Work

Open-domain dialogue generation with deep learning (本文)

Year: 2023
Type: dissertation
Abstract: This thesis aims to generate high-quality responses in an open-domain dialogue using deep learning techniques.Three different solutions are proposed for dialogue systems based on their relevance to the context and topic, level of personalization, and ability to provide a good user experience in long-term conversations, respectively.We begin by introducing Anna, a novel dialogue generation model that is composed of a hierarchical self-attention network and a topic-augmented network.During daily conversations, humans typically respond by understanding the dialogue history and assembling their knowledge regarding the topic.However, existing dialogue generation models are weak at capturing the dependencies among words or utterances, resulting in an insufficient understanding of context and the generation of irrelevant responses.Previous works have largely ignored topic information modeling in multi-turn dialogue, making responses overly generic.Anna captures contextual dependencies and assigns greater weight to important words and utterances to compute context representations.We incorporate topic information into our model as prior knowledge to synthesize topic representations.These representations jointly determine the probability distributions of responses, which effectively simulates how people behave in real conversations.Empirical studies demonstrate that Anna outperforms baseline models in terms of response quality, parameter size and decoding speed.We consider human-like conversations to be essentially based on inferring information about the persona of the other party.Therefore, higher levels of personalized responses can enhance user engagement.However, we found that the responses generated from most previous models were self-centered, with little consideration for the user in the dialogue.Additionally, during interactions, Anna did not model the personas of the different parties.Motivated by these observations, we propose an approach to personalized dialogue generation that detects implicit user personas.Because it is difficult to collect a large amount of detailed personal facts for each user, we attempted to model the potential persona of a user and its representation from the dialogue history with no external knowledge.The perception and fader variables were conceived using conditional variational inference.The two latent variables simulate the process of people becoming aware of each other's personas and producing a corresponding expres-
Institution: Keio University
Language: English
Open Access status: green
OpenAlex ID: W7134726079
Cites: 79
Cited by:
Related to: 10