Learning to Select Knowledge for Response Generation in Dialog Systems (1902.04911v2)

Abstract: End-to-end neural models for intelligent dialogue systems suffer from the problem of generating uninformative responses. Various methods were proposed to generate more informative responses by leveraging external knowledge. However, few previous work has focused on selecting appropriate knowledge in the learning process. The inappropriate selection of knowledge could prohibit the model from learning to make full use of the knowledge. Motivated by this, we propose an end-to-end neural model which employs a novel knowledge selection mechanism where both prior and posterior distributions over knowledge are used to facilitate knowledge selection. Specifically, a posterior distribution over knowledge is inferred from both utterances and responses, and it ensures the appropriate selection of knowledge during the training process. Meanwhile, a prior distribution, which is inferred from utterances only, is used to approximate the posterior distribution so that appropriate knowledge can be selected even without responses during the inference process. Compared with the previous work, our model can better incorporate appropriate knowledge in response generation. Experiments on both automatic and human evaluation verify the superiority of our model over previous baselines.

Learning to Select Knowledge for Response Generation in Dialog Systems

The paper "Learning to Select Knowledge for Response Generation in Dialog Systems" introduces a novel approach to address the issue of generating uninformative responses in dialogue systems by employing a more effective knowledge selection mechanism. The authors propose an end-to-end neural model that incorporates both prior and posterior distributions over external knowledge to enhance the informativeness of responses generated by dialogue systems.

The primary challenge addressed in this paper is the selection of appropriate knowledge from a pool of potential information sources for response generation. Traditional approaches generally rely on semantic similarity between input utterances and available knowledge, often treating this as a prior distribution without considering the actual knowledge used in the response. This method can lead to a substantial discrepancy between prior and posterior knowledge distributions, imposing difficulties in selecting suitable knowledge and thereby limiting the quality of response generation.

This research introduces a distinctive framework where a posterior distribution over knowledge is inferred considering both the dialogue utterances and responses. This framework facilitates a more informed knowledge selection process during training, leveraging the responses to improve the accuracy of knowledge selection. Simultaneously, a prior distribution is utilized, which approximates the posterior to allow accurate knowledge selection during inference when responses are unavailable.

The model architecture consists of various components: utterance encoder, knowledge encoder, knowledge manager, and decoder. The knowledge manager plays a crucial role by employing both prior and posterior information to aid in accurate knowledge sampling. Sampling is achieved via Gumbel-Softmax re-parameterization to maintain differentiability during model training.

Empirically, the model demonstrates substantial improvements over existing methods, as verified through experiments on the Persona-chat and Wizard-of-Wikipedia datasets. Key results indicate that the proposed model outperforms popular Seq2Seq and memory network-based baselines in terms of BLEU scores and Distinct metrics, showcasing its enhanced capacity for generating diverse and informative responses. Importantly, the research highlights surges in knowledge recall, precision, and F1 metrics, signifying improved alignment between selected knowledge and generated responses.

The integration of the knowledge selection mechanism into the Lost in Conversation Transformer model further exemplifies its versatility, yielding remarkable improvements in automatic evaluation metrics when applied across varied datasets.

The implications of this research are manifold. Practically, the method proposed provides a more robust mechanism for enriching dialogue systems with relevant, context-aware knowledge, potentially improving user engagement and satisfaction in conversational AI applications. Theoretically, separating prior and posterior distributions over knowledge enhances the understanding of the role these distributions play in grounded dialogue generation, thus paving the way for future advancements in response generation models.

Future directions, as suggested by the authors, include extending this mechanism for multi-turn conversations, which encompasses broader contexts and presents more complex challenges in dialogue settings. This research contributes significantly to the ongoing endeavor to improve the intelligence and interactivity of dialogue systems through more methodical and informed knowledge integration strategies.