annotator
dict | passage
dict |
|---|---|
{"0":0,"1":0,"2":0,"3":0,"4":0,"5":0,"6":0,"7":0,"8":0,"9":0,"10":0,"11":0,"12":0,"13":0,"14":0,"15"(...TRUNCATED) | {"0":"[NEU] will not yet recognize [NEG] as [NEU], the [NEU] said Monday as world leaders rushed to (...TRUNCATED) |
Task: Aspect based sentiment recognition. Domain: Political news coverage. Named entities have been masked with [NEG], [NEU], or [POS] tokens. Dataset size: 1400 clippits.
An ABSA-BERT model successfully leveraged this dataset, with semi-supervised learning, to get very good results on fine grained sentiment recognition.
Creation: Five of my friends voluntarily used my browser plugin annotation tool I made to send me marked sentences of the news they were reading. How well they understood the assignment is not always clear. There is very little quality control (no agreement tests etc), but it was successful in the end.
Tokens are chosen based on how the sentence paints that entity. Context words are the only signal available to the model for classification, so the model won't learn any embedded sentiment biases within named entities themselves.
samples:
- [NEG] is too incompetent to overturn the election or put democracy in peril
- After [POS] was elected governor in 2010, he and his allies in the state Senate and Assembly fine-tuned tactics that [POS] and national [POS] leaders apparently hope will carry them to another squeaker of a win in the state.
- [NEG] is a living, breathing, walking conflict of interest
- [POS] is the world champion of aging well—physically, intellectually, spiritually, and emotionally.
- Downloads last month
- 4