Fairness modeling for topics with different scales in short texts

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Mining

Background:

• Topic modeling short texts faces challenges like data sparsity and lack of context.
• Existing methods often overlook emerging or small-scale topics, leading to biased results.
• Information is unequally treated based on attention levels in topic modeling.

Purpose of the Study:

• To propose a fairness-oriented topic discovery approach (MixTM-G) for short texts.
• To enable the discovery of topics with varying attention levels.
• To address limitations of traditional topic modeling in identifying minor subjects.

Main Methods:

• Integrated normalized pointwise mutual information (NPMI) within a graph model.
• Utilized graph algorithms to identify semantically related clusters in document graphs.
• Employed a mixed topic modeling (MixTM) approach using bi-grams and tri-grams.

Main Results:

• The proposed MixTM-G approach demonstrates superior performance in topic modeling.
• Effectively addresses data sparsity by leveraging semantic relationships between words.
• Outperforms conventional methods in detecting small-scale topics under equivalent conditions.

Conclusions:

• MixTM-G enhances topic discovery by improving semantic associations in sparse data.
• The fairness-oriented approach ensures more equitable identification of topics.
• The method shows significant efficacy for short text topic modeling.