论文笔记 | An End-to-End Deep Framework for Answer Triggering with a Novel Group-Level Objective

作者:李瑾
单位:燕山大学


论文地址:https://www.aclweb.org/anthology/D17-1131.pdf
代码地址:https://github.com/jiez-osu/answer-triggering


目录

  1. 研究背景
  2. 研究方法
    2.1 模型框架
    2.2 组级目标函数
  3. 实验与评估
    3.1 数据集
    3.2 实验结果
  4. 总结

1. 研究背景

在现实生活和实际的问答中,并不是所有的问题都存在有效的可以充分支持所提出的问题的候选答案,因此在2015年WiKiQA数据集发布的同时首次提出了答案触发这一任务。

答案触发任务可以划分为两个子任务:一是构建模型对候选答案排序,使得正确答案获得最高的分数;二是对候选答案中是否存在正确答案作出预测。

先前工作通过pipline的方法来解决这一问题,但得到的模型并不能令人满意,F1值仅仅在32%到36%之间。本文提出了端到端的深度应答触发框架。

2. 研究方法

本文提出了使用端到端的深度模型框架并提出了新的目标函数惩罚答案触发中存在

2.1 模型框架

模型由三个部分组成:编码层、问答匹配层和最大池化层。

其中:

  • 编码层可以根据所要达成的目标使用不同的网络框架
  • 问答匹配层将问题与候选答案拼接喂入前向神经网络
  • 在max pooling 层作者将同一答案的候选集中预测标签为正和负的答案划分为互不相交的两个子集
2.2 组级目标函数

作者提出了新的目标函数惩罚机制来促使模型作出正确的选择。

  • 对于每个预测标签为负的样本集,最高分数的错误答案将受到铰链惩罚(hinge loss)。公式如下:
    O 1 = 1 N n e g ∑ i : l i = 0 m a x ( 0 , d − − ( 0.5 − m i − ) ) O_{1}=\frac{1}{N_{neg}}\sum_{i:l_{i}=0}^{}max(0,d^{-}-(0.5 - m_{i}^{-})) O1=Nneg1i:li=0<
  • 0
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 1
    评论
Dynamic event-triggered control (DETC) is a control strategy that reduces the number of times control actions are executed by triggering them only when necessary. This approach can save computational resources and improve the efficiency of control systems. In the context of discrete Markov jump systems (DMJS), DETC has been shown to be effective in stabilizing the system and ensuring its performance. DMJS are a class of systems that undergo discrete transitions between different operating modes, each characterized by a different set of dynamics. The transitions between modes are stochastic and determined by Markov processes. The goal of control in DMJS is to stabilize the system and ensure its performance despite the stochastic nature of the transitions. DETC for DMJS involves triggering control actions based on the system's state and the occurrence of certain events. The events can be defined based on the system's behavior or performance specifications. When an event occurs, the control action is executed, and the system transitions to a new mode. The triggering of the control action is based on a dynamic threshold that takes into account the system's state and the event occurrence. DETC has several advantages over traditional time-triggered control strategies in DMJS. It can reduce the number of control actions and improve the system's efficiency. It can also ensure the system's stability and performance in the presence of stochastic transitions. However, the design and implementation of DETC for DMJS can be challenging due to the complex nature of the system's dynamics and the stochastic transitions. Overall, DETC is a promising control strategy for DMJS that can improve the efficiency and performance of control systems. Further research is needed to develop more effective and robust DETC algorithms for DMJS.

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论 1
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值