving multiple speakers. Unlike two-party dialogues, MDG features complex discourse structures where utterances can address non-adjacent speakers, creating intricate dependency chains.

Key Challenges:

Colloquialism and Ambiguity: Real-world multi-party dialogues often contain informal expressions, ellipsis (omissions), and coreference (e.g., "that," "it") that obscure the underlying discourse structure.
Structural Fidelity: Existing methods (both explicit graph-based and implicit latent variable approaches) struggle to encode dialogue structure accurately when the input context is incoherent or ambiguous. This leads to poor response generation, such as addressing the wrong speaker or ignoring previous context.
Lack of Context Optimization: Previous research focused on encoding existing structure but neglected the potential of rewriting the context to improve coherence before generating a response.

2. Methodology: DRCR Framework

The authors propose DRCR (Discourse coherence and Response-guided Context Rewriting), a framework that improves MDG by rewriting the dialogue context to enhance structural clarity. The framework operates in three main stages:

A. Addressee Recognition (AR) as a Coherence Proxy

To quantify discourse coherence without expensive manual annotation, the authors train an Addressee Recognition Classifier.

Rationale: If a model can easily identify who an utterance is addressed to, the discourse is coherent.
Implementation: A RoBERTa-large encoder processes the dialogue context. A biaffine attention mechanism calculates the probability of speaker $j$ being the addressee of utterance $i$ . The average probability of correct addressee identification serves as the Discourse Coherence Score.

B. Preference Data Construction (Warm-up Phase)

To initialize the Rewriter and Responder models, the authors construct preference data using an external Large Language Model (LLM) as a "teacher."

Tree-based Sampling: A top-down sampling method generates multiple candidate rewritten contexts for each dialogue turn.
Semantic Fidelity: Natural Language Inference (NLI) filtering prunes candidates that deviate significantly from the original meaning.
Dual-Feedback Scoring: Candidates are scored on two dimensions:
1. Discourse Coherence ( $C$ ): Measured by the AR classifier's confidence.
2. Response Quality ( $R$ ): Measured by the similarity (BLEU/ROUGE) between the response generated from the rewritten context and the ground truth.
Importance Calibration: A Coefficient of Variation (CoV) is used to dynamically weight $C$ and $R$ . This ensures that if coherence is already high, the model focuses more on response quality, and vice versa.
Training: The highest and lowest scoring samples form preference pairs (Chosen vs. Rejected) to train the Rewriter and Responder using Direct Preference Optimization (DPO).

C. Mutual Self-Evolution (MSE)

To reduce reliance on static external data and adapt to the model's evolving capabilities, DRCR employs an iterative self-evolution loop:

The current Rewriter generates candidate contexts.
The current Responder generates responses for these contexts.
The system scores the new candidates using the internal coherence and response metrics.
Context Replacement: If a rewritten context scores higher than the original, the original context is replaced.
Data Update: New preference data is generated from the current models' outputs to retrain both the Rewriter and Responder.
This loop repeats, allowing the models to continuously refine their alignment and rewriting strategies.

3. Key Contributions

Novel Framework (DRCR): The first method to explicitly incorporate context rewriting to enhance multi-party dialogue generation, addressing the issue of incoherent input structures.
Dual-Feedback Mechanism: Introduces a dual-signal optimization strategy using Discourse Coherence (structural clarity) and Response Quality (generation fidelity) to guide the rewriting process.
Mutual Self-Evolution: Proposes an iterative learning loop where the rewriter and responder collaborate to generate their own preference data, reducing dependence on external static datasets and mitigating alignment biases.
Coherence Proxy: Demonstrates that Addressee Recognition accuracy is a quantifiable and effective proxy for measuring discourse coherence.

4. Experimental Results

The authors evaluated DRCR on four multi-party dialogue datasets: Ubuntu IRC-16, Ubuntu IRC-19, HLA-Chat++, and Friends.

Main Results: DRCR significantly outperformed state-of-the-art baselines (including GSN, HeterMPC, SS-MPC, and fine-tuned LLMs like Llama-3.2 and Qwen3) across all metrics (BLEU, ROUGE, METEOR).
- Example: On Ubuntu IRC-19, Qwen3-8B+DRCR achieved a BLEU-1 of 17.81, surpassing the SFT baseline (14.22) and the previous SOTA SS-MPC (15.60).
Ablation Studies:
- Removing either Discourse Coherence or Response Quality feedback led to performance degradation, proving both signals are essential.
- Removing the CoV calibration (dynamic weighting) reduced performance, highlighting the need to balance the two signals.
- Removing the "Warm-up" phase caused a significant drop, indicating the necessity of initializing model behavior with high-quality external data.
Coherence Analysis: Post-rewriting, the accuracy of addressee recognition increased significantly (e.g., from 75.32% to 84.16% on IRC-16), confirming that rewriting effectively clarifies discourse structure.
Human Evaluation: DRCR outperformed all baselines in Coherence, Fluency, Informativeness, and Helpfulness, achieving scores nearly on par with human-annotated ground truth.

5. Significance and Impact

Paradigm Shift: Moves beyond simply encoding dialogue structure to actively repairing it. By rewriting colloquial or ambiguous inputs into clearer forms, the model can better understand complex multi-party interactions.
Generalizability: The method is model-agnostic and effective across different LLM backbones (Llama, Qwen) and domains (technical support, personalized chat, TV show transcripts).
Efficiency: While rewriting adds some inference time, the Mutual Self-Evolution mechanism reduces the long-term cost of relying on expensive external LLMs for preference data construction.
Future Direction: The paper highlights that selective rewriting (rewriting only necessary utterances) is a promising area for future optimization to further reduce latency.

In conclusion, DRCR demonstrates that improving the structural coherence of the input context through guided rewriting is a critical, yet previously overlooked, factor in achieving high-quality multi-party dialogue generation.