2.1. Customer Requirement Extraction

E-commerce platforms and social media have become key channels for consumers to share product opinions, with user-generated content (UGC) containing valuable insights into product experiences. A comprehensive understanding of CRs is crucial for successful new product design ¹. Traditional requirement mining methods, such as the voice of customer framework by Griffin and Hauser ¹⁹, rely on focus group interviews but suffer from small sample sizes and subjectivity. With the growth of online reviews, unstructured post-purchase UGC offers abundant requirement-related information, although extracting requirements from large volumes of unstructured text remains a challenge ²⁰.

Existing approaches include the use of word frequency counting and co-occurrence analysis to identify core product attributes by Büschken and Allenby ¹¹ that lacks precision and ignores the influence of sentiment. Ferguson et al. ²¹ developed an ergonomics-focused phrase set to extract design-specific requirements (for example, earplug dimensions) from reviews. Latent Dirichlet allocation (LDA), an unsupervised method, is widely used in review analysis for its simplicity. For example, Liu et al. ²² applied LDA to pharmacy reviews for satisfaction analysis. Guo et al. ²³ combined topic modeling and sentiment analysis to prioritize service improvements using more than 20,000 hotel reviews. However, LDA performance degrades significantly on long texts; a root cause lies in its word-bag assumption that ignores contextual dependencies and semantic coherence ²⁰. In long reviews, product attributes and related sentiments are often scattered across sentences, leading to sparse topic distributions and ambiguous clustering results. Recent studies ²⁴ further confirm that LDA fails to capture the hierarchical semantic structure of long texts, resulting in overlapping topic clusters and reduced attribute extraction accuracy.

Timoshenko and Hauser ²⁵ proposed a machine learning-based UGC analysis framework that was more efficient than traditional methods but required manual encoding of 12,000 UGC sentences, limiting its use in low-engagement domains. Tucker and Kim ²⁶ tracked temporal shifts in consumer preferences using machine learning, and Long et al. ²⁷ developed a framework linking core requirements to design strategies. Zhan et al. ²⁸ identified negative themes in pharmacy reviews via sentiment analysis but overlooked critiques in positive or neutral feedback. Zhang et al. ²⁹ used a deep learning-based embedding ensemble to mine innovative ideas from more than 10,000 reviews with 91% accuracy, although dichotomizing requirements into “innovative” and “non-innovative” provided limited actionable insights.

In recent years, lightweight transformer models have emerged as a pivotal solution for resource-constrained scenarios, achieving remarkable advancements via architectural innovations and efficiency-oriented optimizations. Representative models such as CloFormer ³⁰, designed for computer vision tasks, adopt enhanced local perception mechanisms to reduce parameters to merely 42k and FLOPs to 0.6G while attaining a Top-1 accuracy of 77.0% on ImageNet1K, outperforming conventional lightweight vision models ³⁰. TokenFormer ³¹ introduces a parameter-tokenization strategy, replacing fixed linear projections with attention-driven token-parameter interactions to enable seamless incremental scaling from 124M to 1.4B parameters without full retraining, drastically cutting training costs while maintaining performance parity ³². For specialized tasks such as sign language recognition, TSLFormer ³² leverages a compact two-layer Transformer architecture with 144-dimensional features, achieving 90.67% accuracy on the AUTSL dataset and enabling real-time inference on mobile devices with a 0.3-s processing latency per second of video. Despite these gains, trade-offs persist between extreme lightweighting and the modeling of complex contextual or spatiotemporal dependencies, highlighting avenues for further optimization in efficient architecture design.

2.2. Customer Requirement Processing

After extracting the requirements from UGC, processing and analysis are required to translate the insights into product design guidance. The Kano model and its variants CRs. Li et al. ³³ proposed an importance-Kano analysis for online fresh products, highlighting consumer focus on shopping experience and quality. Bi et al. ³⁴ integrated sentiment analysis with the Kano model to identify factors influencing satisfaction. Sun et al. ³⁵ introduced a CM-Kano model incorporating both customer and manufacturer perspectives. However, the Kano model relies on questionnaire data, limiting large-scale review mining, and its qualitative criteria are subjective ³⁶.

