Umeå University's logo

Topic models are powerful tools to get an overview of large collections of text data, a situation that is prevalent in industry applications. A rising trend within topic modeling is to directly cluster dimension-reduced embeddings created with pretrained language models. It is difficult to evaluate these models because there is no ground truth and automatic measurements may not mimic human judgment. To address this problem, we created a tool called STELLAR for interactive topic browsing which we used for human evaluation of topics created from a real-world dataset used in industry. Embeddings created with BERT were used together with UMAP and HDBSCAN to model the topics. The human evaluation found that our topic model creates coherent topics. The following discussion revolves around the requirements of industry and what research is needed for production-ready systems.

Document clustering models serve unique application purposes, which turns model quality into a property that depends on the needs of the individual investigator. We propose a framework, Cluster Interpretation and Precision from Human Exploration (CIPHE), for collecting and quantifying human interpretations of cluster samples. CIPHE tasks survey participants to explore actual document texts from cluster samples and records their perceptions. It also includes a novel inclusion task that is used to calculate the cluster precision in an indirect manner. A case study on news clusters shows that CIPHE reveals which clusters have multiple interpretation angles, aiding the investigator in their exploration.

A frequent problem in document clustering and topic modeling is the lack of ground truth. Models are typically intended to reflect some aspect of how human readers view texts (the general theme, sentiment, emotional response, etc), but it can be difficult to assess whether they actually do. The only real ground truth is human judgement. To enable researchers and practitioners to collect such judgement in a cost-efficient standardized way, we have developed the crowdsourcing solution CIPHE -- Cluster Interpretation and Precision from Human Exploration. CIPHE is an adaptable framework which systematically gathers and evaluates data on the human perception of a set of document clusters where participants read sample texts from the cluster. In this article, we use CIPHE to study the limitations that keyword-based methods pose in topic modeling coherence evaluation. Keyword methods, including word intrusion, are compared with the outcome of the thorougher CIPHE on scoring and characterizing clusters. The results show how the abstraction of keywords skews the cluster interpretation for almost half of the compared instances, meaning that many important cluster characteristics are missed. Further, we present a case study where CIPHE is used to (a) provide insights into the UK news domain and (b) find out how the evaluated clustering model should be tuned to better suit the intended application. The experiments provide evidence that CIPHE characterizes clusters in a predictable manner and has the potential to be a valuable framework for using human evaluation in the pursuit of nuanced research aims.

Document clustering is frequently used in applications of natural language processing, e.g. to classify news articles or create topic models. In this paper, we study document clustering with the common clustering pipeline that includes vectorization with BERT or Doc2Vec, dimension reduction with PCA or UMAP, and clustering with K-Means or HDBSCAN. We discuss the inter- actions of the different components in the pipeline, parameter settings, and how to determine an appropriate number of dimensions. The results suggest that BERT embeddings combined with UMAP dimension reduction to no less than 15 dimensions provides a good basis for clustering, regardless of the specific clustering algorithm used. Moreover, while UMAP performed better than PCA in our experiments, tuning the UMAP settings showed little impact on the overall performance. Hence, we recommend configuring UMAP so as to optimize its time efficiency. According to our topic model evaluation, the combination of BERT and UMAP, also used in BERTopic, performs best. A topic model based on this pipeline typically benefits from a large number of clusters.

Validating the accuracy of a classifier in its application environment is a labor-intensive and often subjective task. Adapting large language models (LLMs) to classification tasks makes reliable human evaluation increasingly important due to their generative capabilities. Acceptance Sampling is a method long used in traditional production lines to validate products before shipment. We adapt Acceptance Sampling concepts to the validation of text classifiers and conduct a case study focusing on news. We show how Acceptance Sampling can efficiently optimize the required amount of manual labor and discuss how the subjective elements of classifying texts can be handled by defining acceptance criteria.