Aspect-Guided Review Summarization

• Aspect-guided review summarization is a method that organizes reviews around specific aspects like battery and quality using predefined or induced taxonomies.
• It integrates diverse pipelines—extractive, abstractive, and evidence-cluster approaches—to aggregate opinions and generate concise, structured summaries.
• The field leverages supervised, unsupervised, and LLM-based strategies to balance controllability, scalability, and multi-lingual application in review analysis.

Searching arXiv for recent and foundational papers on aspect-guided review summarization. Aspect-guided review summarization is the generation of review summaries organized around product, service, or domain-specific aspects, with content selection and synthesis constrained by those aspects rather than driven solely by overall salience. In review settings, an aspect is a salient evaluative dimension such as battery, screen, service, cleanliness, or novelty, depending on the domain. The task typically decomposes into identifying or inducing aspects, associating opinions or evidence with those aspects, estimating sentiment or argumentative polarity, and producing a structured output that preserves aspect coverage and grounding. Across the literature, aspect guidance serves several distinct functions: it can act as a predefined control signal, an induced latent structure, a user query, a target for evidence aggregation, or an explicit intermediate representation used to constrain downstream generation (Wu et al., 2015, Amplayo et al., 2021, Zhou et al., 11 Jun 2025).

1. Conceptual scope and task formulation

Aspect-guided review summarization targets concise, structured summaries that report what is said about specific aspects of an entity. In product-review settings, the canonical formulation asks which aspects are discussed and what sentiments are expressed toward them, often at sentence level (Wu et al., 2015). In more recent work, the task is broadened to include evidence-backed arguments, user-selected aspects, multilingual outputs, and modular pipelines that explicitly verify intermediate aspect-specific content (Zhou et al., 11 Jun 2025, Mukku et al., 2024, Li et al., 27 Jan 2025).

A central distinction in the literature concerns the source of aspects. Some systems assume a predefined aspect taxonomy and directly map text segments to those aspects. This is the case in early convolutional approaches that use fixed labels such as battery, screen, camera, speaker, and running speed for smartphone reviews (Wu et al., 2015). Other systems induce aspects automatically from review corpora using unsupervised topic models, clustering, or weak supervision, then summarize opinions within the discovered aspect structure (Mukherjee et al., 2020, Mitcheltree et al., 2018, Zhao et al., 2019, Zhou et al., 11 Jun 2025). A third line of work treats aspects as user controls supplied at inference time, enabling personalized summaries focused on selected facets such as location and room for hotels, or staff and food for restaurants (Suhara et al., 2020, Mukherjee et al., 2020, Amplayo et al., 2021).

The output format also varies. Some systems are extractive and return representative sentences or snippets grouped by aspect (Mitcheltree et al., 2018, Mukherjee et al., 2020, Zhao et al., 2019). Others are abstractive and generate fluent aspect-conditioned summaries from selected evidence, opinion phrases, or controllers (Suhara et al., 2020, Amplayo et al., 2021, Tian et al., 2020). More recent systems emphasize attributable intermediate structures, such as aspect-sentiment-evidence tuples, representative evidence clusters, or per-aspect meta-reviews, before producing the final summary (Zhou et al., 11 Jun 2025, Li et al., 27 Jan 2025).

2. Aspect representation and control regimes

A major axis of variation is whether aspect guidance is predefined, induced, or refined dynamically.

In predefined-aspect systems, aspect mapping is a supervised classification problem. "Aspect-based Opinion Summarization with Convolutional Neural Networks" formalizes this as direct sentence-to-aspect mapping rather than unsupervised discovery, arguing that predefined aspects are less general but more precise for particular products (Wu et al., 2015). The same principle appears in weakly supervised and user-guided classification settings, where a small number of seed words or manually defined aspect schemas support downstream summarization (Li et al., 2020, Li et al., 27 Jan 2025).

