A Simple Theoretical Model of Importance for Summarization

Abstract

Research on summarization has mainly been driven by empirical approaches, crafting systems to perform well on standard datasets with the notion of information Importance remaining latent. We argue that establishing theoretical models of Importance will advance our understanding of the task and help to further improve summarization systems. To this end, we propose simple but rigorous definitions of several concepts that were previously used only intuitively in summarization: Redundancy, Relevance, and Informativeness. Importance arises as a single quantity naturally unifying these concepts. Additionally, we provide intuitions to interpret the proposed quantities and experiments to demonstrate the potential of the framework to inform and guide subsequent works.

Earlier Isn’t Always Better: Sub-aspect Analysis on Corpus and System Biases in Summarization

“…Our study extends the hypothesis to various corpora as well as systems. With a specific focus on importance aspect, a recent work (Peyrard, 2019a) divided it into three subcategories; redundancy, relevance, and informativeness, and provided quantities of each to measure. Compared to this, ours provide broader scale of sub-aspect analysis across various corpora and systems.…”

ToC-RWG: Explore the Combination of Topic Model and Citation Information for Automatic Related Work Generation

“…This criterion casts summarization as finding a set of summary sentences which closely match the doc distribution. When selecting sentences to constitute the summary, this optimization objective penalizes redundancy while maximizing relevance [17]. Because the problem of finding the subset of sentences from a collection that minimizes the KL divergence is NP-hard, a greedy algorithm is often used in practice.…”

KLearn: Background Knowledge Inference from Summarization Data

“…Despite its fundamental role, background knowledge has received little attention from the summarization community. Existing approaches largely focus on the relevance aspect, which enforces similarity between the generated summaries and the source documents (Peyrard, 2019). Figure 1: A summary (S) results from the combination of the background knowledge (K) and the source document (D).…”

“…Figure 1: A summary (S) results from the combination of the background knowledge (K) and the source document (D). Following Peyrard (2019), S is similar to D (Relevance measured by a small KL(S||D)) but also brings new information compared to background knowledge (informativeness measured by a large KL(S||K)). We can infer the unobserved K from the choices unexplained by the Relevance criteria.…”

“…Indeed, choices made by human summarizers and human annotators provide implicit information about their background knowledge. We build upon a recent theoretical model of information selection (Peyrard, 2019), which postulates that information selected in the summary results from 3 desiderata: low redundancy (the summary contain diverse information), high relevance (the summary is representative of the document), and high informativeness (the summary adds new information on top of the background knowledge). The tension between these 3 elements is encoded in a summary scoring function θ K that explicitly depends on the background knowledge K. As illustrated by Fig.…”