On the Trade-off between Redundancy and Local Coherence in Summarization

TL;DR

This work analyzes the trade-off between redundancy and local lexical cohesion in extractive single-document summarization, focusing on long, highly redundant scientific texts. It introduces two optimization paradigms: a reward-guided reinforcement learning framework that balances informativeness, redundancy, and cohesion, and two unsupervised KvD-based models (TreeKvD and GraphKvD) that simulate human memory processes to jointly optimize relevancy, non-redundancy, and cohesion. Automatic and human evaluations show that cohesion-focused optimization improves content organization and cohesion without sacrificing informativeness in the RL setting, while KvD-based unsupervised systems achieve highly cohesive summaries across varying document redundancy but may reduce informativeness. The study provides extensive evidence that cognitive-inspired memory mechanisms shape the balance among informativeness, redundancy, and cohesion, with practical implications for long-document summarization in technical domains.

Abstract

Extractive summaries are usually presented as lists of sentences with no expected cohesion between them and with plenty of redundant information if not accounted for. In this paper, we investigate the trade-offs incurred when aiming to control for inter-sentential cohesion and redundancy in extracted summaries, and their impact on their informativeness. As case study, we focus on the summarization of long, highly redundant documents and consider two optimization scenarios, reward-guided and with no supervision. In the reward-guided scenario, we compare systems that control for redundancy and cohesion during sentence scoring. In the unsupervised scenario, we introduce two systems that aim to control all three properties -- informativeness, redundancy, and cohesion -- in a principled way. Both systems implement a psycholinguistic theory that simulates how humans keep track of relevant content units and how cohesion and non-redundancy constraints are applied in short-term memory during reading. Extensive automatic and human evaluations reveal that systems optimizing for -- among other properties -- cohesion are capable of better organizing content in summaries compared to systems that optimize only for redundancy, while maintaining comparable informativeness. We find that the proposed unsupervised systems manage to extract highly cohesive summaries across varying levels of document redundancy, although sacrificing informativeness in the process. Finally, we lay evidence as to how simulated cognitive processes impact the trade-off between the analyzed summary properties.

Figures (10)

• Figure 1: Sections of a scientific article taken from the arXiv dataset showcasing information redundancy and cohesion. Repeated content is marked by text chunks with the same color and symbol, whilst consecutive sentences present cohesive phrases underlined.
• Figure 2: Simulation of KvD reading during three cycles. Each row shows the sentence consumed (top), the propositions extracted (left), and memory trees before (1a, 2a, 3a) and after (1b, 2b, 3b) applying a memory constraint of 5 nodes. Argument $N means that proposition N is used as argument. Squared nodes are recalled propositions. Solid lines connect nodes selected to keep in memory, and dotted lines connect nodes to be pruned.
• Figure 3: Pipeline of KvD reading simulation and sentence scoring using the simulation example in Fig.2.
• Figure 4: Step-by-step construction of proposition tree from an input sentence, starting from obtaining its dependency tree in UD format (a), merging dependent nodes into head nodes (b), promoting coordinating conjunctions to head status (c), to finally build propositions from non-leaf nodes (d).
• Figure 5: Simulation example of TreeKvD (left) and GraphKvD (right). Each memory cycle shows the input sentence, extracted propositions, and the derived memory tree. Function aP refers to subroutine attachPropositions in Algorithm \ref{['alg:kvd']}. Solid line: edge in final memory tree; dotted line: pruned edge; red dotted line: edge connecting $T$ and $P$. Squared nodes: propositions recalled from long-term memory; underlined node: new root of memory tree. Relevant content common in propositions and the gold summary is coloured in blue.