Query-Focused Comparative Explainable Summaries

• The paper introduces QF-CES, a task that produces query-conditioned comparative summaries by generating a table of aligned product features and a final recommendation verdict.
• It employs a two-stage pipeline where multi-source product data is first compressed into opinion summaries (M-OS) and then used to generate a structured comparative explanation.
• Empirical results show a 40% reduction in latency and improved clarity, faithfulness, and informativeness, making QF-CES effective for explainable e-commerce recommendations.

Searching arXiv for the core paper and closely related QFS/QF-CES work to ground the article with current citations. Query-Focused Comparative Explainable Summaries (QF-CES) is a query-conditioned comparative summarization task in which the input is a user query together with the top-$k$ recommended items and the output is a single, side-by-side comparative summary that both aligns with the query and explains a recommendation decision. In the formulation introduced for e-commerce, $k=3$, the summary consists of a tabular comparison across the recommended products and a natural-language “final verdict” that explicitly states which product best suits the query and why. The task is positioned as personalized, privacy-preserving, recommendation engine-agnostic, and category-agnostic, and it is operationalized through a multi-stage pipeline that first compresses heterogeneous product evidence and then generates a structured comparative explanation (Attri et al., 7 Jul 2025).

1. Definition and scope

The formal input to QF-CES is a query $q$, a set of top-3 recommended products $\mathcal{P} = \{p_1,p_2,p_3\}$, and, for each product $p_i$, a multi-source evidence bundle

$$X_i = \{ \text{title}_i, \text{description}_i, \text{key\_features}_i, \text{specifications}_i, \text{reviews}_i, \text{rating}_i, \text{price}_i \}.$$

The task is to compute

$f:(q,\mathcal{P},\{X_i\}_{i=1}^3)\mapsto S,$

where $S$ is a structured comparative output composed of a comparison table and a final verdict explanation. The table aligns the three products by shared attributes; the final verdict is an NLE that directly answers the query, states which product best fits it, and explains the relevant trade-offs (Attri et al., 7 Jul 2025).

QF-CES sits at the intersection of query-focused summarization, comparative summarization, opinion summarization, and explainable recommendation. Earlier query-sensitive comparative summarization over web search results already extracted feature-relevant sentences from multiple selected pages and organized them for cross-site comparison, using concept-based segmentation over HTML DOM structure (Chitra et al., 2012). QF-CES generalizes the comparative impulse of that earlier work but differs in several respects stated explicitly in the task definition: it is query-conditioned, operates over top-3 recommended products rather than arbitrary selected URLs, consumes heterogeneous product metadata and reviews, and requires both a side-by-side table and a final verdict explanation (Attri et al., 7 Jul 2025).

The task is motivated by a limitation repeatedly noted across adjacent literatures: traditional opinion summarization usually summarizes one item at a time, often emphasizes frequent aspects in reviews, and typically omits objective metadata such as specifications and prices. QF-CES instead treats recommendation as an intrinsically comparative act. The query determines which attributes should be foregrounded, the table enforces explicit alignment across products, and the verdict converts that alignment into a recommendation decision (Attri et al., 7 Jul 2025).

2. Input representation and generation pipeline

The core methodology uses a two-step architecture. First, each product is compressed through Multi-Source Opinion Summarization (M-OS): $\text{MOS}_i = g(X_i).$ Second, the QF-CES generator consumes the query and the three product-level M-OS summaries: $S = f\big(q,\{p_i,\text{MOS}_i\}_{i=1}^3\big).$ This factorization is not incidental. It is introduced to reduce the input footprint, preserve salient subjective and objective evidence, and make downstream comparative generation tractable over long, heterogeneous source bundles (Attri et al., 7 Jul 2025).

In the M-OS stage, the input includes product title, description, key features, specifications, reviews, average rating, and price. An LLM prompt synthesizes these into a short opinion summary that integrates objective attributes with review-derived pros and cons. In the subsequent QF-CES stage, another LLM is instructed to read the query and the three M-OS summaries, select attributes dynamically based on query relevance and product specifics, produce a comparison table with the three products as columns, and generate a final verdict (Attri et al., 7 Jul 2025).

The output format is partly fixed and partly query-adaptive. Mandatory rows include product title, base price, final price, average rating, pros, and cons. Additional rows are dynamically selected according to the query and the product evidence; examples given in the task description include rows such as “Camera quality”, “Battery life”, “Gaming performance”, “Build quality & design”, and “Ease of use for elderly”. Missing values are explicitly marked as “N/A”. This combination of fixed schema and dynamic attribute selection is one of the defining technical features of QF-CES, because it couples structural comparability with query sensitivity (Attri et al., 7 Jul 2025).

The paper contrasts this pipeline with a Direct Input Approach (DIA), in which raw multi-source product data for all three products is passed directly to the comparative generator. The M-OS intermediate is introduced specifically to avoid the token and latency burden of that direct setup and to regularize the evidence presented to the final generator (Attri et al., 7 Jul 2025).