In induced-aspect systems, aspects emerge from the corpus. Tourist-review summarization uses Attention-based Aspect Extraction (ABAE) to learn coherent aspects unsupervised, then manually consolidates them into eight coarse classes such as Attractions, Access, Amenities, and Culture (Mukherjee et al., 2020). Airbnb summarization and personalization similarly use ABAE or k-means over sentence embeddings to discover aspect distributions that can drive canonical sentence selection and user profiles (Mitcheltree et al., 2018). ASESUM goes further by avoiding any predefined taxonomy: it initializes aspect candidates with product-category prompts, normalizes them with DBSCAN, and treats the resulting clusters as the active aspect set for argument extraction and ranking (Zhou et al., 11 Jun 2025).

In controllable systems, aspect guidance is an explicit user input. OpinionDigest filters extracted aspect-opinion-sentiment triples before generation, so the summary can focus on a specified aspect and/or sentiment (Suhara et al., 2020). AceSum introduces binary aspect queries over a learned aspect inventory and conditions a pretrained summarizer on aspect codes, aspect-relevant sentences, and aspect keywords (Amplayo et al., 2021). In tourism, the ILP objective changes as the selected aspect set $A^r$ changes, because the sentence relevance term is computed against user-selected aspect embeddings (Mukherjee et al., 2020).

A plausible implication is that aspect guidance functions simultaneously as a semantic decomposition mechanism and as a controllability interface. In systems built around fixed taxonomies, control is explicit but domain maintenance is required. In systems based on induced aspects, domain adaptability improves, but aspect granularity and label stability become a primary concern (Zhou et al., 11 Jun 2025, Mukherjee et al., 2020).

3. Core pipeline architectures

Despite considerable architectural diversity, most aspect-guided summarization systems instantiate some variant of a four-stage pipeline: aspect assignment, opinion or evidence extraction, aggregation or ranking, and summary construction.

The sentence-classification pipeline is exemplified by the CNN-based AOS system. A tokenized sentence is embedded as $X = [x_1, \dots, x_n]$, processed by 1D convolutions and max-over-time pooling, then classified with a softmax layer. The cascaded CNN uses aspect classifiers at level 1 and a shared sentiment classifier at level 2; the multitask CNN shares word embeddings across aspect and sentiment tasks (Wu et al., 2015). This design treats summarization as downstream aggregation over sentence-level aspect and sentiment predictions.

The extractive optimization pipeline appears in tourist-review summarization. Sentences are scored by

$$\operatorname{Opin\_Score}(s) = \operatorname{Readability}(s) \times |\operatorname{Sentiment}(s)-2| \times \operatorname{Relevance}(s),$$

where relevance is the maximum cosine similarity between sentence words and user-selected aspect embeddings. Summary selection is then formulated as an ILP with length, redundancy, and demographic-balance constraints (Mukherjee et al., 2020). AspMem uses a related extractive formulation, but replaces explicit aspect labels with an aspect-aware memory initialized from external product descriptions; salience combines aspect relevance and sentiment, and selection is solved with ILP under redundancy and budget constraints (Zhao et al., 2019).

The extract-then-generate pipeline is prominent in abstractive systems. OpinionDigest first extracts tuples $(o_i, pol_i, a_i)$ from reviews using ABSA, clusters similar opinion phrases, selects the most popular clusters, textualizes the selected opinions with separators, and trains a Transformer to reconstruct reviews from these opinion sets. At inference time, the generator is conditioned only on selected aspect-sentiment opinions, which directly implements aspect control (Suhara et al., 2020). AceSum similarly constructs synthetic training pairs with aspect controllers derived by hierarchical multi-instance learning, then fine-tunes T5-small to generate summaries conditioned on aspect codes, keywords, and sentences (Amplayo et al., 2021).

The evidence-cluster and argumentation pipeline is best represented by ASESUM. It extracts tuples $\langle a, s, x\rangle$ using a Review Argument Scheme, clusters arguments by evidence similarity, replaces each argument’s evidence with a representative evidence snippet $X_c$ selected by TextRank, and ranks clusters using support and contradiction relations defined within each evidence cluster (Zhou et al., 11 Jun 2025). This yields summaries composed of top-ranked, attributable evidence statements rather than free-form generation.

The modular verified pipeline decomposes the problem further. "Decomposed Opinion Summarization with Verified Aspect-Aware Modules" separates aspect identification, opinion consolidation, and meta-review synthesis, with each intermediate output evaluated independently (Li et al., 27 Jan 2025). This suggests a movement away from monolithic end-to-end generation toward inspectable, aspect-specific intermediate reasoning.

