标题:Neural Machine Reading Comprehension: Methods and Trends
-
Author: Shanshan Liu, Xin Zhang, Sheng Zhang, Hui Wang, Weiming Zhang
-
Link: https: //arxiv.org/pdf/1907.01118.pdf
Abstract: Over the past few years, with the advent of deep learning, machine reading comprehension (which requires a machine based on a given context to answer questions) have won more and more attention. While reading based on machine learning to understand the depth of research is booming, but there is no comprehensive research articles to summarize methods have been proposed in this field and recent trends. So, this recent research work in this field full of potential in a comprehensive overview.
Specifically, the researchers first compared the machine reading comprehension tasks in different dimensions, and describes the overall architecture. Then, they further classify the SOTA method commonly used model used in the field. Finally, the researchers discussed the new trends in the field, and made a number of outstanding issues at the end of the article.
Recommendation: This article National University of Defense Technology comprehensive introduction to the status of the machine reading comprehension research, development and new trends, is rare in the field review articles. Reading comprehension machine has great significance in terms of the machine Q & A, information search, etc., recommend interested readers to read this article.
Article structure:
1. Introduction MRC
2. Tasks and evaluation matrix
MRC can be divided into four tasks: |
Corresponding to the data set |
loze Test, cloze |
CNN & Daily Mail,CBT (The Children’s Book Test),LAM- BADA dataset (LAnguage Modeling Boardened to Account for Discourse Aspects),Who-did-What,CLOTH,CliCR |
Multiple Choice, Radio |
MCTest,RACE |
Span Extraction, extract span |
SQuAD,NewsQA,TriviaQA,DuoRC |
Free Answering, free answer |
bAbI,MS MARCO,SearchQA,NarrativeQA,DuReader |
评价维度:construction, understanding, flexibility, evaluation and application
评价标准:ACC,F1, ROUGE-L, ROUGE (Recall-Oriented Understudy for Gisting Evaluation),BLEU (Bilingual Evaluation Understudy)
3.MRC system general structure:
step |
For example |
Embeddings |
Word2vec |
Feature Extraction |
RNN, CNN |
Context-Question Interaction (find some of the most relevant to the issue in the text) |
Attention mechanism (attention mechanism), unidirectional or bidirectional |
Answer Prediction |
|
4. Application of deep learning system in MRC
The latest developments
Reading Comprehension knowledge-based machines
Can not answer questions
Multi-stage machine reading comprehension
Q & A session
6. Unresolved Issues
Integration of external knowledge
MRC system robustness
Limitations in a given context
Reasoning skills shortage
7. Conclusions