Reading the paper "learning span level interactions for aspect sentimental triple extraction"

An overview of the article

   At present, for ASTE Most task processing methods are based on the interaction between words , As a result, if the aspect entity or opinion word contains multiple token Poor performance at . Therefore, this article proposes a span level （Span-ASTE） Methods , The interaction of the whole span between goals and opinions is explicitly considered . It can be predicted by using the semantics of the whole span , So as to better ensure emotional consistency .

Article model

   This paper proposes Span-ASTE The model is shown above , It consists of three parts ： Sentence coding , Method module and triplet module .

Sentence coding

   For sentence coding, the author adopts 2 Ways of planting , One is to use Glove, The other is BERT, Of course it must be BERT Better . After obtaining the hidden layer state representation of each word , The author passes the predefined span length , Word by word, the span representation of its corresponding length , As shown below ：

Among them the Fwidth Represents the width of this span , start 2 The first is the starting and receiving position of splicing , Of course, the author also made a comparison between average pooling and span pooling .

Method module

   Here the author represents the span obtained in the first step , Input to a feedforward neural network for processing ATE Task or OTE Mission .

But there is a problem , That is, there are too many spans , If everyone has to judge like this, the time complexity is too high , Therefore, the author set up a two-way pruning strategy , That is to judge whether each extracted span is an aspect entity or opinion word

   Finally, according to the obtained aspect entity span and opinion span , Judge the corresponding emotional polarity

Invalid It represents invalid emotion , Such as invalid pairing .
   Last self feeling , It is also similar to a multi task processing .
Article address ：Learning Span-Level Interactions for Aspect Sentiment Triplet Extraction