3. Relation to query-focused summarization research

QF-CES inherits its query-conditioning requirement from the broader QFS literature, but it adds explicit cross-item comparison and explanation. A useful antecedent is the coarse-to-fine QFS framework that decomposes summarization into relevance estimation, answer-likelihood estimation, and centrality estimation over explicit segment units, with final selection under length and redundancy constraints (Xu et al., 2020). That decomposition is not itself a QF-CES model, but it establishes a modular view of query focus that is congenial to explainable comparative systems: relevance can justify why a segment is on-topic, answer-likelihood can justify why it addresses the information need, and centrality can justify why it is worth including.

Neural QFS work further clarifies the design space between two-stage extractive–abstractive pipelines and end-to-end long-context models. RELREG-style two-stage systems provide explicit extracted evidence as rationales, whereas SEGENC-style long-input models fuse query-conditioned segments at decode time and reach stronger ROUGE on QMSum (Vig et al., 2021). This distinction maps naturally onto QF-CES. A two-stage design yields hard evidence and stronger explicit rationales; an end-to-end design yields higher-quality comparative narratives over longer, noisier inputs. The QF-CES formulation in e-commerce chooses a hybrid kind of modularity: an intermediate M-OS stage produces compressed evidence, and a second stage generates the comparative explanation (Attri et al., 7 Jul 2025).

Other QFS lines sharpen specific components that are relevant to QF-CES even when they are not comparative in their original form. Dual-CES decomposes saliency and focus through a dual-cascade optimization with saliency-based pseudo-feedback distillation, explicitly addressing the tension between global coverage and query specificity in unsupervised multi-document extractive summarization (Roitman et al., 2018). MaRGE learns query modeling from generic summarization corpora by using a Unified Masked Representation and Masked ROUGE Regression for evidence ranking, then conditions a generator on ranked evidence (Xu et al., 2020). QontSum augments long-input QFS with a segment scorer and InfoNCE-based contrastive learning to separate relevant segments from salient-but-not-relevant distractors (Sotudeh et al., 2023). Taken together, these methods show that modern QFS increasingly treats query alignment as an explicit latent variable rather than a by-product of generic summarization.

QF-CES also belongs to a smaller lineage of explicitly comparative summarization. Earlier work on selected web pages produced query-sensitive comparative summaries by matching feature keywords against concept blocks and extracting high-scoring sentences under section headings (Chitra et al., 2012). That earlier formulation was extractive and DOM-centric. QF-CES replaces webpage structure with heterogeneous recommendation evidence, replaces sentence extraction with LLM generation, and makes explanation a first-class output through the final verdict (Attri et al., 7 Jul 2025).

4. Data resources and evaluation protocol

The principal dataset introduced for the task is MS-Q2P (Multi-Source Query-2-Product), a proprietary e-commerce dataset. It contains 7,752 unique queries and 23,256 total product instances, corresponding to top-3 products per query. Reported corpus statistics include an average of 10 reviews per product, average specification length of 242.6 words, average review length of 17.99 words, average description length of 105.79 words, and average key-features length of 24.64 words. The covered domains include electronics, home and kitchen, sports, clothing, shoes, and jewelry (Attri et al., 7 Jul 2025).

MS-Q2P supports two linked tasks. At the product level, it supports M-OS, where multi-source product evidence is compressed into a single opinion summary. At the query level, it supports QF-CES, where three such products are compared under a user query. The mapping from queries to top-3 products is treated as an external recommendation input rather than part of the QF-CES learning problem, which is why the framework is described as recommendation engine-agnostic (Attri et al., 7 Jul 2025).

For evaluation of comparative summaries, the paper introduces CES-EVAL. Fifty queries from MS-Q2P are paired with ten generated QF-CES outputs each, and three expert annotators score every output on five dimensions: clarity, faithfulness, informativeness, format adherence, and query relevance. This yields

$k=3$0

human ratings. Annotation proceeds in two rounds: Round I yields Krippendorff’s $k=3$1, and a disagreement-resolution Round II raises agreement to $k=3$2 (Attri et al., 7 Jul 2025).

The automatic evaluation framework, QF-CES-PROMPT, uses dimension-specific LLM evaluation prompts. Rather than reading a single discrete score as definitive, the paper estimates a distribution over possible scores by repeated generation and computes the expected score

$k=3$3

where $k=3$4 is a possible discrete score and $k=3$5 is estimated from approximately 100 evaluator outputs. Metric–human alignment is then measured using summary-level correlation averaged over queries: $k=3$6 with $k=3$7 instantiated as Spearman’s $k=3$8 or Kendall’s $k=3$9. The paper explicitly uses summary-level correlation in the style of Bhandari et al. to assess whether LLM judges reproduce human ranking behavior (Attri et al., 7 Jul 2025).

