Multi-source Opinion Summarization

• Multi-source Opinion Summarization is a technique that aggregates diverse opinion-bearing texts and metadata to produce summaries reflecting consensus rather than individual views.
• It employs various representations—extractive, abstractive, and rationale-based—to handle contradictions and emphasize key aspects and sentiments in large review sets.
• Evaluation strategies focus on source-grounded, aspect-aware measures and human judgments to tackle faithfulness, representativeness, and minority opinion challenges.

Searching arXiv for recent and foundational papers on multi-source opinion summarization to support the article. Multi-source Opinion Summarization (M-OS) is the task of generating a summary from multiple opinion-bearing sources about the same entity or topic, such as user reviews for a product, hotel, restaurant, or movie, so that the output reflects salient, representative, or consensus opinions rather than the content of any single source (Suhara et al., 2020, Elsahar et al., 2020, Bražinskas et al., 2019). In the literature, the canonical setting is multi-review summarization, but the scope has widened to include additional product metadata, comparative summaries over multiple entities, large-scale multi-perspective summaries, microblogs, and scientific peer review (Attri et al., 7 Jul 2025, Iso et al., 2021, Jiang et al., 2023, Bilal et al., 2022, 2305.14647). Across these settings, the field is organized by a small set of recurring technical questions: how to represent dispersed and contradictory opinions, how to train without large gold-summary corpora, how to control aspect or sentiment focus, and how to evaluate summaries when reference summaries may themselves be incomplete or unrepresentative (Bražinskas et al., 2020, Malon, 2023, Shen et al., 2023).

1. Scope, definitions, and problem variants

A standard formulation treats an entity $e$ as associated with a set of reviews $R_e=\{r_i\}_{i=1}^{|R_e|}$, where each review is free-form text and the desired output is a textual summary $s_e$ that captures the most salient opinions expressed across the reviews (Suhara et al., 2020). In grouped-review generative formulations, the same setting appears as a review collection $r_{1:N}=(r_1,\dots,r_N)$ tied by shared entity-level semantics, with summary generation cast as producing a “consensus review” or “copycat review” from the group (Bražinskas et al., 2019). Self-supervised formulations make the same multi-source assumption explicit by treating an individual review $s_i$ as a pseudo-summary for a set of similar reviews $D_i$ about the same entity (Elsahar et al., 2020).

The literature distinguishes M-OS from generic multi-document summarization by the nature of the source material. The inputs are subjective, noisy, redundant, and often contradictory, and good summaries must aggregate opinions rather than merely compress topic facts (Elsahar et al., 2020, Shen et al., 2023). In the e-commerce literature, a stronger formulation extends M-OS beyond review-only summarization by integrating customer reviews with product title, description, key features, specifications, and average rating, so that summaries capture both subjective evidence and objective product attributes (Attri et al., 7 Jul 2025). This broader definition is not universal across all papers, but it has become an important recent direction.

Task variants differ mainly in the target output. Some systems produce a single generic summary for one entity (Amplayo et al., 2019, Bražinskas et al., 2020). Others generate perspective-specific outputs, such as pros, cons, and verdict summaries from the same review pool (Jiang et al., 2023). Comparative formulations generate two contrastive summaries and one common summary from two review sets, one for each entity-specific difference and one for shared content (Iso et al., 2021). Domain-specific extensions include microblog opinion summarisation, where summaries of tweets separate factual information from author opinions (Bilal et al., 2022), and scientific opinion summarization, where reviewer reports are synthesized into meta-reviews that summarize strengths, weaknesses, consensus, controversy, and decision support (2305.14647).

