-
title
-
Summary of contents
This article mainly applies reinforcement learning methods to automatically learn the sentence structure related to the task, so as to achieve the purpose of auxiliary text classification. The common sense used has two points: (1) The information is redundant, so delete some information to eliminate interference. (2) The specific phrase structure has an auxiliary effect on classification.
The author defines three types of networks as shown in the figure below:
they are respectively a strategy network , a presentation network and a classification network , and their functions are as follows:
Regarding the policy network and the presentation network, you can refer to the following diagram:
- The formula is not posted, just look at the author's conclusion: the
text representation is still good.
The results have improved, and I feel that the improvement is not very significant on some tasks.