Text-Based Explainable Recommendation

• Text-based explainable recommendation is a paradigm that generates human-readable justifications for user–item pairs, boosting trust and user satisfaction.
• Modern methods integrate neural generative models and retrieval-augmented techniques to produce detailed, multi-aspect, and factually-grounded explanations.
• Emerging trends focus on sentiment alignment, fairness through counterfactual reasoning, and adaptive explanation strategies to address data sparsity and bias.

Text-based explainable recommendation encompasses algorithmic frameworks and methodologies for generating natural language rationales that justify recommendation decisions, with the goal of enhancing transparency, trust, and user satisfaction in recommender systems. The evolution of this field has transitioned from template- and feature-based justification to neural generative models—including LLMs—and now foregrounds concerns of factuality, fairness, sentiment alignment, and multi-aspect reasoning as critical elements in explanation quality.

1. Foundations and Definitions

Text-based explainable recommendation refers to models that, given a user–item pair (and often associated context such as ratings, reviews, or features), produce a personalized, human-readable explanation for why a particular item is recommended to the user. This process typically involves three core steps: user and item representation learning, preference prediction, and conditional natural language generation. Early frameworks focused on structured, template-based output or ranking predefined textual rationales, whereas contemporary approaches leverage neural architectures for free-form generation, often in a multi-task regime where recommendation and explanation modules are co-optimized (Zhang, 2017, Peng et al., 2024, Wang et al., 2018).

The main objectives of text-based explainable recommendations are to:

• Increase user satisfaction and trust through transparent reasoning (Zhang, 2017, Kabongo et al., 30 Dec 2025).
• Provide actionable insight into the underlying user–item match, enhancing domain understanding and supporting user decision-making (Wang et al., 2018).
• Support regulatory or fairness requirements by exposing model logic and potential biases (Wang et al., 2022).

2. Core Modeling Paradigms

The dominant modeling paradigms for text-based explainable recommendation can be categorized as follows:

a. Template and Feature-Level Rationale

Template systems use event-driven slot filling based on extracted product or service features and user-item feature-specific preference scores. For instance, Explicit Factor Models (EFM) leverage phrase-level sentiment analysis to connect feature-level user attention and item quality into concise rationales (Zhang, 2017). Multi-task models such as MTER use tensor factorization to jointly predict preferences and generate opinionated content by scoring the likelihood of a user using a specific opinion phrase about an item’s feature (Wang et al., 2018).

b. Generative Neural Approaches

Hierarchical sequence-to-sequence architectures (e.g., HSS) and Transformer-based models facilitate free-form generation of personalized explanations (Chen et al., 2021, Ma et al., 2024). These often integrate user and item embeddings (derived from collaborative filtering, review aggregation, or interaction summaries) with content-aware mechanisms such as feature-aware attention, aspect enhancement modules, and retrieval augmentation (Cheng et al., 2023, Liu et al., 8 Jul 2025). Newer designs can exploit LLM prompt engineering (continuous prompts, MoE adapters) to natively condition language generation on collaborative signals, ratings, and aspect profiles (Peng et al., 2024, Ma et al., 2024, Liu et al., 8 Jul 2025).

c. Sentiment and Consistency Alignment

Modern frameworks increasingly couple the prediction of ratings and the generation of explanations, enforcing alignment of output sentiment with the predicted preference via dedicated latent vectors, auxiliary loss terms, or constrained decoding (Yang et al., 2021, Liu et al., 21 Feb 2025, Shimizu et al., 2024). Recent models verify not only surface-level fluency but also match to underlying ratings and post-purchase sentiment extracted from user reviews (Shimizu et al., 2024).

d. Counterfactual and Fairness-Based Explanations

COFFEE and counterfactual generation frameworks approach fairness in explanation by enforcing parity of explanation quality or informativeness across protected attributes (e.g., gender, price tier), using adversarial or policy-gradient regularization (Wang et al., 2022, Ranjbar et al., 2023). Counterfactual explanations identify minimal edits to user or item features (including text) required to alter the recommendation outcome; such approaches provide both model-oriented and user-understandable rationale for item ranking (Ranjbar et al., 2023).

3. Advanced Architectures and Emerging Trends

a. LLM-Driven and Profile-Based Reasoning

Recent designs, including PGHIS, XRec, and MADRec, leverage the language understanding capabilities of large pretrained LLMs by embedding detailed user and item textual profiles—generated through hierarchical interaction summarization or unsupervised aspect extraction—directly into the LLM’s input prompt (Liu et al., 8 Jul 2025, Ma et al., 2024, Park et al., 15 Oct 2025). This allows the LLM to reason natively over high-level preferences and item attributes without information loss from dense embeddings. Contrastive prompting (CPEG) is used to produce high-quality, discriminative ground-truth rationales (Liu et al., 8 Jul 2025).

b. Multi-Aspect and Adaptive Explanation

Frameworks like MADRec and HAG emphasize extracting multiple aspects (categories) from reviews using unsupervised clustering and attention-based mechanisms, summarizing these aspects and feeding them into downstream LLMs or hierarchical decoders for both ranking and explanation tasks (Park et al., 15 Oct 2025). Multi-aspect summaries support both static and sequential recommendation subtasks, with self-feedback loops used to iteratively refine candidate selection based on missing ground-truth matches.

c. Retrieval-Augmented Generation and Information Bottlenecking

Hybrid retrieval–generation systems such as ERRA enhance generation quality and stability by retrieving aspect-relevant or semantically similar review sentences using dense retrieval models (e.g., Sentence-BERT), injecting those into the generative model to overcome data sparsity and capture nuanced context (Cheng et al., 2023). Geometric Information Bottleneck methods (GIANT) inject graph-induced cluster priors into variational structures, constraining latent factors with global interaction structure and enabling coherent topic-based explanation extraction (Yan et al., 2023).

