Skip to main content

phd project chenghao

Escape from information cocoon with language models

Recommender system of question-and-answering platform nowadays has faced with the problem of trapping users in the information cocoon, in which people keep being recommended similar contents that they have shown interests in. We identify that those contents, however, typically involving academic-, professions- or personal problems-related contents based on users’ searching history, are likely to cause anxiety, especially during the periods of a day that users are in urgent need of escaping from their social roles and mental burdens.

This project tries to propose a new framework, in which users are enabled to customize their personalization through a measurement of “fun”, without any user-related contexts taken into considerations, or after user-related contexts are taken into considerations.

This project involves four stages of research.

  1. The collection of relevant textual data and the annotation of “Fun”. We plan to experiment two approaches. a) Identify platforms that explicitly collect and produce fun contents, and develop unique metric to automatically label the data. b) First collect relevant data, manually annotate a small portion of the data, then pseudo label the rest of the data using state-of-the-art semi-supervised learning framework.
  2. Exploratory analysis of what makes “fun” using NLP analysis framework.
  3. Training of state-of-the-art contextualized language models to understand and predict fun.
  4. Evaluation and user studies of applications of the framework. This might involve the development of a user interface that enables users to interact with the models through customization of the scale of fun, through which we can a) collect users’ time on page, clicking data, etc., b) study whether users’ perception of fun aligns with language models’ predictions. c) understand users’ experience through questionnaire and qualitative interview.

Deliverables We aim to outcome two lines of papers, one concerning contributions to the NLP community based on the extensive study of Fun (targeting EMNLP, etc.), one concerning contributions to the human-AI interaction community based on the user study of the language model-facilitated customized recommender system (targeting CHI, IUI, etc.).


Baicheng Sun

Sun Yat-sen University/ Tsinghua University