4. Modeling strategies

4.1 Supervised neural aspect mapping

Supervised aspect mapping is characteristic of early neural work. In the CNN formulation, each predefined aspect is modeled with a binary classifier, and sentiment classification is shared because few sentences express opposite sentiments about different aspects (Wu et al., 2015). The multitask variant optimizes a combined objective

$$L_{\text{total}} = \sum_{a=1}^A \lambda_a L_{\text{aspect}}^{(a)} + \lambda_s L_{\text{sent}} + \gamma \lVert \Theta \rVert_2^2,$$

with shared embeddings across tasks, yielding better generalization under limited aspect-specific data (Wu et al., 2015).

Typed abstractive decoding offers another explicit aspect mechanism. In "Aspect and Opinion Aware Abstractive Review Summarization with Reinforced Hard Typed Decoder", the decoder predicts a word type $t_t \in \{A,O,C\}$ at each step—aspect, opinion, or context—and restricts generation and copy to the corresponding vocabulary partition. The reinforced hard typed decoder then improves type prediction with REINFORCE, producing better ROUGE than PGNet and softer typed baselines (Tian et al., 2020). Although this model does not classify review text into aspects, it operationalizes aspect guidance as lexical control during generation.

4.2 Unsupervised aspect induction

Unsupervised topic induction remains central in domains with limited annotation. The tourist-review system trains ABAE on pooled TripAdvisor reviews, selects $K=25$ by maximum average coherence, and then maps fine-grained topics into eight coarse aspect classes (Mukherjee et al., 2020). Airbnb work compares ABAE with k-means and LDA, showing that a simple k-means baseline is strong for frequent aspects while ABAE helps on rarer aspects such as cleanliness (Mitcheltree et al., 2018).

Weakly supervised induction can also come from external product knowledge. AspMem mines seed words from feature descriptions with TF-IDF, stores them in aspect memory cells, computes word-level attention from cosine similarity to those memories, and uses the resulting aspect-aware segment embeddings both for aspect identification and extractive summary selection (Zhao et al., 2019). ARYA, although centered on aspect classification rather than summarization, addresses the challenge of a noisy misc aspect through entropy-based supervision induction and iterative seed filtering, which is directly relevant when aspect labels feed an aspect-wise summarizer (Li et al., 2020).

4.3 LLM-based decomposition and prompting

Recent systems rely heavily on prompting rather than supervised training. ASESUM uses LLMs to instantiate the Review Argument Scheme and extract aspect, sentiment, and evidence in constrained JSON form (Zhou et al., 11 Jun 2025). MARS decomposes multilingual extraction into four prompting phases—aspect identification, sentiment classification, verbatim extraction, and translation—before summarization (Mukku et al., 2024). Prompted Opinion Summarization with GPT-3.5 uses pipelines such as topic clustering, extractive filtering, and recursive chunk summarization to produce aspect-specific summaries for hotel reviews (Bhaskar et al., 2022). PAKPA prompts LLMs both for aspect-based sentiment analysis and for single-aspect key-point generation, yielding quantitative summaries whose prevalence comes directly from cluster sizes (Tang et al., 2024).

A notable trend is that prompting often reintroduces explicit structure: JSON schemas, aspect definitions, sentence IDs, representative evidence, or exact aspect counts. This suggests that LLM-based aspect-guided summarization performs best when the aspect layer is not merely implied by the prompt but represented as a concrete intermediate object.

5. Evidence aggregation, ranking, and summary construction

A defining feature of aspect-guided review summarization is that the summary is not simply generated from raw reviews; it is built from aspect-organized evidence.

In sentence-level AOS, aggregation is typically statistical. Given the set $\mathcal{D}_a$ of sentences mapped to aspect $X = [x_1, \dots, x_n]$0, one can compute discrete or probability-weighted sentiment scores over those sentences, plus positive and negative counts and representative quotes (Wu et al., 2015). This creates aspect hubs or dashboards rather than only free text.