Alternative studies ^37,38 have used machine learning to construct knowledge graphs from fragmented data, supporting conceptual design knowledge acquisition. Knowledge graphs connect diverse data by representing entities and relationships, integrating sources, and visualizing information flows ³⁹. Although useful for storing requirement data, they require extensive expert input and manual annotation, and their word/phrase-level nodes capture limited semantic information. Thus, this study uses deep learning for requirement classification to preserve UGC semantics.

Sentiment and attention analysis of UGC is common ⁴⁰, and assessing customer satisfaction with each attribute is critical ⁴¹. Traditional methods (for example, questionnaires and rule-based approaches) ^42,43 lack generalizability owing to domain-specific lexicons. Machine and deep learning dominate sentiment analysis. Imtiaz and Islam ⁴⁴ used mobile phone reviews to link product attributes to satisfaction, informing design strategies. Zhao et al. ⁴⁵ prioritized attributes by considering sentiment and attention levels. This study builds on this by integrating inter-attribute relationships to provide more comprehensive design recommendations.

2.3. Fundamental Techniques

Hybrid neural networks have attracted considerable attention in NLP. Li et al. ⁴⁶ proposed a BERT–BiLSTM model for sentiment analysis of Chinese buzzwords, whereas the RoBERTa–BiLSTM framework by Rahman et al. ⁴⁷ achieved competitive results on three sentiment analysis benchmarks. Another Li et al. ⁴⁸ used a BiLSTM–CNN hybrid model with multi-feature fusion to enhance community question answering text matching. Deng et al. ⁴⁹ developed an attention-based BiLSTM–CNN model with a gating mechanism for medium-to-long text classification.

Building on these, this study proposes BERT–BiLSTM–hierarchical position-aware network (HPANet), integrating BERT for deep contextual understanding and BiLSTM for capturing sequential dependencies to extract fine-grained CRs from reviews.

Among pre-trained models (for example, GPT, ERNIE, XLNet, and T5), BERT and its derivatives excel in text understanding tasks. This framework uses the BERT-base-Chinese model for its balance of accuracy and efficiency in processing Chinese text.

BERT is a deep multi-layer bidirectional transformer encoder. As shown in Fig. 1, it processes inputs through embedding and encoding, where tokens are represented as the sum of token (semantic content), segment (distinguishing paired sequences), and position (token order) embeddings.

Composite embeddings pass through transformer encoder layers, each with multi-head self-attention and a position-wise feed-forward network, enhanced by residual connections and layer normalization to capture bidirectional contextual dependencies. The final hidden state of a [CLS] token represents sequences for classification, whereas the states of other tokens support downstream tasks (for example, named entity recognition). Fig. 1 illustrates the transformer blocks, where the embeddings are refined using attention and feed-forward networks.

BiLSTM follows the BERT encoder to model sequential dependencies and long-range contexts, capturing both forward and backward information that is critical for understanding the surrounding words in a text.

For a BERT-generated sequence of contextualized token embeddings \(\{h_{1}, h_{2}, \dots, h_{n}\}\), BiLSTM processes it in two directions as follows:

The final hidden state at time \(t\) concatenates forward and backward states as follows:

This bidirectional structure captures dependencies in past and future contexts, enhancing detection of nuanced CRs in free-form reviews.

3.1. CRs Definition

3.1.2. Product Requirement Mining

To filter product-irrelevant terms and identify CR categories, clustering analysis is applied to the extracted keywords, enabling the recognition of distinct product characteristic classes. For example, in smartphone reviews, terms such as “performance,” “speed,” and “processor” may cluster into a “performance” category, while “appearance,” “design,” and “weight” may form a “design aesthetics” category.

The \(K\)-means algorithm, a widely used unsupervised learning method, groups vocabulary into clusters based on semantic similarity. Its goal is to partition data into \(k\) clusters by maximizing the intra cluster cohesion and minimizing inter-cluster divergence that is achieved by iteratively optimizing the objective function.

\begin{equation} J(\mu, C) = \sum_{i=1}^{n} \min_{j \in \{1, 2, \dots, K\}} \left\lVert \chi_{i} - \mu_{j} \right\rVert^{2}, \label{eq:kmeans} \end{equation}

where \(\mu = \{\mu_{1}, \mu_{2}, \dots, \mu_{k}\}\) denotes the set of cluster centroids, and \(C = \{C_{1}, C_{2}, \dots, C_{k}\}\) denotes the cluster assignments. The algorithm proceeds as follows:

• Initialization: Randomly select \(k\) data points as initial centroids \(\{\mu_{1}, \mu_{2}, \dots, \mu_{k}\}\).