2. Core formulations and summary representations

The field has used several summary representations, each corresponding to a different view of what constitutes an opinion summary. Early weakly supervised extractive work represents an opinion as $o_s=(s,A_s,\mathit{pol}_s)$, where $s$ is a review segment, $A_s$ is an aspect set, and $\mathit{pol}_s\in[-1,+1]$ is segment polarity, and defines the summary for product $R_e=\{r_i\}_{i=1}^{|R_e|}$0 as a subset $R_e=\{r_i\}_{i=1}^{|R_e|}$1 of such opinion units (Angelidis et al., 2018). Fine-grained phrase-based work instead treats the output as aspect–sentiment–cluster structure: for each target, and for each aspect–sentiment pair, the system outputs semantically coherent phrase clusters rather than a paragraph summary (Ge et al., 2021). This representation is designed to expose sub-aspect structure inside broad aspect categories.

Abstractive systems typically verbalize a latent or selected content representation into fluent text. OpinionDigest defines review-level opinion sets as $R_e=\{r_i\}_{i=1}^{|R_e|}$2, aggregates them into an entity-level opinion inventory $R_e=\{r_i\}_{i=1}^{|R_e|}$3, selects a salient subset $R_e=\{r_i\}_{i=1}^{|R_e|}$4, and feeds a textualized version $R_e=\{r_i\}_{i=1}^{|R_e|}$5 to a Transformer trained to reconstruct reviews from extracted opinions (Suhara et al., 2020). Self-supervised abstractive models instead train on pseudo-summary pairs $R_e=\{r_i\}_{i=1}^{|R_e|}$6, where $R_e=\{r_i\}_{i=1}^{|R_e|}$7 is a set of similar reviews and $R_e=\{r_i\}_{i=1}^{|R_e|}$8 is one review treated as the target; the learning objective is the conditional negative log-likelihood of the pseudo-summary given the input review set and control tokens (Elsahar et al., 2020). Copycat-style models interpret the summary as a low-novelty representative review generated from a grouped latent-variable model, with test-time novelty suppression implemented by replacing sampled latent variables with their means (Bražinskas et al., 2019).

A distinct line of work makes evidence part of the summary itself. Rationale-based opinion summarization defines the output as $R_e=\{r_i\}_{i=1}^{|R_e|}$9, where $s_e$0 is a set of representative opinions and $s_e$1 is a set of rationale groups, one group per opinion (Li et al., 2024). This changes the target from a summary-only artifact to a summary-plus-support artifact. In comparative summarization, the target is relational rather than absolute: given two entities $s_e$2 and $s_e$3, the model generates $s_e$4, $s_e$5, and $s_e$6, explicitly decomposing opinions into entity-specific and shared content (Iso et al., 2021).

These representational choices imply different notions of fidelity. Extractive segment and phrase outputs are naturally source-grounded (Angelidis et al., 2018, Ge et al., 2021). Review-like abstractive outputs better match human-written summary style but can drift toward hallucination or pseudo-review language (Bražinskas et al., 2019, Bražinskas et al., 2020). Rationale-based outputs improve interpretability by attaching supporting snippets to each representative opinion (Li et al., 2024). Comparative and multi-perspective outputs replace a single global summary with a structured family of summaries whose content depends on viewpoint or comparison context (Iso et al., 2021, Jiang et al., 2023).

3. Learning paradigms and architectural patterns

Weak supervision is a foundational paradigm in M-OS. One influential extractive framework combines a weakly supervised aspect extractor with a sentiment predictor trained from user ratings, then ranks segments by a salience function that multiplies sentiment magnitude by aspect-specificity (Angelidis et al., 2018). FineSum reduces supervision further to aspect names, sentiment names, and a few seed keywords, then uses an opinion-oriented spherical word embedding space to generate pseudo labels for a contextual phrase classifier (Ge et al., 2021). These systems show that aspect and sentiment structure can be induced without gold summaries or token-level opinion annotations.

