Evaluating the Performance to Detecting Topic Shift of Large Language Models: A Study

Topic shift detection aims to identify whether there is a change in the current topic of conversation or if a change is needed. The study found previous work did not evaluate the performance of large language models like ChatGPT on the task of topic shift. Therefore, this paper's main task and innovation lie in analyzing ChatGPT's performance on topic shift detection. To provide a more comprehensive evaluation, we conducted topic shift detection tasks on ChatGPT from three aspects: single utterances, adjacent utterances, and contextual levels. Additionally, to gauge the performance of large language models, we conducted experiments on multiple small-scale models and compared the results of the two models. Experimental results on the publicly available English TIAGE dataset showed that small-scale models exhibited lower recall in all three aspects, while ChatGPT performed better in the recall. This suggests that compared to small-scale models, large models are more capable of accurately detecting topic shifts. However, large models also exhibited lower precision, indicating that while ChatGPT can recognize content differences in utterances, its judgment on whether these different contents belong to the same topic is poor.