• Assignment Step: For each data point \(\chi_{i}\), compute its distance to all centroids \(\mu_{j}\) and assign it to the nearest cluster:
  \begin{equation} \mathrm{cluster{\_}assignment} \bigl(\chi_{i}\bigr) = \mathop{\rm arg~min}\limits_{j \in \{1, 2, \dots, K\}} \left\lVert \chi_{i} - \mu_{j} \right\rVert^{2}. \label{eq:cluster_assignment} \end{equation}
• Update Step: Recalculate the centroid of each cluster as the mean of its assigned data points:
  \begin{equation} \mu_{j} = \frac{1}{\vert C_{j} \vert} \sum_{\chi_{i} \in C_{j}} \chi_{i}, \label{eq:center_update} \end{equation}
  where \(C_{j}\) represents the set of data points in cluster \(j\), and \(|C_{j}|\) denotes the size of \(C_{j}\).

• Termination: Repeat assignment and update until the centroids stabilize (change \(<\) threshold) or the maximum iteration count is reached.

This process effectively extracts product characteristic information from large-scale customer reviews and uncovers the specific requirements of users for these characteristics, enhancing review data interpretability and providing valuable references for product design and market optimization.

3.2. CRs Extraction

The CRs are classified in fine-grained detail using the proposed BERT–BiLSTM–HPANet model (architecture in Fig. 3). The model preprocesses the requirement texts into embedded vectors, feeds them into BERT to generate context-rich word representations, and uses BiLSTM to capture the sequential features bidirectionally. Subsequently, it employs two key attention mechanisms.

• Hierarchical cross-head gating (HCG) attention dynamically recalibrates BiLSTM outputs to focus on critical semantic information.

• Position-aware attention integrates absolute and relative positional encodings to identify key semantic positions.

A detailed comparison of these mechanisms is presented in Table 1. Finally, average pooling and a multilayer perceptron are used to produce the classification logits. The proposed model combines BERT, BiLSTM, and dual attention to address the limitations in semantic capture and position sensitivity. The enables the accurate extraction of CRs from online reviews to optimize the product design.

3.2.2. Position-Aware Attention

Position-aware attention is a deep neural network module designed to enhance the explicit representation and interaction of positional information in sequence modeling. By leveraging a dual-modal position awareness mechanism that integrates absolute positional encoding and relative positional bias, combined with a dynamic multi-head attention paradigm, this module enables the cooperative modeling of local contextual relationships and global positional dependencies within text sequences. Its core lies in constructing a position-sensitive attention weight distribution, allowing the model to simultaneously capture semantic associations and spatial structural features simultaneously.

In terms of implementation, the module first performs position enhancement on the input sequence using learnable absolute positional encoding parameters. It injects independent position embedding representations for each sequence position, thereby establishing a benchmark spatial index. On this basis, a linear projection is employed to generate multi-head query, key, and value vectors. The multi-head attention mechanism is used to decompose the hidden space into multiple semantic subspaces, thereby facilitating the learning of diverse features. To strengthen the position awareness capability, a relative positional bias parameter matrix is introduced, that acts directly on the attention calculation process to explicitly model the relative distance association between any two positions. This bias matrix adaptively adjusts the interaction intensity between positions through parameterized learning, enabling the semantic associations of adjacent positions to receive stronger weight incentives. A scaling factor is introduced to normalize the dot product attention results, effectively controlling the stability of the attention weight distribution and avoiding gradient anomalies. When handling variable-length inputs, the module incorporates an attention masking mechanism to dynamically shield the interference from invalid positions, ensuring that the model focuses on the computation of effective semantic regions.

The calculation method can be expressed as follows:

\begin{equation} \mathrm{Attention}(Q, K, V) = \mathrm{softmax}\left(\frac{QK^{T}}{\sqrt{d_{k}}} + B_{\mathrm{rel}}\right) V, \label{eq:attention} \end{equation}

where \(B_{\mathrm{rel}}\) denotes the multi-head relative positional bias tensor. Serving as a key feature interaction layer in the BERT–BiLSTM hybrid architecture, this module significantly enhances the ability of the model to jointly represent complex semantic patterns and position-sensitive features through hierarchical position-enhanced attention operations.

