search

teapot123 / JASen

Code and Data for our EMNLP-2020 paper Weakly-Supervised Aspect-Based Sentiment Analysis via Joint Aspect-Sentiment Topic Embedding.
49 stars 5 forks source link

Only one aspect can be predicted? #4

Closed TeTeTang closed 2 years ago

TeTeTang commented 2 years ago

Hi, I noticed the sample test set only has one aspect/sentiment labeled for each review/doc, but the paper says your model aims to output a set of joint topics for each review, how to adjust the test set to achieve that goal? Thanks.

teapot123 commented 2 years ago

Hi, thanks for your interest! The setting of my paper is to predict the sentiment and aspect simultaneously for each review sentence, so it is still a single-label problem. My method may not work well on multi-label problems (if I understand you correctly), but since it outputs prediction logits for all the aspects, perhaps you can try to set a threshold to get multiple aspects for each review. However, it would still be better to use loss functions designed specifically for multi-label classification to deal with this problem.

TeTeTang commented 2 years ago

Thanks for the response, that resolved my question.