In optimization-based extractive systems, aggregation is driven by salience and diversity. Tourist-review summarization uses sentence-level readability, sentiment strength, and aspect relevance, then penalizes redundancy and fairness deviation in the ILP objective (Mukherjee et al., 2020). AspMem likewise combines aspect relevance and sentiment into a salience score and performs ILP selection under redundancy and length constraints (Zhao et al., 2019). SOLAR prioritizes app reviews with a two-level ranking system: first topics are scored by topic proportion, negativity, average rating, and freshness; then reviews within topics are scored by rating, freshness, sentiment probabilities, length, and topic influence (Gao et al., 2022).

In evidence-grounded LLM pipelines, aggregation often occurs through clustering. ASESUM clusters semantically similar evidence, unifies citations by replacing all evidence in a cluster with a TextRank-selected representative snippet, and ranks clusters by the best argument score within the cluster (Zhou et al., 11 Jun 2025). PAKPA similarly clusters aspect terms under fixed sentiment, prompts the LLM to generate one concise key point per cluster, and defines prevalence as cluster size rather than via post hoc sentence–summary similarity (Tang et al., 2024). AaKOS constructs weighted knowledge graphs from review clusters, where edge weights encode the proportion of mentions of attributes within aspect clusters, then filters the graph by user-selected aspects before generation (Wang et al., 2023).

In abstractive controllers, construction is conditioned on selected support units. OpinionDigest textualizes top opinion phrases with [SEP] markers and trains a Transformer on the reconstruction objective

$X = [x_1, \dots, x_n]$1

so that summarization becomes verbalization of aggregated aspect-opinion content (Suhara et al., 2020). AceSum’s control prefix serializes aspect codes, keywords, and aspect-relevant sentences into the encoder input, making summary construction explicitly controller-driven (Amplayo et al., 2021).

6. Evaluation dimensions and empirical findings

Evaluation in this area spans aspect identification, sentiment accuracy, summary overlap, grounding, diversity, controllability, and sometimes downstream utility.

The early CNN AOS system reports sentence-level aspect mapping F1 and sentiment accuracy on Amazon smartphone reviews. Across aspects, CNNs improve over SVM by 0.5–2.4 F1 points, pre-trained embeddings add +0.7 to +1.7 F1, and multitask CNN with word2vec is best overall. Sentiment accuracy rises from 80.3 for SVM to 84.1 for M-CNN with pre-trained embeddings (Wu et al., 2015). These results established that neural aspect mapping could materially improve the front end of aspect-guided summarization.

Tourism work evaluates both automatic ROUGE and human judgments. The ILP-based controllable method performs strongly among extractive baselines and is preferred in crowdsourced evaluation for aspect coverage, readability, and diversity; micro-averaged human preferences reach 63% for aspect coverage, 61% for readability, and 65% for diversity on all-aspect summaries (Mukherjee et al., 2020). Airbnb summarization instead evaluates whether retrieved sentences actually match their assigned aspect, with Fleiss’ $X = [x_1, \dots, x_n]$2 for sentence-level evaluation, and shows that k-means surprisingly outperforms ABAE on frequent aspects while ABAE helps on rarer ones (Mitcheltree et al., 2018).

Abstractive systems show more mixed automatic and human outcomes. OpinionDigest improves over LexRank and MeanSum on Yelp, achieving ROUGE-1 29.30, ROUGE-2 5.77, and ROUGE-L 18.56, and also improves content support relative to MeanSum, with 29.77% of summary sentences fully supported versus 23.25% (Suhara et al., 2020). The reinforced hard typed decoder consistently outperforms Seq2Seq, PGNet, STD, and HTD on two Amazon datasets, reaching ROUGE-1 31.97 and ROUGE-L 30.11 on Electronics (Tian et al., 2020).

Grounding-oriented systems emphasize consistency and faithfulness metrics beyond ROUGE. ASESUM reports higher SummaC SCin than HERCULES in all tested domains, with an average improvement of +5.84, and also generally higher diversity (Zhou et al., 11 Jun 2025). MARS reports large gains over multilingual baselines at both aspect and overall levels, for example ROUGE-1 41.7 and BERTScore 0.81 at aspect level, along with strong human ratings for aspect-specificity, factuality, coverage, fluency, and brevity (Mukku et al., 2024). PAKPA reaches quantification precision 0.95 on Yelp, substantially above ABKPA, SKPMBase(IC)+, RKPA-Base, and Enigma+, and scores best in human evaluation across coverage, faithfulness, redundancy, validity, sentiment clarity, informativeness, and single-aspect quality (Tang et al., 2024).

