Work

User-guided dynamic topic discovery in large texts

Year: 2023
Type: article
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.I would like to express my deepest gratitude to my adviser, Professor Jiawei Han, for his unwavering kindness, excellent guidance, sound judgment, and constant encouragement throughout this journey.His systematic approach has significantly changed how I think about research.I am also immensely grateful to my collaborators, Shivam Agarwal, Nishant Balepur, and Susik Yoon, for their dedication and enthusiasm.Brainstorming ideas and working together with them has been a truly enjoyable experience, reinforcing my passion for teamwork.
Language: English
Open Access status: green
OpenAlex ID: W7220779482
FWCI:
Cites: 87
Cited by:
Related to: 10