Unsupervised and self-supervised abstractive learning replaced gold-summary supervision with surrogate training signals. Copycat trains a hierarchical variational autoencoder to reconstruct each review from the rest of its group, using a product-level latent variable $s_e$7, review-level latent variables $s_e$8, and a pointer-generator decoder with access to the other reviews $s_e$9 (Bražinskas et al., 2019). Self-Supervised and Controlled Multi-Document Opinion Summarization constructs pseudo-summary pairs by selecting, for each target review, the $r_{1:N}=(r_1,\dots,r_N)$0 most relevant reviews of the same entity according to tf-idf cosine similarity, and trains a multi-source Transformer with standard maximum likelihood (Elsahar et al., 2020). OpinionDigest avoids gold summaries by reconstructing the original review from extracted opinion phrases, then reuses the same generator at test time with selected multi-review opinions as input (Suhara et al., 2020).

Few-shot transfer adds a small amount of gold-summary supervision to self-supervised pretraining. Few-Shot Learning for Opinion Summarization first trains a conditional Transformer LLM to generate a review given other reviews of the same product, conditioned on automatically derived properties such as content coverage, writing style, rating deviation, and length deviation, then learns a plug-in network that predicts summary-like property values from only a handful of human summaries (Bražinskas et al., 2020). This design is motivated by the observation that unsupervised abstractive models are not exposed to actual summaries and therefore fail to capture formal style, informativeness, and sentiment-preserving aggregation (Bražinskas et al., 2020).

Architecturally, multi-source conditioning has been implemented in several distinct ways. CondaSum encodes each review independently, condenses the review set into a fused review vector $r_{1:N}=(r_1,\dots,r_N)$1 by attentive pooling over review representations, and uses fused word-level representations with attention and copy in the decoder (Amplayo et al., 2019). Self-Supervised and Controlled Multi-Document Opinion Summarization extends Transformer decoding with multi-source cross-attention and compares a Parallel strategy, which attends to each review independently before averaging, with a Mean strategy, which averages keys and values first (Elsahar et al., 2020). CoCoSum uses two base LED summarizers and collaborative decoding, combining token distributions from entity-specific and common-summary models to enforce contrastiveness or commonality at decoding time (Iso et al., 2021).

Recent work has also expanded the input space. MultimodalSum adds entity-level images and metadata to multi-review summarization, using one encoder per modality and gated decoder-side fusion after a staged training pipeline in which non-text encoders are pretrained against a frozen text decoder as a pivot (Im et al., 2021). MEDOS introduces separate encoders for reviews, product description, and question-answer pairs, with gated fusion $r_{1:N}=(r_1,\dots,r_N)$2, and pairs this with a synthetic dataset creation strategy that selects pseudo-summaries using similarities to both description and QA (Siledar et al., 2024). SUBSUMM addresses very large review pools by selecting only $r_{1:N}=(r_1,\dots,r_N)$3 reviews from around $r_{1:N}=(r_1,\dots,r_N)$4 reviews on average via sentiment-aware and information-aware subset selection, then training first on sub-optimal subsets and later on higher-quality subsets with a contrastive objective (Jiang et al., 2023).

4. Evaluation resources, metrics, and benchmark design

Evaluation has become a central research problem in M-OS because reference summaries are scarce and may not faithfully represent opinion distributions. OpoSum was introduced as an opinion summarization dataset with training reviews from six domains and human-annotated development and test sets containing gold standard aspect annotations, salience labels, and opinion summaries (Angelidis et al., 2018). OpinSummEval later assembled human judgments and outputs from 14 opinion summarization models on the Yelp test set, with 100 test instances, 8 reviews per instance, and 11,200 individual scores across Aspect Relevance, Self-Coherence, Sentiment Consistency, and Readability (Shen et al., 2023). Scientific opinion summarization introduced ORSUM, with 10,989 meta-reviews and 40,903 individual reviews from 39 conference venues, split into 9,890 training, 549 validation, and 550 test examples (2305.14647). For metadata-rich M-OS, M-OS-EVAL contains 50 products, 14 model-generated summaries per product, and 14,700 ratings across seven dimensions (Attri et al., 7 Jul 2025).

