Incremental Extractive Opinion Summarization Using Cover Trees

TL;DR

This work tackles incremental extractive opinion summarization for streaming reviews by extending centroid-based methods (CentroidRank) with CoverSumm, an algorithm that uses cover-tree indexing and a small reservoir to update summaries efficiently as new reviews arrive. The approach maintains the centroid $oldsymbol{\mu}_t = \frac{1}{t}\sum_{i=1}^t x_i$ and retrieves the $k$ nearest neighbors from a reservoir to form the current summary, while guaranteeing exact nearest neighbors through principled reservoir searches with radius $r = d_k + \lambda$. Theoretical results show the reservoir search yields exact NN results, with bounds like $n_{rs} = O(D \log n)$ queries and a maximum reservoir size $|\mathcal{R}| = O(k)$, and the interval between reservoir searches grows as $O(t)$ in stable phases and $O(\sqrt{t \log t})$ under drift. Empirically, CoverSumm delivers up to 36x speedups over baselines on real and synthetic data while maintaining informative, non-redundant summaries, and human evaluations confirm alignment with the underlying review content. These findings suggest CoverSumm is a practical, scalable solution for up-to-date opinion summaries in dynamic review streams.

Abstract

Extractive opinion summarization involves automatically producing a summary of text about an entity (e.g., a product's reviews) by extracting representative sentences that capture prevalent opinions in the review set. Typically, in online marketplaces user reviews accumulate over time, and opinion summaries need to be updated periodically to provide customers with up-to-date information. In this work, we study the task of extractive opinion summarization in an incremental setting, where the underlying review set evolves over time. Many of the state-of-the-art extractive opinion summarization approaches are centrality-based, such as CentroidRank (Radev et al., 2004; Chowdhury et al., 2022). CentroidRank performs extractive summarization by selecting a subset of review sentences closest to the centroid in the representation space as the summary. However, these methods are not capable of operating efficiently in an incremental setting, where reviews arrive one at a time. In this paper, we present an efficient algorithm for accurately computing the CentroidRank summaries in an incremental setting. Our approach, CoverSumm, relies on indexing review representations in a cover tree and maintaining a reservoir of candidate summary review sentences. CoverSumm's efficacy is supported by a theoretical and empirical analysis of running time. Empirically, on a diverse collection of data (both real and synthetically created to illustrate scaling considerations), we demonstrate that CoverSumm is up to 36x faster than baseline methods, and capable of adapting to nuanced changes in data distribution. We also conduct human evaluations of the generated summaries and find that CoverSumm is capable of producing informative summaries consistent with the underlying review set.

Figures (11)

• Figure 1: Overview of the incremental CentroidRank-based summarization task and the utility of nearest neighbour (NN) data structures. (Top): We show the CentroidRank task in the incremental setting, where the centroid ($\mu_i$) evolves over time. Additionally, we illustrate the benefits of maintaining a reservoir of candidate samples for NN retrieval. (Bottom): We show how utilizing efficient NN retrieval data structures, such as the SG Tree, can enhance the efficiency of incremental summarization using reservoir search (discussed in Section \ref{['sec:rs']}), particularly in scenarios where the reservoir does not have all the NNs the centroid.
• Figure 2: An illustration of CoverSumm's operation over three consecutive time steps. Red circles represent centroids at different times, green circles indicate current nearest neighbors in reservoir $\mathcal{R}$, and blue circles denote representations outside $\mathcal{R}$. The figure displays the last query $\mu_{\mathrm{last}}$ with radius $(d_{\mathrm{k}}+\lambda)$. In the first two cases, the current centroid is within distance $\lambda$ of $\mu_{\mathrm{last}}$, and the summary can be retrieved from the reservoir. The rightmost figure shows a boundary case where a representation is just outside the reservoir's boundary and summary computation requires querying the cover tree.
• Figure 3: Time required by CoverSumm compared to baseline algorithms with an increasing number of reviews. We observe that the processing time of brute-force CentroidRank and Naive CT gradually increases, while the processing time of CoverSumm only slightly increases during summarization.
• Figure 4: Average ROUGE scores obtained by different incremental summarization systems on SPACE dataset. R1, R2, RL denote the average ROUGE-1, ROUGE-2, and ROUGE-L scores respectively. We also report the time taken for incremental summarization per entity by different algorithms.
• Figure 5: Evolution of user ratings in CoverSumm's summary and user reviews during summarization in an incremental setting. The goal of this experiment is to determine if the user ratings can be accurately reflected in the incremental summaries from CoverSumm. We report the results in three settings when reviews arrive in their: (a) original temporal order; (b) ascending order of their ratings; (c) descending order of their user ratings. We observe that CoverSumm's summary can track drifts in the ratings.

Theorems & Definitions (18)

• Proposition 1: Bounding distance to centroid
• Proposition 2: Bounding distance between subsequent centroids
• Proposition 3: Correctness of Reservoir Search
• proof
• Proposition 4: Exact nearest neighbours
• proof
• Proposition 5: Number of reservoir search queries
• Proposition 6: Maximum reservoir size
• Theorem 1: Hoeffding-Azuma inequality hoeffding1963probability
• proof