[Paper Review] Discovering topics in text datasets by visualizing relevant words

Franziska Horn, Leila Arras+3 more|arXiv (Cornell University)|Jul 18, 2017

Advanced Text Analysis Techniques14 references3 citations

TL;DR

This paper proposes a method for discovering trending topics in text datasets by identifying and visualizing relevant words that distinguish document clusters. Using DBSCAN clustering and a novel relevancy score based on true positive and false positive rates, the approach generates informative word clouds that reveal salient topics—demonstrated by uncovering key themes like Trump's inauguration and related events in New York Times snippets.

ABSTRACT

When dealing with large collections of documents, it is imperative to quickly get an overview of the texts' contents. In this paper we show how this can be achieved by using a clustering algorithm to identify topics in the dataset and then selecting and visualizing relevant words, which distinguish a group of documents from the rest of the texts, to summarize the contents of the documents belonging to each topic. We demonstrate our approach by discovering trending topics in a collection of New York Times article snippets.

Motivation & Objective

• To enable rapid exploratory analysis of large, unstructured text datasets by identifying salient topics.
• To address the limitation of traditional keyword extraction methods that fail to capture topic-specific distinctions across document groups.
• To develop a robust, fast method for identifying words that distinguish one cluster of documents from others, without requiring labeled data.
• To demonstrate the utility of word cloud visualization for summarizing and interpreting topic clusters in real-world text collections.
• To provide an open-source tool for replicating and applying the method to new text datasets.

Proposed method

• Preprocess text by lowercasing, removing non-alphanumeric characters, and computing tf-idf vectors for individual words and bigrams.
• Use DBSCAN clustering with a minimum cosine similarity threshold of 0.55 and a minimum cluster size of three documents to group related articles.
• Compute a relevancy score for each word based on its true positive rate (TPR) within a cluster and its false positive rate (FPR) across other clusters.
• Combine TPR and FPR scores into a final relevancy score using the formula: $ r_c(t_i) = 0.5(r_{c\_diff}(t_i) + r_{c\_quot}(t_i)) $, where $ r_{c\_quot} $ is the rate quotient to normalize for imbalance.
• Select top-ranking words per cluster and visualize them in word clouds to summarize each topic.
• Highlight relevant words in individual documents to support interpretation and navigation.

Experimental results

Research questions

• RQ1 Can a clustering-based approach with word visualization effectively reveal trending topics in large text collections?
• RQ2 How can relevant words be identified that distinguish one document cluster from others, especially in the absence of labeled data?
• RQ3 To what extent does the proposed TPR-FPR relevancy score outperform traditional methods like tf-idf in identifying topic-defining terms?
• RQ4 Can the method detect meaningful, real-world events (e.g., political inaugurations, sports events) from news snippets without prior topic knowledge?
• RQ5 How robust is the method to variations in cluster size and document distribution?

Key findings

• The method successfully identified the week of Trump’s inauguration as a central topic, with relevant words such as 'Trump', 'inauguration', and 'president' showing strong clustering.
• The approach detected secondary events such as the Australian Open and an avalanche in Italy, demonstrating its ability to uncover nuanced topics beyond the main event.
• Word clouds generated from top relevancy-scored words clearly summarized each cluster, enabling quick visual interpretation of document content.
• The use of TPR and FPR-based scoring reduced the risk of selecting meaningless words (e.g., 'Tuesday') that might otherwise be falsely elevated due to temporal frequency.
• Over 50 meaningful clusters were identified in the week of the inauguration, including topics related to political appointments and corrections, validating the method’s sensitivity to diverse themes.
• The open-source library 'textcatvis' enables full reproducibility and reuse of the method for other text datasets.