A recurrent empirical result is that standard overlap metrics are unreliable for opinion summarization. OpinSummEval reports that ROUGE-1 has system-level Kendall correlations of $r_{1:N}=(r_1,\dots,r_N)$5 on Aspect Relevance, $r_{1:N}=(r_1,\dots,r_N)$6 on Self-Coherence, $r_{1:N}=(r_1,\dots,r_N)$7 on Sentiment Consistency, and $r_{1:N}=(r_1,\dots,r_N)$8 on Readability, while BARTScore$r_{1:N}=(r_1,\dots,r_N)$9 reaches $s_i$0, $s_i$1, $s_i$2, and $s_i$3 on the same dimensions (Shen et al., 2023). Scientific opinion summarization arrives at a related conclusion from a different angle: standard metrics such as ROUGE-L, BERTScore, FactCC, SummaC, and DiscoScore do not directly evaluate discussion involvement, opinion faithfulness, or decision consistency, so the paper adapts G-EVAL and GPTLikert as reference-free evaluators (2305.14647). This suggests that source-aware and task-specific evaluation is not optional in M-OS.

A second evaluation theme is representativeness rather than mere support. Automatically Evaluating Opinion Prevalence in Opinion Summarization defines a reference-free prevalence score

$s_i$4

where $s_i$5 tests whether review $s_i$6 is consistent with summary sentence $s_i$7, $s_i$8 discounts trivial statements, and $s_i$9 discounts redundant statements (Malon, 2023). On the Amazon benchmark from Bražinskas et al. (2020), the paper shows that human-authored summaries have prevalence scores of $D_i$0, $D_i$1, and $D_i$2, only slightly above random extracts at $D_i$3, $D_i$4, and $D_i$5, while a greedy extractive construction reaches $D_i$6 (Malon, 2023). The finding is not that human summaries are useless, but that reference imitation and ROUGE do not guarantee representative aggregation over many reviews.

Newer M-OS work has moved toward explicit multidimensional human evaluation. M-OS-EVAL scores Fluency, Coherence, Relevance, Faithfulness, Aspect Coverage, Sentiment Consistency, and Specificity, and prompt-based evaluators built on Omni-Prompt with GPT-4o achieve average Spearman correlation $D_i$7 with human judgments (Attri et al., 7 Jul 2025). In the comparative explainable setting, QF-CES-PROMPT evaluates clarity, faithfulness, informativeness, format adherence, and query relevance, and Llama-3.1-70B reaches an average Spearman correlation of $D_i$8 with human ratings (Attri et al., 7 Jul 2025). The broader pattern is that evaluation increasingly treats M-OS as a source-grounded, aspect-aware, sentiment-aware task rather than a conventional reference-overlap task.

5. Control, comparison, and domain expansion

Control has been a persistent theme because different users often need different slices of the same opinion pool. CondaSum proposes zero-shot customization by replacing the default fusion query $D_i$9 with a query vector $o_s=(s,A_s,\mathit{pol}_s)$0 computed from background reviews representing the desired need, such as acting, plot, positive sentiment, or negative sentiment (Amplayo et al., 2019). OpinionDigest implements customization by filtering selected opinions by aspect and/or sentiment before generation, so the same generator can produce aspect-specific or sentiment-specific summaries (Suhara et al., 2020). SUBSUMM operationalizes multi-perspective summarization directly at the dataset level: AmaSum provides pros, cons, and verdict summaries, and sentiment-aware subset selection is used to build perspective-specific inputs (Jiang et al., 2023).

Comparative summarization generalizes control into relational generation. CoCoSum formulates comparative opinion summarization as generating a contrastive summary for entity $o_s=(s,A_s,\mathit{pol}_s)$1 against $o_s=(s,A_s,\mathit{pol}_s)$2, another for $o_s=(s,A_s,\mathit{pol}_s)$3 against $o_s=(s,A_s,\mathit{pol}_s)$4, and a common summary for the pair (Iso et al., 2021). Query-Focused Comparative Explainable Summarization pushes this further by using M-OS as an intermediate representation: each product is first summarized from title, description, key features, specifications, reviews, average rating, and pricing details, and those product-level summaries are then used to generate a query-conditioned comparative table plus a final verdict (Attri et al., 7 Jul 2025). In this setting, M-OS is not the final task but a precomputable evidence abstraction layer.

