Personalized Text Summarization Based on Gaze Patterns

With the explosive growth of online content, summarization has become important for users to grasp information quickly. However, existing summarization only provides static and one-size-fits-all results, e.g., a paragraph is summarized to the same short sentence, for different users, failing to satisfy users' diverse preferences. In our subjective study, such preference diversity on content summarization can be quite significant, e.g., the ground-truth summaries provided by different users on the same long text only have an average cosine similarity value of 0.383. This paper fills this gap by a personalized content summarization framework. First, we have collected a Chinese long text summarization dataset with gaze behavior of different users. Based on our measurement study on the dataset, we reveal that people's gaze behavior and their preferred content summary have strong correlation, e.g., the preference of different parts of speech (POS); Second, we propose to incorporate such gaze-based"attention patterns" into text summarization, by designing a gaze-based key sentence extracting strategy, PRank (short for Perso alRank) which is then integrated with a conventional pointer generator model to satisfy different individuals. Our trace-driven experiments on the dataset verify the effectiveness of our design: our model outperforms baselines by at least 3 ROUGE-1 points.