III: Small: A Holistic Approach to Sentiment Analysis

Lifelong Learning for Sentiment Analysis

Award Information

People Involved

In addition to the PI, the following graduate students work on the project.

Project Goals

This project aims to propose novel lifelong learning (LL) techniques for 4 core sub-tasks of sentiment analysis (SA) and 1 joint model for holistic SA of all sub-tasks: (1) Learning on the job for aspect extraction, which learns while working after model building. In traditional learning, after model is built, the model is only applied without any further improvement until the next training cycle. (2) Lifelong aspect topic modeling for aspect grouping, which uses past grouping of aspects to help the new grouping. (3) Lifelong attention and self-supervised learning for aspect sentiment classification, which retains the previous attention distributions and leverages them to generate more accurate attentions in a self-supervised manner for the new task classification. (4) Continuous association learning for coreference resolution, which helps discover coreference relations more accurately. (5) Lifelong multi-heterogeneous-task learning to jointly learn all sub-tasks of SA in a holistic manner to exploit the synergy of the tasks in order to improve the overall accuracy of SA.

Broader Impacts

SA is an important research area. Anyone who wants to gain information or knowledge from social media needs it. SA has been studied in many areas of computer science, e.g., natural language processing, data mining, and information retrieval. It is also increasingly applied and studied in social sciences. Thus, making a major improvement in SA can have an important impact in these areas. Apart from research, commercial applications of SA is also flourishing. Successful completion of the project will have a positive impact on these applications. Since LL has not been researched extensively, this project will also make fundamental contributions to LL itself. This project will train two graduate students. Annotated data and implemented software from the project will be made publicly available. Mature algorithms will be included in PI’s data mining and text mining course. Many parts of this project can be blended into the course project. PI will also actively recruit minority and female students, and engage undergraduates.



  1. Zixuan Ke, Bing Liu, Hao Wang, and Lei Shu. Continual Learning with Knowledge Transfer for Sentiment Classification. to appear in Proceedings of European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD-2020), Ghent, Belgium, 14-18, September 2020.

  2. Nianzu Ma, Sahisnu Mazumder, Hao Wang, Bing Liu. Entity-Aware Dependency-based Deep Graph Attention Network for Comparative Preference Classification. to appear in Proceedings of Annual Meeting of the Association for Computational Linguistics (ACL-2020, short paper), Seattle, Washington, July 5th to July 10th, 2020.

Annotated Data

Will be added once available

Software Download

Code for (Ke et al., 2020) can be downlloaded here.

Code for (Ma et al., 2020) can be downlloaded here.


Under construction


Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Date of Last Update: August 1, 2019