Abstract:Neural response generation can automatically produce replies for open-domain dialogue systems without hand-crafted rules or templates. Current studies follow a non-context-adaptive paradigm that employs a single response generator to deal with all dialogues. However, as a dialogue progresses, its textual characteristics (e.g., context length, information volume, involving topics) are changing, so are the issues challenging its response generation. Non-context-adaptive response generators are inflexible and may fail to achieve globally good performance without considering the differences existing among dialogues. In this paper, we propose a novel framework named as C ontext- A daptive R esponse G eneration ( CARG ), which assembles two different response generators to respectively handle long and short dialogues. Specifically, given a dialogue, CARG first classifies it into short or long types according to the number of its containing utterances. For a short dialogue, CARG employs a sequential reader (SR) to concatenates all utterances into a sequence aiming to construct the dialogue context by limited semantics. For a long dialogue where irrelevant noises and relevant contexts both exist, CARG uses a jumping reader (JR) to generate the response, which treats the latest utterance as the anchor and further performs selective context utilization under its guidance. We introduce ensemble learning strategy to conduct the training and testing of CARG. Extensive experimental results on two benchmark chat corpora show that the proposed CARG framework can outperform various competitive baselines, validating its effectiveness on response generation.

Hard-style Selective Context Utilization for dialogue generation based on what user just said

Towards Efficient Coarse-grained Dialogue Response Selection

Improving Contextual Language Models for Response Retrieval in Multi-Turn Conversation

A Survey on Response Selection for Retrieval-based Dialogues.

How To Make Context More Useful? An Empirical Study On Context-Aware Neural Conversational Models

Deep context modeling for multi-turn response selection in dialogue systems

Sequential or jumping: context-adaptive response generation for open-domain dialogue systems

Dialogue Generation Model with Hierarchical Encoding and Semantic Segmentation of Dialogue Context

Context Directed Speech Recognition in Dialogue Systems

Forward Creation, Reverse Selection: Achieving Highly Pertinent Multimodal Responses in Dialogue Contexts

Context Retrieval via Normalized Contextual Latent Interaction for Conversational Agent

Topic Aware Context Modelling For Dialogue Response Generation

Response Selection from Unstructured Documents for Human-Computer Conversation Systems

Context-aware Multi-Feature Fusion for Open-Domain Dialogue Generation.

Dynamic Working Memory for Context-aware Response Generation

Leveraging Context for Neural Question Generation in Open-domain Dialogue Systems.

A New Dialogue Response Generation Agent for Large Language Models by Asking Questions to Detect User's Intentions

Learning to Detect Relevant Contexts and Knowledge for Response Selection in Retrieval-based Dialogue Systems

Exploring Dense Retrieval for Dialogue Response Selection

Contextual Fine-to-Coarse Distillation for Coarse-grained Response Selection in Open-Domain Conversations

Keep and Select: Improving Hierarchical Context Modeling for Multi-Turn Response Generation