The domain of application has also broadened. Template-based Abstractive Microblog Opinion Summarisation introduces microblog opinion summarisation with a dataset of 3,100 gold-standard summaries of tweets spanning a 2-year period, created by journalists following a template that separates factual information from author opinions (Bilal et al., 2022). Scientific opinion summarization defines the task of generating a meta-review from a paper’s title, abstract, and reviewer reports, explicitly requiring synthesis of strengths, weaknesses, consensus, controversy, and support for the final decision (2305.14647). Metadata-enriched e-commerce work defines M-OS over product description and QA in addition to reviews (Siledar et al., 2024), while multimodal work uses images and metadata alongside reviews (Im et al., 2021). These developments indicate that “source” in M-OS increasingly means heterogeneous evidence, not only multiple texts of the same type.

A neighboring but not fully opinion-native direction is bias-aware multi-source news summarization, where the goal is to summarize multiple news documents while sampling equally from different aspects rather than letting a majority viewpoint dominate (Rajan et al., 2023). This suggests a conceptual overlap with M-OS on viewpoint diversity and anti-majority-bias design, even though explicit opinion structure, stance modeling, and source attribution are not central there.

6. Persistent limitations and open directions

Several limitations recur across the literature. Popularity-based aggregation can privilege majority opinions and suppress minority but potentially important issues. OpinionDigest explicitly ranks clusters by size, so minority opinions may be dropped if they do not make the top $o_s=(s,A_s,\mathit{pol}_s)$5, and contradictory opinions are not explicitly reconciled (Suhara et al., 2020). Opinion prevalence evaluation likewise notes that a metric optimized for prevalence may overlook rare but critical events, such as tail risks, because it rewards statements supported by many reviews (Malon, 2023). Scientific opinion summarization makes a related point in a different domain: reviewer disagreement and minority positions matter because the final decision need not follow the majority (2305.14647).

Hallucination and faithfulness remain central difficulties for abstractive systems. Self-Supervised and Controlled Multi-Document Opinion Summarization identifies hallucination as a major problem in self-supervised M-OS and introduces metadata and inferred control tokens to reduce it (Elsahar et al., 2020). CGI$o_s=(s,A_s,\mathit{pol}_s)$6 for meta-review generation also frames hallucination as an open problem, especially when summarizing consensus and controversy over long scientific reviews (2305.14647). Multimodal and metadata-enriched settings add another failure mode: auxiliary sources can improve specificity, but they also create more opportunities to misstate unsupported facts if integration is not careful (Im et al., 2021, Siledar et al., 2024).

The quality of human references is itself contested. Opinion prevalence work shows that human-authored summaries can have only slightly better prevalence than random extracts on the tested benchmark (Malon, 2023). ORSUM reports that only 20.7% of meta-reviews contain detailed discussion on both strengths/weaknesses and agreements/disagreements, and concludes that human-written summaries are “not always reliable” (2305.14647). These results do not invalidate human evaluation, but they do weaken the assumption that a single human reference is an adequate gold standard for M-OS.

Open directions follow directly from these limitations. ORSUM points to adding author rebuttals and better hallucination detection for long scientific opinion summarization (2305.14647). SUBSUMM suggests richer aspect-aware sampling and integration with LLMs (Jiang et al., 2023). M-OS benchmark work highlights temporal patterns, multimodal content, more complete review corpora, and cross-lingual or cross-cultural evaluation as future directions (Attri et al., 7 Jul 2025). Taken together, these papers indicate that the next phase of M-OS research is likely to focus less on generic summarization performance and more on source-grounded faithfulness, controversy preservation, multi-perspective control, and evaluation protocols that explicitly model what an opinion summary is supposed to preserve.