Notes on the literature review of the dialogue system

The original address of the paper: Https://arxiv.org/pdf/1711.01731.pdf I. Introduction

This paper summarizes the dialogue system and discusses the possible future research directions.

The dialogue system is broadly divided into two categories:
(1) Task-oriented dialogue system
Task-oriented dialogue system is to help users to complete specific tasks, such as looking for goods, booking accommodation, booking restaurants and so on. There are two main ways to achieve a task-oriented dialogue system: (1) Pipeline method, (2) End-to-end method.

(2) Non-mission-oriented dialogue system
The non-mission-oriented dialogue system is to interact with the user and provide answers, simply speaking, in the open field of gossip. The realization of non-task-oriented dialogue system can be divided into two main categories: (1) Generative method, (2) Retrived-based method.

In this paper, the above two types of dialogue systems, as well as their respective implementation methods are reviewed and explained in detail. II. Mission-oriented dialogue system Pipeline Method

Pipeline method steps can be divided into 4, the process as shown below, detailed description please continue to read the following 1.Natrual Language Understanding (NLU)

Convert user input statements into pre-defined semantic slots (semantic slots)

Let's take a first example:
In the task-based dialogue system, the user wants to check the hotel information, and then say a sentence: "Show restaurant at New York tomorrow."
It takes two steps to understand this sentence:
(1) First to determine the user is required to book hotels, rather than booking tickets, shopping, check Express, then this is a classification problem, that is, the identification of user intent category
(2) The hotel category will have a pre-set semantic slot (semantic slot) corresponding to it, such as New York is the slot value of location. The process of filling the slot values is to extract the word information in the sentence.

The above two steps can also be called intent recognition (intent detection) and slot filling (slot filling), respectively:
intent to identify:
Classification issues, classifying user-issued statements into pre-defined intent categories.

At present, deep learning techniques are successfully applied in the recognition of intentions [15;84;112]. [25] using convolutional neural networks to extract vector representations of query statements (vector representations). [29] and [74] there are similar applications based on the CNN classification framework

Slot Fill:
Sequence labeling issues, which are semantic tags for each word in the sentence. The input is a sentence consisting of a sequence of words, and the output is the semantic category (Slot/concept ID) of the word and word corresponding to the sequence of the group.

[17] and [15] the use of deep belief networ (DBNs), achieved better than CRF baseline effect;
[51;115;66;113] uses RNN. 2.Dialogue State Tracking

Manages the input of each round of conversations based on the history of the conversation and outputs the status of the current conversation. The dialog state HT represents the characterization of a conversation on a time t (also known as a slot or semantic frame).

Traditional way: Use manual rules to choose the most likely results. But the error rate is high.
Statistical-based dialogue system: the probability distribution of each slot is calculated for each round of conversations. The literature has robust sets of hand-crafted rules[93],conditional random Fields[26;25;63],maximun entropy models[98], WEB-STYLE-RANKING[101]
Based on deep learning: [26] The application of deep learning in faith tracking is introduced. On a domain can be migrated to the new domain, [58] proposed Multi-domain RNN dialog state tracking models; [59] Neural Belief Tracker (NBT) was proposed to identify the slot-value pair. 3.Policay Learning

Make the next reaction based on the current conversation state. For example, in the online shopping scene, if the conversation status identified in the previous step is "recommendation", then this step will give the recommended action, that is, get the product from the database.

Supervised learning can be applied: first a rule-based agent is used to do a hot start, and then supervised learning is performed on the action generated by rule [111].
End-to-end reinforcement learning can also be applied: [14]

4.Natural language Generation (NLG)
The response given by policy learning is translated into the corresponding natural language form of the answer provided to the user. A good answer generator should have 4 features: adequacy, fluent,readability and variation[78].

traditional method [90;79]: Use sentence planning to convert the input semantic features into intermediate forms (such as tree or template form) and then convert the intermediate form to the final answer via surface realization.
Deep learning methods: [94;95] describes the structure of NN in lstm, similar to rnnlm;[94] using the forward RNN generator, as well as the backwards of CNN and RNN Reranker, all sub-models for joint optimization; [95;83] Added a control celll to gate the Dialogue Act; [96] The former is improved in the application of the Multiple-domain, [123] The use of Encode-decode lstm-based, and the combination of attention machanism; [20] Use the Sequence-to-sequence method. End-to-end Methods

Features: Using a single model, interacting with an external structure database

[97;7] introduces a network-based, end-to-end, and trained conversation system. Convert the problem of the dialog system to learning a map (answer from the historical dialogue –>). However, it requires a lot of training data and lacks robustness.
[120] For the first time, End-to-end intensified learning methods were demonstrated, and joint training dialogues were followed by strategic learning.
[45] Training End-to-end model as a task completion nueral Gialogue system.

The task dialog system requires external Knowledge base support.
The earlier approach was to parse the query statement semantically and symbolize it and get the object from the knowledge base based on the symbolic query statement [97,103,45]. However, there are two drawbacks: The obtained result does not carry the uncertainty information of semantic parsing; The statement parsing in the dialog policy needs to be trained separately.
The improved practice is: [21] using RNN and combining the attention mechanism; [18] Substituting symbolic query statements as "soft" posterior disttribution; [102] combined with RNN in domain-specific knowledge, and encode for templates.