3.3. CRs Analysis

4.1. Experiment Data

4.1.2. CRs Dataset Construction

To enable a systematic requirements analysis for laptop design, we constructed a structured dataset from UGC. The dataset was built using the reviews of Dell, HP, and ASUS laptops scraped from JD.com, a leading Chinese e-commerce platform. These reviews contained multi-dimensional feedback on product characteristics, forming the basis for requirement extraction and model training. To ensure dataset quality, a four-stage filtering process was implemented, as presented in Table 2.

Following preprocessing (encompassing the aforementioned filtering procedures), valid sentences were manually annotated with requirement categories. Specifically, nine primary labels were defined for laptop design requirements that corresponded to the product attributes that users cared the most about. The final dataset comprised 3,504 valid user requirement data points, with structured themes directly mapped to the outcomes of attribute extraction, laying a solid foundation for subsequent requirement analysis and product optimization endeavors.

4.2. Loss Function and Evaluation Metric

We adopted a cross-entropy loss function to quantify the discrepancy between the predicted probabilities and true labels. For our classification task, it is formulated as follows:

Accuracy (Acc), precision (P), recall (R), and F1-score were adopted as evaluation metrics to quantify the classification performance of the model. Their mathematical formulations are as follows.

Accuracy measures the overall proportion of correctly classified samples against the total number of samples. Its calculation is defined by Eq. \(\eqref{eq:accuracy}\):

Precision quantifies the ratio of truly positive samples to all samples predicted as positive by the model. This is formulated as shown in Eq. \(\eqref{eq:precision}\):

Recall (or sensitivity) reflects the proportion of truly positive samples that are correctly identified from all actual positive samples. Its formula is given by Eq. \(\eqref{eq:recall}\).

The F1-score is the harmonic mean of precision and recall, balancing the precision of the model in positive prediction and its ability to retrieve actual positives. It is calculated using Eq. \(\eqref{eq:f1}\).

4.3. Experimental

The text classification results across three datasets (CNews, ChnSentiCorp, and CRs), as are presented in Table 4, showed that BERT–BiLSTM–HPANet outperformed the baselines (for example, TextCNN and FastTEXT) and advanced pre-trained models (MacBERT, ChineseBERT, and RoBERTa). It achieved superior accuracy and F1-scores, demonstrating strong semantic capture and adaptability to domain-specific texts such as CRs.

For parameter efficiency, the model balanced the performance and computational cost, as shown in Fig. 5. Its parameter count was competitive with mainstream architectures (BERT, ERNIE, and RoBERTa), avoiding excessive parameters that inflated training/inference costs. This allowed deployment in resource-constrained environments (for example, edge devices) without sacrificing the quality.

In summary, the BERT–BiLSTM–HPANet model leveraged the contextual embeddings of BERT and the sequential modeling of BiLSTM to outperform single-architecture models. It achieved state-of-the-art accuracy with leaner parameters than heavyweight models (for example, RoBERTa), reducing the computational overhead. These results validated its suitability for industrial applications requiring high-precision text classification and resource efficiency.

4.4. Ablation Experiments

To determine the individual contributions of the core components, we conducted ablation experiments using the ChnSentiCorp and CRs datasets. In particular, this section examines the influence of the three key modules (BiLSTM, HCG attention, and position-aware attention) on the performance of the model, shedding light on their distinct roles in the full BERT–BiLSTM–HPANet architecture.

As presented in Table 5, removing the BiLSTM layer (w/o BiLSTM) caused consistent performance degradation across both datasets. This outcome underscored the pivotal function of BiLSTM in capturing sequential dependencies that is essential for enriching semantic representations in text classification tasks. Similarly, omitting HCG attention (w/o HCG Attn) or position-aware attention (w/o Position-Aware Attn) led to declines in accuracy, precision, recall, and F1-score. These results highlighted that HCG attention refined multi-head attention weights to prioritize critical semantic information, whereas position-aware attention bolstered the utilization of positional features, both of which were indispensable for effectively modeling complex contextual and structural nuances within texts.

The full model outperformed all ablated configurations, validating the synergistic design: the sequential modeling prowess of BiLSTM, semantic calibration ability of HCG attention, and positional enhancement capacity of position-aware attention collectively optimized the classification performance. This ablation study clarified the necessity of each component and underscored the robustness of the model architecture for text-intensive applications such as CR analysis.

