Neurosymbolic AI for Personalized Sentiment Analysis

Overview

🔥 This is a repository for our paper (Neurosymbolic AI for Personalized Sentiment Analysis) accepted in HCII'24 (26th International Conference on Human-Computer Interaction). (Author list: Luyao Zhu, Rui Mao, Erik Cambria, and Bernard J. Jansen.)

In this work, we propose a framework for personalized sentiment analysis that performs automatic user profiling by modeling users based on different levels of personalization, before performing sentiment analysis. In particular, such framework leverages seven levels of personalization (from bottom to top), namely: Entity, to distinguish between humans and other intelligent agents; Culture, to take into account how different cultures perceive the same concept as positive or negative; Religion, to consider how specific religious beliefs may affect an individual’s opinion about certain topics; Vocation, to better gauge people’s opinion based on their job and education level; Ideology, to take into account political beliefs as well as social, economic, or philosophical viewpoints; Personality, to better classify certain concepts as positive or negative based on personality traits; finally, Subjectivity, to take into account personal preferences and experiences.

Dataset

The original dataset, Harry Potter Dialogue (HPD), can be downloaded from the link.

Usage

Query different aspects

There are 7 different aspects, i.e., subjectivity, personality, ideology, vocation, religion, culture, entity. Change the flag aspect to get the user profile with respect to 7 different levels of personalization.

python run_query.py --aspect "entity" --proposer_model "gpt-4-turbo-2024-04-09"

Sentiment analysis based on different levels of personalization

There are 9 values for flag aspect, including 7 aforementioned aspects, all, and 0. all means utilizing all the 7 aspects for sentiment analysis. 0 means conduct sentiment analysis without any user profiles.

python run_query_sa.py --aspect "all" --proposer_model "gpt-4-turbo-2024-04-09"

Evaluation

1. eval_by_category_whole.py shows the personalization utility analysis by entity and culture aspects. The results are adjusted by answer rate.

python eval_by_category_whole.py

2. eval_by_category.py shows the personalization utility analysis by entity and culture aspects. The results are not adjusted by answer rate, which means the metrics are calculated for only the queries that are successfully answered.

python eval_by_category.py

Issues and Usage Q&A

To ask questions, report issues or request features 🤔, please use the GitHub Issue Tracker. Before creating a new issue, please make sure to search for existing issues that may solve your problem.

Citation

Please cite as

@inproceedings{zhu2024neurosymbolic,
  title={Neurosymbolic AI for personalized sentiment analysis},
  author={Zhu, Luyao and Mao, Rui and Cambria, Erik and Jansen, Bernard J},
  month = {jul},
  year = {2024},
  address = {Washington DC, USA},
  booktitle={International Conference on Human-Computer Interaction},
  organization={Springer}
}