《Interactive Classification by Asking Informative Questions》读后感

目录

一、文章阅读

【更多论文阅读新进展欢迎关注统计人贞子微信公众号】

1 摘要和总结

摘要部分
  • where users provide an initial query using natural language, and the system asks for additional information using binary or multi-choice questions.本文的交互方式采用的是自然语言问答的形式,机器对用户的提问方式是二选一或者多选一的模式。主要针对的是“分类问题”。
  • 提出了“继续发问”和“返回结果”两种问题之间权衡的一种决策系统(使用众包任务代替实时交互?)
  • 评估方法,评测上面的停止发问系统的好处
总结部分
  • 专家指导的问题和答案增量设计——可扩展性差
  • We demonstrate that the system can be bootstrapped without any interaction data and show effectiveness on two tasks.用非实时交互数据进行训练

2 Introduction介绍

  • 简单介绍了单轮问答的进展阻碍——(用户对领域本身不理解&系统对用户的问题理解不正确)
  • 本文研究目的(评估interaction的好处&交互数据获得的成本和复杂性控制)——交互系统提出一系列二项和多项选择题
    《Interactive Classification by Asking Informative Questions》读后感
  • 设计的方法不依赖于学习过程中的用户交互(不依赖于垃圾的交互系统和被导向的行为(引导回答))
  • 后验分布的贝叶斯分解(对意图标签和用户响应),使用后验来计算问题的预期信息增益平衡准确性和新一轮交互成本
  • 非交互式注释任务
  • We evaluate with both a simulator and human users.

3 Technical Overview

  • Goal:to classify a natural language query to a label through an interaction.
  • 《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感

《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感

  • 模型选择具有最大信息增益的问题。给定用户响应,模型将更新对分类标签的置信度。
  • 根据yi初始化x,用户给yi加标签,根据用户加的标签生成{q,r}
    《Interactive Classification by Asking Informative Questions》读后感

《Interactive Classification by Asking Informative Questions》读后感

4 Related Work

放在参考文献那里

  • 使用分类目标,问题和答案的自然语言描述来计算我们的分布,而不是将它们视为分类数据或结构数据。

  • 使用联合分布的贝叶斯分解,可以将其轻松扩展到其他模型驱动的选择方法。

  • (某文)用户将其与真实图像进行比较,并使用相关性得分或描述它们之间差异的自然语言*提供反馈

5 Method(重要***)

《Interactive Classification by Asking Informative Questions》读后感

  • 两个假定【(1)r只取决于q和y,与过去的互动无关。】【qt很大程度取决于xt-1】X是交互

  • 《Interactive Classification by Asking Informative Questions》读后感

  • 《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感

  • 这种分解使得能够利用单独的注释来直接学习这两个组件,从而减少了收集昂贵的用户交互记录的需求。
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感

  • 我们不只是将标签,问题和答案视为分类变量。利用它们的自然语言内容来估计它们的相关性。这减少了对大量注释的需求,并改善了资源匮乏情况下的模型。Similarity:

  • 《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    average across annotators to estimate:
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    《Interactive Classification by Asking Informative Questions》读后感
    Loss
    《Interactive Classification by Asking Informative Questions》读后感

《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感
用户模拟器:
《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感
策略选择器:(策略梯度反向传播确定参数)
《Interactive Classification by Asking Informative Questions》读后感

  • 奖励功能为在交互结束时预测正确的目标提供了正向奖励,为预测错误的目标提供了负向奖励,并为每个提出的问题提供了少量的负向奖励。

6 Data Collection

  • Amazon Mechanical Turk发布众包任务
  • 收集每个FAQ文档的初始查询和标签。附录A.1描述了工人培训过程。
  • Initial query的收集(X):给定目标FAQ(yi),我们要求他们提供给这样的系统的初始查询
  • FAQ tag生成:标签不限于预定义的本体,可以是描述文档主题的短语或单个单词————convert tags into questions(q)
  • 将q和y关联起来(通过回答r)————人工、加上S(·)提供初步筛选
    《Interactive Classification by Asking Informative Questions》读后感

7 Experimental Setup

  • 无交互:仅使用初始查询来预测分类标签。我们考虑以下四个实现
    • BM25:一种基于关键字的通用检索方法评分模型
    • RoBERTaBASE:微调的该模型编码文本
    • 简单的RNN+fastText单词嵌入层
    • RNN + self-attention
  • Random Interaction:随机选择某个T的问题进行交互。
  • No Initial Query Interaction:不会使用最大信息标准来限制初始用户查询。
    《Interactive Classification by Asking Informative Questions》读后感
  • we use one encoder for user initial queries and question-answer pairs and a second encoder for bird names.

8 Results

用户评分:
《Interactive Classification by Asking Informative Questions》读后感
accuracy
《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感
Figure 3: [email protected] (y-axis) against turns of interactions (x-axis) for FAQ (left) and Birds (right) tasks
《Interactive Classification by Asking Informative Questions》读后感

where x-axis is the number of episodes (400 trials per episode). The results are compared on different turn penalty ra.

《Interactive Classification by Asking Informative Questions》读后感

二、参考文献选读

《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感
《Interactive Classification by Asking Informative Questions》读后感