Foody-bert results from the second round of fine-tuning on the text classification task.
Continuation of fine-tuning of senty-bert, which is fine-tuned on yelp reviews and Stanford sentiment treebank with ternary labels (neutral, positive, negative).
The primary intended use is in sentiment analysis of the texts of product and service reviews, and this is the domain in which the model has been evaluated to date.
We urge caution about using these models for sentiment prediction in other domains. For example, sentiment expression in medical contexts and professional evaluations can be different from sentiment expression in product/service reviews.
Bias, Risks, and Limitations
Significant research has explored bias and fairness issues with language models (see, e.g., Sheng et al. (2021) and Bender et al. (2021)). Predictions generated by the model may include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups.
Recommendations
We recommend careful study of how these models behave, even when they are used in the domain on which they were trained and assessed. The models are deep learning models about which it is challenging to gain full analytic command; two examples that appear synonymous to human readers can receive very different predictions from these models, in ways that are hard to anticipate or explain, and so it is crucial to do continual qualitative and quantitative evaluation as part of any deployment.
We advise even more caution when using these models in new domains, as sentiment expression can shift in subtle (and not-so-subtle) ways across different domains, and this could lead specific phenomena to be mis-handled in ways that could have dramatic and pernicious consequences.
Training Details
Training Data
The model was trained on product/service reviews from Yelp, reviews from Amazon, reviews from IMDB (as defined by this dataset), sentences from Rotten Tomatoes (as given by the Stanford Sentiment Treebank), the Customer Reviews dataset, and on subsets of the DynaSent dataset. The dataset mainly contains restaurant review data.
For extensive details on these datasets are included in the associated Paper.
@article{potts-etal-2020-dynasent,
title={{DynaSent}: A Dynamic Benchmark for Sentiment Analysis},
author={Potts, Christopher and Wu, Zhengxuan and Geiger, Atticus and Kiela, Douwe},
journal={arXiv preprint arXiv:2012.15349},
url={https://arxiv.org/abs/2012.15349},
year={2020}}
