Bio annotation of emotion analysis aste triples extraction

BIO brief introduction

   In natural language processing tasks , Sequence tagging is a common treatment scheme , For example, single element represented by information extraction / Tuple extraction task . Common sequences are labeled BIO, among B representative Begin, Is the beginning of the target phrase ;I representative Inside In the middle of the phrase for the target ,O Then it means that it is not the target phrase . Examples are as follows [1]：

   Than BIO Further is BIOES, They represent respectively begin, inside, outside, end and single, That is to say, bi BIO More specific ,single Indicates that the target phrase contains only one word , When the target phrase has only two words , It can be used directly begin and end To mark .

A unified BIO mark

   stay BIOES Based on the task , In order to better handle ASTE Mission , In the literature [2] Based on the literature [3] Establish a unified BIO Labeling scheme , Combine emotion with aspect term Combination ,opinion It's the same as before , As shown in the figure below ：

Location aware BIO Labeling scheme

   But the above scheme ignores the treatment NLP An important factor of the task is the distance and connection between words . So the literature [4] Based on the above annotation scheme, a location aware annotation scheme is proposed , the aspect term and opinion The positional relationship between them is also taken into account , As shown in the figure below ：

   So that we can put aspect term The polarity of and the position of the opinion word associated with it are taken into account .
   Of course, there is a problem with this scheme aspect term（ Aspect terms ） Yes, it can correspond to multiple opinion Of , So the literature [4] The author of has proposed another variant , That is to say, the opinion words aspect term The standard group method of is interchangeable , Thus the above problems are solved .

reference

[1]A More Fine-Grained Aspect-Sentiment-Opinion Triplet Extraction Task
[2]A Unified Model for Opinion Target Extraction and Target Sentiment Prediction
[3]Knowing What, How and Why: A Near Complete Solution for Aspect-Based Sentiment Analysis
[4]Position-Aware Tagging for Aspect Sentiment Triplet Extraction