User-guided dynamic topic discovery in large texts

Venkat Ramanan, Karthik

Permalink

https://hdl.handle.net/2142/120590 Copy

Description

Title

User-guided dynamic topic discovery in large texts

Author(s)

Venkat Ramanan, Karthik

Issue Date

2023-05-04

Director of Research (if dissertation) or Advisor (if thesis)

Han, Jiawei

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Information Retrieval
• Question Answering
• Natural Language Processing
• Data Mining
• Pretrained Language Models
• Topic Modeling

Abstract

Dynamic topic models (DTMs) play a crucial role in generating insights from large timestamped corpora of text by capturing the evolution of topics over time. Despite their popularity, existing DTMs are fully unsupervised, resulting in generated topic evolutions that often do not cater to a user’s needs. Additionally, the topic evolutions produced by DTMs tend to contain generic terms that do not accurately represent their designated time steps. This is particularly problematic as DTMs are frequently employed for analyzing the evolution of specific topics within a corpus. To address these challenges, we propose ReGenT, a framework for Dynamic, Discriminative Topic Discovery. This task aims to discover topic evolutions from temporal corpora that align with a set of user-provided category names while uniquely capturing topics at each time step. We accomplish this by (1) utilizing a retrieval-QA framework to retrieve relevant words for seeds with high granularity, (2) automatically generating and ranking strong questions to probe future words to expand our initial word set, (3) ensuring that the mined words are distinctly popular at a given time, and (4) iteratively refining our word list through ensemble ranking. We conduct experiments on two diverse datasets and demonstrate that ReGenT achieves state-of-the-art performance through extensive evaluations.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120590

Copyright and License Information

Copyright 2023 Karthik Venkat Ramanan

Owning Collections