Controllability itself has become an evaluation target. AceSum reports that, on Space, 50.9% of aspect-specific outputs discuss the requested aspect exclusively and 61.3% of two-aspect outputs cover both requested aspects (Amplayo et al., 2021). Prompted GPT-3.5 work argues that ROUGE and BERTScore inadequately capture opinion summarization quality and introduces entailment-based metrics for factuality, faithfulness, and semantic genericity, showing higher correlation with human judgments than overlap metrics (Bhaskar et al., 2022).

This suggests a broad methodological shift: aspect-guided review summarization is increasingly evaluated not only as summarization, but as structured evidence compression under controllability and grounding constraints.

7. Extensions, applications, and open issues

The literature has expanded aspect-guided review summarization far beyond its original e-commerce formulation.

One extension is domain generalization. Scientific-review summarization adapts the aspect concept to rhetorical facets such as Challenge, Approach, and Outcome. ACLSum provides expert-annotated extractive and abstractive references for these facets and shows that end-to-end aspect-based summarization is generally strong, while extract-then-abstract can be advantageous when extraction is accurate (Takeshita et al., 2024). The modular verified pipeline extends this further to scientific papers, businesses, and products, showing that aspect-aware decomposition yields high aspect coverage and grounded intermediate outputs that also help humans write meta-reviews more efficiently (Li et al., 27 Jan 2025).

Another extension is multilingual and large-scale deployment. MARS handles English, Spanish, French, German, and Italian by extracting structured multilingual insights and summarizing translated verbatims in a target language, while also achieving average latency improvements of 92.5% relative to an off-the-shelf LLM baseline (Mukku et al., 2024). Wayfair’s large-scale production system combines LLM-based ABSA, aspect consolidation, stratified review sampling, and evidence-guided prompting, and reports online improvements of +0.3% visit-level Add to Cart Rate, +0.5% conversion rate, and a 0.13% drop in customer-level bounce rate at $X = [x_1, \dots, x_n]$3 (Boytsov et al., 30 Sep 2025). It also releases 11.8M anonymized reviews across 92,000 products with extracted aspects and summaries (Boytsov et al., 30 Sep 2025).

A recurring controversy concerns predefined versus induced aspect sets. Predefined aspects offer interpretability, stable control, and efficient supervised modeling, but they require domain maintenance and may miss emergent facets (Wu et al., 2015, Suhara et al., 2020). Induced-aspect systems adapt better across products and domains, but risk fragmentation, unstable labels, and mismatched granularity (Zhou et al., 11 Jun 2025, Mitcheltree et al., 2018). A plausible implication is that hybrid systems—predefining coarse taxonomies while allowing controlled emergence of fine-grained subaspects—may offer a better practical trade-off, and several systems already move in that direction through clustering, normalization, or taxonomy expansion (Zhou et al., 11 Jun 2025, Mukku et al., 2024).

Another persistent issue is multi-aspect sentences. Early AOS assumes that few sentences express opposing sentiments about different aspects, which justifies a single sentiment classifier for all aspects (Wu et al., 2015). Later work explicitly notes that this assumption can fail and recommends aspect-conditioned sentiment models or token-level alignment when multi-aspect conflicts are common (Wu et al., 2015, Tang et al., 2024). This remains a nontrivial modeling problem because aspect guidance is often imposed at sentence or snippet granularity, while real review language is frequently entangled.

Finally, aspect-guided summarization increasingly intersects with quantitative summarization, personalization, and human-in-the-loop analysis. PAKPA turns aspect clusters into prevalence-counted key points (Tang et al., 2024). Tourist-review systems encode user-selected aspects, length budgets, and demographic balance (Mukherjee et al., 2020). Verified modular pipelines provide intermediate artifacts that reduce human summarization time by 14.7% in one study (Li et al., 27 Jan 2025). These developments position aspect guidance not merely as a summarization constraint, but as a general interface for making large review corpora inspectable, steerable, and operationally useful.