search

salesforce / fewshot_absa

A Generative Language Model for Few-shot Aspect-Based Sentiment Analysis
https://arxiv.org/abs/2204.05356
BSD 3-Clause "New" or "Revised" License
28 stars 10 forks source link

what is your contribution？ #7

Open Heart-beatsss opened 2 years ago

Heart-beatsss commented 2 years ago

sorry, i don't know your contribution just reformulating the tasks? and proving GPT-2 is powerful???

ehosseiniasl commented 2 years ago

Hi Thanks for your comment. all of previous baselines used Bert based model for a single task, aspect term extraction, or polarity prediction. we are the first one to reformulate this task as language generation, and achieved very good few shot performance results, along with multi tasking. There are extensive research on how to get the best few-shot performance from unidirectional and bidirectional models. I hope this answers your question

Heart-beatsss commented 2 years ago

Hi Thanks for your comment. all of previous baselines used Bert based model for a single task, aspect term extraction, or polarity prediction. we are the first one to reformulate this task as language generation, and achieved very good few shot performance results, along with multi tasking. There are extensive research on how to get the best few-shot performance from unidirectional and bidirectional models. I hope this answers your question

Thanks for your reply. but so sorry, i disagree with you.

  1. "all of previous baselines used Bert based model for a single task, aspect term extraction, or polarity prediction.we are the first one to reformulate this task as language generation,"you said. As far as me know, in ACL2021, earlier than 2022, there are two papers reformulate this task as language generation(A Unified Generative Framework for Aspect-Based Sentiment Analysis; Towards Generative Aspect-Based Sentiment Analysis). And Your Tasks are defined in a similar way to these works. The formulation is too common, it's not innovative.
  2. "few shot performance", I just have seen that GPT-2 is powerful, you haven't done a lot of work on "few-shot" / model / tricks. Admittedly few-shot works great, but I don't think there's much to learn from this paper. Congratulations on being accepted by COLING2022, I'm just giving my opinions, hope to be corrected if there are mistakes, and hope to see your better works and more papers.