Towards Interactive Construction of Topical Hierarchy: A Recursive Tensor Decomposition Approach

Area of Science:

• Natural Language Processing
• Machine Learning
• Information Retrieval

Background:

• Automatic topical hierarchy construction is crucial for organizing large text volumes.
• Existing methods struggle with consistent topic preservation and slow inference during interactive modifications.
• User-driven exploration of text data requires flexible and responsive hierarchical structures.

Purpose of the Study:

• To develop a novel method for efficient and consistent interactive construction of user-desired topical hierarchies.
• To address the limitations of existing hierarchical topic modeling techniques in terms of topic preservation and inference speed.
• To enable users to freely modify topical hierarchies through operations like expanding and merging branches.

Main Methods:

• Proposes STROD (Scalable Tensor decomposition for Recursive topic Organization and Discovery), a novel method.
• Utilizes a recursive generative model for topic hierarchy representation.
• Employs a scalable tensor decomposition inference algorithm with theoretical performance guarantees.

Main Results:

• STROD allows for efficient and consistent modification of topic hierarchies.
• Empirical evaluations demonstrate a reduction in construction runtime by several orders of magnitude.
• The method generates consistent and high-quality topical hierarchies.

Conclusions:

• STROD offers a significant advancement in automatic topical hierarchy construction.
• The method provides a scalable and efficient solution for interactive user-driven topic organization.
• Future work may involve further optimization and application to diverse text corpora.