This research introduces a new family of topic coherence metrics called Contextualized Topic Coherence Metrics (CTC) that benefits from the recent development of Large Language Models (LLM). CTC includes two approaches that are motivated to offer flexibility and accuracy in evaluating neural topic models under different circumstances. Our results show automated CTC outperforms the baseline metrics on large-scale datasets while semi-automated CTC outperforms the baseline metrics on short-text datasets.
CTC is implemented as a service for researchers and engineers who aim to evaluate and fine-tune their topic models. The source code of this python package is provided in ./ctc and a notebook named example.ipynb is prepared to explain how to use this python package as follows.