5. Empirical findings

The empirical results separate naturally into three layers: M-OS quality, QF-CES generation quality, and QF-CES evaluation quality. For M-OS, six open-source LLMs are assessed with an adapted OP-PROMPT protocol over seven dimensions: fluency, coherence, aspect coverage, faithfulness, relevance, sentiment consistency, and specificity. The best-performing M-OS generator in those experiments is Mistral-7B-Instruct-0.3, and that model is then used to provide product-level summaries for downstream QF-CES generation (Attri et al., 7 Jul 2025).

For QF-CES generation itself, human evaluation shows that GPT-4o achieves the strongest overall performance, with dimension scores of 4.81 for clarity, 4.53 for faithfulness, 4.50 for informativeness, 4.61 for format adherence, and 4.32 for query relevance, for an overall average of 4.55. Among open-source systems, the strongest Qwen-based model reaches an overall average of 4.49, with particularly strong informativeness and query relevance. The same Qwen model under the Direct Input Approach scores lower, with an overall average of 4.10, indicating that the M-OS intermediate is not only a latency optimization but also a quality-preserving representation choice (Attri et al., 7 Jul 2025).

The latency comparison reinforces that interpretation. Over 50 queries, each generated 50 times, average QF-CES generation time is reported as 9.99 seconds for the M-OS pipeline and 16.55 seconds for DIA, corresponding to an inference reduction of approximately 40%. The comparison explicitly excludes one-time offline M-OS generation, because M-OS is assumed pre-computable. This result is important because QF-CES is framed not merely as an analysis benchmark but as a plausible user-facing recommendation interface (Attri et al., 7 Jul 2025).

For automatic evaluation, LLaMA-3.1-70B-Instruct is the strongest evaluator under QF-CES-PROMPT, achieving an average Spearman correlation of approximately 0.74 with human judgments across clarity, faithfulness, informativeness, format adherence, and query relevance. GPT-4o is reported as second best. The result is significant because the task is reference-free and structurally complex: the evaluator must assess a table, a verdict, cross-product alignment, and query conditioning without a gold comparative summary (Attri et al., 7 Jul 2025).

6. Explainability, limitations, and research directions

Explainability in QF-CES is partly architectural and partly presentational. Architecturally, the M-OS intermediate externalizes per-product evidence before comparative generation. Presentationally, the table makes the attribute basis of the recommendation explicit, and the final verdict converts those attribute comparisons into a query-specific recommendation. In contrast to one-sentence recommendation rationales, QF-CES exposes both the comparison substrate and the decision explanation. This is why the framework is better described as comparative explainable summarization than as a conventional recommendation explanation layer (Attri et al., 7 Jul 2025).

The present formulation nevertheless has clear limitations. The reported dataset contains only about ten reviews per product, evaluator experiments cover a limited set of LLM judges, QF-CES-PROMPT is task-specific, and complex spec-heavy products can cause incomplete or stalled outputs. The latency study, while systematic, is reported on 50 queries. The paper also notes that LLM evaluators may hallucinate judgments, particularly in complex cases, which places a practical ceiling on fully automatic evaluation without human auditing (Attri et al., 7 Jul 2025).

Several adjacent QFS strands suggest concrete extensions. Segment-level relevance and answer-likelihood decomposition would offer a more explicit rationale layer for why particular attributes or claims enter the final verdict, because the same evidence could be scored for query relevance, answer-bearingness, and centrality (Xu et al., 2020). Contrastive segment learning over long inputs suggests a way to separate query-relevant comparative evidence from salient but non-comparative distractors (Sotudeh et al., 2023). Reader-centric re-ranking with an answer reconstruction objective suggests another direction: a QF-CES output could be evaluated not only for surface faithfulness or format adherence, but for whether a modeled reader can reconstruct the answer to the original query from the generated table and verdict (Piano et al., 2024). A plausible implication is that future QF-CES systems may converge toward multi-stage architectures with explicit evidence selection, contrastive attribute grouping, and reader-aware evaluation.

A second plausible direction is domain generalization. The current formulation is e-commerce specific, but the task structure is not inherently tied to products. Earlier comparative summarization of webpages already organized feature-aligned content across multiple sources (Chitra et al., 2012), and sentiment-oriented QFES work later introduced multi-bias graph formulations that combine query relevance, sentiment alignment, and information-content regularization (Moubtahij et al., 15 Sep 2025). This suggests that QF-CES could be re-instantiated for other comparative settings—such as service plans, policy options, or scientific systems—provided that the input can be normalized into aligned evidence units and that the output preserves the dual requirement of comparison and explanation.

In that sense, QF-CES is best understood not as a narrow prompt template, but as a task formulation that fuses three previously separate requirements: query conditioning, explicit comparison, and recommendation-oriented explanation. The defining technical move is to make all three visible in the output schema itself—table for aligned evidence, verdict for decision explanation—rather than leaving comparison and explanation implicit in an unstructured summary (Attri et al., 7 Jul 2025).