4. Evaluation Protocols and Metrics

A broad suite of metrics has developed to assess explanation quality, including:

• Textual Overlap: BLEU, ROUGE, and BERTScore measure n-gram or embedding overlap with gold explanations (Chen et al., 2021, Shimizu et al., 2024).
• Feature and Aspect Coverage: Feature Matching Ratio (FMR), Feature Coverage Ratio (FCR), and Feature Diversity (DIV) quantify how well generated texts mention salient item aspects or features (Peng et al., 2024, Shimizu et al., 2024).
• Factual Consistency: Recent work highlights the limitations of overlap-based metrics, proposing statement-level alignment via LLM entailment or NLI models for factual precision and recall, and constructing augmented datasets to facilitate such evaluation (Kabongo et al., 30 Dec 2025).
• Sentiment Alignment: The alignment between explanation sentiment and predicted or true ratings is measured quantitatively (e.g., via a sentiment regressor or LLM-based labeling) (Shimizu et al., 2024, Liu et al., 21 Feb 2025, Yang et al., 2021).
• Fairness: Individual and group-level counterfactual fairness metrics (e.g., Ind-CF, Grp-CF, DDP) assess explanation performance parity across protected groups (Wang et al., 2022).
• User-Centric and Persuasiveness: Human evaluation tasks, such as agreement rate and helpfulness up-vote, assess how well explanations support user decision-making or correspond to the model's actual ranking (Yang et al., 2021, Park et al., 15 Oct 2025).

5. Challenges: Factuality, Fairness, Coherence, and Personalization

While neural methods generate fluent explanations and achieve high scores on semantic-similarity metrics, statement-level evaluations reveal persistent issues:

• Factual Consistency: Models frequently hallucinate, omit, or misattribute critical aspects, as revealed in large-scale statement-alignment benchmarks—LLM-based statement-level precision on Amazon Reviews remains as low as 4–33% for SOTA models, independent of high BERTScore (Kabongo et al., 30 Dec 2025).
• Sentiment and Attribution: Generated rationales do not consistently align with user sentiment or rating unless explicit conditioning is imposed, and sentiment alignment remains a separate challenge from general fluency (Shimizu et al., 2024, Liu et al., 21 Feb 2025).
• Fairness and Bias: Attribute-specific disparities in explanation informativeness (e.g., FeatCov bias by gender) persist without explicit fairness constraints (Wang et al., 2022).
• Personalization and Data Sparsity: Retrieval and aspect-enhancement modules have proven critical for boosting both accuracy and explanation quality, especially for users or items with limited observed interactions (Cheng et al., 2023).

A summary of key models and challenges is provided below:

Model/Class	Salient Mechanism	Addressed Challenge
HSS, NRT	Hierarchical GRU/RNN	Free-text explainability, denoising
PEPLER, PETER, Attn2Seq	Transformer decoder	Scaling, context-awareness
ERRA, XRec, PGHIS	Retrieval, profile-injection	Cold-start, context richness
COFFEE, Gumbel	Counterfactual/fairness	Demographic parity, influence identification
MADRec, HAG, CPEG	Aspect extraction, contrastive prompting	Multi-aspect, high-quality ground truth
SAER, CIER, CER	Sentiment/rating alignment	Sentiment coherence, numerical–textual alignment
GIANT	Geometric prior IB	Global feature integration, topic faithfulness
(Kabongo et al., 30 Dec 2025, Shimizu et al., 2024)	Factual alignment and sentiment disentanglement	Factual consistency, evaluative validity

6. Limitations and Future Directions

Despite advances, the field confronts several fundamental issues:

• Surface-Level vs. Factual Validation: BERTScore, BLEU, and ROUGE do not proxy for factual accuracy or sentiment alignment; explicit statement-level or aspect-coverage metrics are necessary to avoid hallucinated or misleading rationales (Kabongo et al., 30 Dec 2025, Shimizu et al., 2024).
• Factuality-Aware Training: Future systems are expected to incorporate factual grounding objectives (e.g., contrastive losses, NLI-based filtering), end-to-end joint learning with statement-supervision, and retrieval-augmented generation grounded in actual user history or evidence (Kabongo et al., 30 Dec 2025, Liu et al., 8 Jul 2025).
• Multi-aspect, Multi-modal, and Adaptive Explanation: Ongoing work focuses on unsupervised aspect extraction, adaptation to sequential preference evolution, and integration of explanations across modalities (e.g., images + text).
• Scalability and Efficiency: LLM and contrastive methods, while effective, are computationally intensive. Research into lighter summarization, context management, and precomputation is underway (Liu et al., 8 Jul 2025, Park et al., 15 Oct 2025).
• Fairness, Debiasing, Personalization: Future approaches will further explore disentanglement of protected attributes, group-level evaluation, and causal recording of model decisions to support personalization without bias (Wang et al., 2022).

Recent methodologies and benchmarks now provide the means to objectively and rigorously evaluate text-based explanations along multiple technically substantive axes, laying the groundwork for scalable, trustworthy, and user-centric explainable recommender systems.