From Key Points to Key Point Hierarchy: Structured and Expressive Opinion Summarization

KPA extracts the main points in the data as a list of concise sentences or phrases, termed key points, and quantifies their prevalence. While key points are more expressive than word clouds and key phrases, making sense of a long, flat list of key points, which often express related ideas in varying levels of granularity, may still be challenging. To address this limitation of KPA, we introduce the task of organizing a given set of key points into a hierarchy, according to their specificity. Such hierarchies may be viewed as a novel type of Textual Entailment Graph.

 “Key Point Analysis (KPA) is a recent opinion summarization framework that aims to address the above limitations (Bar-Haim et al., 2020b). KPA extracts concise sentences and phrases termed Key Points (KPs), which represent the most salient points in the data, and quantifies the prevalence of each KP as the number of its matching input sentences. One remaining shortcoming of KPA, however, is that it generates a flat list, which does not capture the relations between the key points. For example, consider the sample set of key points in Figure 1 (left), which was automatically extracted from reviews of one of the hotels in the Yelp Open Dataset1. The results do not provide a high level view of the main themes expressed in the reviews. It is hard to tell which key points convey similar ideas, and which key points support and elaborate on a more general key point. As the number of key points in the summary increases, such output becomes even harder to consume.

In this work we introduce Key Point Hierarchies (KPH) as a novel structured representation of opinion summaries. Organizing the key points in a hierarchy, as shown in Figure 1 (right), allows the user to quickly grasp the high-level themes in the summary (the hotel is beautiful, the shows are great, comfortable rooms, great service), and drill down on each theme to get more fine-grained insights, e.g., from “The personnel were great” to “check-in was quick and easy”. Furthermore, key points that (nearly) convey the same meaning (e.g., “Housekeeping was fantastic”, and “The cleaning crew is great”) are clustered together and represented as a single node in the hierarchy. This structured output makes KPA results more consumable, informative, and easier to navigate. KPH can be viewed as a new type of textual entailment graph (§2)”