In this paper we propose a new framework and new methods for the reference-free evaluation of topic segmentation systems directly in the embedding space. Specifically, we define a common framework for reference-free, embedding-based topic segmentation metrics, and show how this applies to an existing metric. We then define new metrics, based on a previously defined cohesion score, Average Relative Proximity. Using this approach, we show that Large Language Models (LLMs) yield features that, if used correctly, can strongly correlate with traditional topic segmentation metrics based on costly and rare human annotations, while outperforming existing reference-free metrics borrowed from clustering evaluation in most domains. We then show that smaller language models specifically fine-tuned for different sentence-level tasks can outperform LLMs several orders of magnitude larger. Via a thorough comparison of our metric’s performance across different datasets, we see that conversational data present the biggest challenge in this framework. Finally, we analyse the behaviour of our metrics in specific error cases, such as those of under-generation and moving of ground truth topic boundaries, and show that our metrics behave more consistently than other reference-free methods.

Recent works on linear text segmentation have shown new state-of-the-art results nearly every year. Most times, however, these recent advances include a variety of different elements which makes it difficult to evaluate which individual components of the proposed methods bring about improvements for the task and, more generally, what actually works for linear text segmentation. Moreover, evaluating text segmentation is notoriously difficult and the use of a metric such as Pk, which is widely used in existing literature, presents specific problems that complicates a fair comparison between segmentation models. In this work, then, we draw from a number of existing works to assess which is the state-of-the-art in linear text segmentation, investigating what architectures and features work best for the task. For doing so, we present three models representative of a variety of approaches, we compare them to existing methods and we inspect elements composing them, so as to give a more complete picture of which technique is more successful and why that might be the case. At the same time, we highlight a specific feature of Pk which can bias the results and we report our results using different settings, so as to give future literature a more comprehensive set of baseline results for future developments. We then hope that this work can serve as a solid foundation to foster research in the area, overcoming task-specific difficulties such as evaluation setting and providing new state-of-the-art results.