Skip to content

drzqb/ABSA

Repository files navigation

细粒度情感分析 Aspect-Based Sentiment Analysis

pipline

joint

absa_joint_parallel_bertcrf:两个标注任务并行
    优点:一个模型
    缺点:两个标注任务难易程度不一,合并loss权重不易调节
    
absa_joint_series_bertcrf:两个标注任务串行
    优点:两个标注任务分别训练,互相不会干扰
    缺点:两个模型,较大;解决办法可以采用bert前11层不训练且共享

unified

absa_bertbigrucrf
absa_bertcrf
absa_bertgru
absa_bertlinear
论文:
    Exploiting BERT for End-to-End Aspect-based Sentiment Analysis
    A Unified Model for Opinion Target Extraction and Target Sentiment Predictio

mrc: 借鉴阅读理解任务的方法,此方法来源于嵌套命名实体识别最新成果

“find the term" + sentence --> 标注任务为寻找实体项目
"tag the sentiment" + sentence --> 标注任务为标记情感

About

细粒度情感分析(ABSA)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages