ProxAnn: Use-Oriented Evaluations of Topic Models and Document Clustering
This work addresses the challenge of scalable and human-aligned evaluations for topic models and document clustering, which is incremental by providing a new protocol and proxy method.
The authors tackled the problem of evaluating topic models and document clustering by designing a scalable human evaluation protocol and an automated LLM-based proxy that aligns with practitioner usage, finding that the best LLM proxies are statistically indistinguishable from human annotators.
Topic model and document-clustering evaluations either use automated metrics that align poorly with human preferences or require expert labels that are intractable to scale. We design a scalable human evaluation protocol and a corresponding automated approximation that reflect practitioners' real-world usage of models. Annotators -- or an LLM-based proxy -- review text items assigned to a topic or cluster, infer a category for the group, then apply that category to other documents. Using this protocol, we collect extensive crowdworker annotations of outputs from a diverse set of topic models on two datasets. We then use these annotations to validate automated proxies, finding that the best LLM proxies are statistically indistinguishable from a human annotator and can therefore serve as a reasonable substitute in automated evaluations. Package, web interface, and data are at https://github.com/ahoho/proxann