Reading of papers on "towards generative aspect based sentimental analysis"

The article introduces

   Currently for ASTE Most tasks are processed for different parts, that is, aspect entity recognition , Emotional word extraction , Emotional analysis designs different ways to deal with . But such processing is difficult to take into account the rich semantic information of tags , But also more complex . Therefore, this question proposes a unified generation framework to deal with different ABSA Mission , There are two types of paradigms , Annotation style （annotation-style） And extract styles （extraction-style）.

Method

   Examples of the two styles are as follows

We can know that the annotation style is to add the corresponding emotional polarity and opinion words after the aspect entity of the original sentence . The extracted style is similar to the standard ASTE Triple form of （what,why,how）.
   In fact, it's easy for us to know , Although there is no special design in the way of generation , But the generated words are difficult to control , For example, the case and plural , Sometimes even one character is wrong , But this will be regarded as a wrong example in judgment , It's a pity . So the question was made trick（ Prediction regularization ）, That is, by calculating the editing distance between the predicted words and the words in the Library （Levenshtein distance）, Convert it to edit the nearest point , As shown in the figure below .

Article address ：Towards Generative Aspect-Based Sentiment Analysis