deepdialog

Papers

Natural Language Understanding 自然语言理解

自然语言理解任务(Natural Language Understanding),也被称为SLU(Spoken LU)或者(LU),或语义解码(Semantic Decoding)。

NLU一般分为两个子任务:意图识别(Intent Detection)、槽识别(Slot Filling)。

意图识别是一个分类任务,所以总体思路和一般的文本分类(Text Classification)或者短文本分类(Short TC)的解决思路基本一致。

槽识别是一个识别任务,类似抽取任务,所以一般方法和命名实体识别(Named Entity Recognization)任务基本一致。

发展(2019)

主流的发展方向有以下几个:

  1. 联合意图识别和槽识别,放入同一模型,互相增加准确率
  2. 从语音开始,而不是从文字开始,因为对话系统往往基于语音,语音识别本身就带来了很多错误,联合语音信息可以保留更多信息
  3. End-to-End的方式进行Fine-Tune,即不仅仅联合意图识别和槽识别,也同时连接后面的DST、DPL甚至NLG模型,共同训练。

Dialog State Tracking 对话状态跟踪

也叫Belief Tracking

Neural Belief Tracker: Data-Driven Dialogue State Tracking [Nikola Mrksić et al., 2016]

Fully Statistical Neural Belief Tracking [Nikola Mrksić et al., 2018]

In slot-based spoken Dialogue systems, tracking the entities in context can be cast as slot carryover task - only the relevant slots from the dialogue context are carried over to the current turn.

Improving Long Distance Slot Carryover in Spoken Dialogue Systems [Chen et al., 2019]

Dialog Policy 对话策略

Natural Language Generation 自然语言生成

Question Answering System 问答系统

Training Neural Response Selection for Task-Oriented Dialogue Systems [Henderson et al., 2019]

对于简单的FAQ问答,也可以被称为 Question Answering Pairing 或者 Neural Response Selection 问题

One-Shot Learning for Text-to-SQL Generation [Lee et al., 2019]

FAQ Retrieval using Query-Question Similarity and BERT-Based Query-Answer Relevance [Wataru Sakata et al., 2019]

A Context-aware Natural Language Generator for Dialogue Systems [Duˇsek et al., 2016]

Knowledge Based Question Answering 基于知识的问答

End-to-End System 端对端系统