Stance_Detection/AnswerStance

• Question-answer stance detection based on recurrent interactive attention
• • Abstract
  • Model
  • Experimental Results
  • Usage
  • • • 1）Baike Embedding
      • 2）glove.6B.100
      • 3）ltp data
      • 4）Dataset
      • 5）Packages
      • 6）Model Train & Test
  • Cite
  • Acknowledge

Question-answer stance detection based on recurrent interactive attention

Abstract

Most existing answer stance detection methods ignore the interactive dependence between question and answer. This paper describes an answer stance detection method based on a recurrent interactive attention (RIA) network. This method simulates the human interactions in question answer reading comprehension using an interactive attention mechanism and iterations to simulate the interactive dependence between question and answer for detecting the answer stance. In addition, since the question text cannot explicitly express its stance, the question text is transformed into declarative sentences. Tests on a Chinese social media question answer dataset
show that this method outperforms existing answer stance detection methods due to the effective representation of the interactive dependence between the question and the answer.

Model

The architecture consists of three main modules, a question-dependent answer-attention based module (q2a), an answer-dependent question-attention based module (a2q) and a recurrent iteration module.

In contrast to the traditional unidirectional construction of question-to-answer stance representation relations, this architecture is based on question-to-answer and answer-to-question stance consistency, and uses two text-dependent modules q2a and a2q to interactively represent question and answer texts, with the aim of further introducing additional question-answer text interaction information and enhancing the deeper interaction representation of question-answer texts.

In addition, based on the strategy of repeated reading for deeper understanding, this architecture optimises the Q & A feature representation by iterating the same question and answer interaction representation several times based on a circular iteration module, which modifies the attention weight of each word in the stance representation to obtain a Q & A stance representation that combines information from multiple Q & A understandings.

Experimental Results

Usage

1）Baike Embedding

file name：baike-50.vec.txt, GoogleDrive or BaiduPan(Extraction Code: 9xr6)
download file and put it ./data/nlp_res/embeddings/baike/

2）glove.6B.100

download file and put it ./data/nlp_res/embeddings/glove/

3）ltp data

download file and put it ./data/nlp_res/ltp/ltp_data

->ltp_data  
  ->cws.model  
  ->pos.model  
  ->.etc. 

4）Dataset

The Chinese social question and answer stance dataset is in /data/processed/, '3k' is the test set and '10k' is the training set.

Before training the model, you need to perform vectorization:
Run the cd command to the pre_processing/ folder, and execute the following program:

python vectorize.py

5）Packages

* python 3.6.9
* torch 1.3.1
* numpy 1.16。0
* pyltp 0.2.1

6）Model Train & Test

Run the cd command to the nnet/ folder and execute the following command to execute the program:
Among them --nhops==3 means repeat reading 3 times.

python main_batch.py --nhops=3
python main_batch.py --is_test --nhops=3

Cite

骆旺达, 刘宇瀚, 梁斌, 徐睿峰. 基于循环交互注意力网络的问答立场分析. 清华大学学报(自然科学版), 2021, 61(9): 913-919.
Yuan, J., Zhao, Y., Xu, J., & Qin, B. (2019). Exploring Answer Stance Detection with Recurrent Conditional Attention. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 7426-7433.

Acknowledge

The data set and part of the code in this work are modified from the AAAI19 paper Exploring Answer Stance Detection with Recurrent Conditional Attention. (pdf)