Dialogue systems are becoming chief human-computer interface systems and natural language processing tools. In recent times, dialogue systems have attracted more research attention. Dialogue system helps to communicate with a human naturally.
Deep learning has been applied in a broad range of big data applications such as computer vision, natural language processing, and recommendation systems. In such a way, deep learning technology tremendously establishes the current advancement in dialogue systems. Deep learning for dialogue systems can influence numerous amounts of data to learn meaningful feature representations and response generation approaches with minimal hand-crafting.
According to the applications, dialogue systems are categorized into task-oriented and non-task-oriented systems. Task-oriented dialogue systems are regarded as a significant branch of spoken dialogue systems. The non-task-oriented dialogue system, also known as chatbots, aims to converse with humans on open fields. The improvement of big data and deep learning technology has highly advanced both task-oriented and non-oriented dialogue systems.
Dialogue systems promote a wide range of applications in business, education, government, healthcare, and entertainment. The knowledge base of customer service agents, Guided selling, Help desk, Website navigation, Technical support, personalized service, training or education, Call centers, and interactive voice response, are notable real-world application tasks of dialogue systems.
• Pretrained Models for Natural Language Understanding - Recently, pretrained language models such as BERT-based and GPT-based systems, Pretrained TOD-BERT, and Span-ConveRT have been applied to task-oriented dialogue systems.
• Domain Transfer for Natural Language Understanding - For domain transfer, multitask training samples from multiple domains apply to learn. This concept helps solve slot name encoding, description encoding, and slot-filling.
• Domain Transfer for Dialogue State Trackers - Domain adaptability is also an important topic for dialogue state trackers. Model-agnostic meta-learning, zero-shot transfer learning, and value normalization are recently applied to resolving domain transfer issues.
• Tracking Efficiency for Dialogue State Trackers - Tracking efficiency is a hot topic in dialogue state tracking. Presently, slot connection mechanism and slot attention to analyzing slot and dialogue context in dialogue state tracking. Multi-Agent Dialog Policy Learning is established to solve the Policy Learning problem environment in dialogue state tracking.
• Response Consistency for Natural Language Generation - For the response consistency problem, methods such as semantic planning gate, gating mechanism, and iterative rectification network were developed.
• End-to-end Task-oriented Dialogue Systems - End-to-end systems are completely data-driven, contributing to their strong and natural responses. It necessary focuses on raising the response quality of end-to-end task-oriented dialogue systems with finite data. Linearized tree-structured representation, reinforcement learning framework, and optimization of learning strategy are recently applied to these topics.
• Retrieval Methods for Task-oriented Dialogue Systems - Retrieval-based methods are less in task-oriented systems for the lack of candidate entries to span all responses.
• Context Awareness - Dialogue context comprises user and system messages. Dialogue context awareness determines the conversation topic and user goal. Several deep learning models have evolved to increase retrieval models context-awareness ability.
• Response Coherence - Response Coherence helps maintain logic and consistency in a dialogue and resembles an interaction process. In retrieval-based systems, responses are naturally coherent. For the generation system, response coherence needs to be considered.
• Response Diversity - Diverse training strategies have been developed to raise response diversity. Recently, entropy-based algorithms, response ranking frameworks, and more methods have been developed for increasing response diversity.
• Empathetic Response - It is necessary to build an empathetic response-based dialogue system concerning emotional changes and sentiments.
• Controllable Generation - In open-domain dialogue systems, controllable dialogue generation needs to design from data sample distributions, causing various uncertain responses.
• Conversation Topic - It is significant to consider the conversation topic in the chatbot. A good conversation topic model assists in retrieving relevant knowledge and guides the conversation rather than passively reacting to the user-s messages.
• Interactive Training - It is human-in-loop training for dialogue systems designed to enhance dialogue system interactions with users.
• Visual Dialogue - There is a need for research focus on visual question answering based upon the content of a picture or video for an open domain dialogue system.