5.1. Data Collection and Processing

The Mechrevo Aurora X laptop was selected for analysis based on the customer reviews collected from JD.com, a leading e-commerce platform for electronic device purchases in China. Using established web crawling technologies, we systematically retrieved 1,700 customer reviews (including positive, neutral, and negative evaluations) spanning September to December 2024. A sample of the data is shown in Fig. 6.

Let \(\mathrm{Review} = \{\mathrm{CR}_{1}, \mathrm{CR}_{2}, \dots, \mathrm{CR}_{N}\}\), where \(N\) represents the number of CRs per review. We used the “re” module in Python to split long customer texts into sentences using special symbols such as periods and exclamation marks. This process separated different product-related requirements within the CR texts, laying the foundation for subsequent requirement extraction and impact analysis. Through this process, we processed 1,700 customer reviews to yield 3,044 valid requirement instances.

5.2. Application of REAM

After data preprocessing, we employed the trained BERT–BiLSTM–HPANet requirement classification model to systematically categorize customer reviews into eight specific product design requirement categories and one miscellaneous category. Table 6 comprehensively presents the representative examples corresponding to eight specific requirement categories.

The pre-trained RoBERTa model captures rich linguistic features and contextual information, enabling it to effectively identify sentiment orientations in complex texts. The statistical results are shown in Fig. 7.

To analyze the influence of CRs influence on each other, we first extracted and structured CRs from reviews. We then computed the Jaccard similarity for all CR pairs—that measured the semantic overlap (ratio of intersection to union of CR descriptions). For our identified CRs (\(\mathrm{CR}_{1}\) to \(\mathrm{CR}_{8}\)), these similarities form the heatmap, as shown in Fig. 8. Warmer colors (for example, red) indicated stronger semantic links, revealing the CRs that customers often associated together. This helped prioritize fulfilling interrelated requirements in the product design, ensuring that we aligned with overlapping CRs.

5.3. Results

Drawing on insights from the sentiment distribution (Fig. 7), requirement interaction patterns (Fig. 8), and classification statistics (Table 6), targeted product enhancements for Mechrevo Aurora X were proposed to address critical pain points and amplify competitive strengths. Focusing on high-impact categories, as presented in Table 6, where “Processor (CR2)” and “Display (CR4)” dominate customer feedback (503 and 313 mentions, respectively), the 12th gen i7 + RTX 4070 configuration in CR2, that garners positive sentiment, should be paired with refined thermal designs (for example, upgraded heatsinks or liquid metal cooling for power users) to sustain high performance. For CR4, screen defects such as dead pixels and backlight leakage can be mitigated through 100% automated optical inspection, complemented by “zero-defect” display warranties to reinforce trust in visual quality.

Critical pain points in “Internal Storage (CR3, 220 instances)” and “Software System (CR7, 215 instances)” require focused solutions. Underperforming SSDs in CR3 should be phased out, with high-speed PCIe 4.0/5.0 NVMe drives (\(>\)7,000/6,000 MB/s) sourced through partnerships, accompanied by firmware optimizations (for example, DRAM-cache integration) to reduce latency. For CR7, a dedicated “Gaming Mode” can disable conflicting background processes (for example, Windows Update) and lock stable GPU drivers, with top-cited issues (for example, blue screens) resolved within 48 hours via user-feedback loops.

Leveraging the sentiment trends, as shown in Fig. 7, strong positives in CR2 and CR5 (for example, metal chassis) can be amplified through custom finishes (for example, anodized aluminum) or RGB accents. For mixed/negative signals in CR1 and CR3, the left-side USB port should be reengineered (for example, rear placement) to fix “tight fit” issues, whereas discounted SSD upgrade kits can turn storage complaints into loyalty-building opportunities.

Aligning with requirement dependencies in the Jaccard heatmap, as shown in Fig. 8, high-correlation pairs such as CR2 \(\leftrightarrow\) CR4 and CR2 \(\leftrightarrow\) CR5 can inform “performance bundles” for example, RGB-lit vents (CR5) optimized for GPU thermals (CR2) or high-refresh displays (CR4) paired with CPU overclocking presets (CR2). Latent links such as CR1 \(\leftrightarrow\) CR8 (USB port and heat dissipation) can be addressed by relocating ports to avoid blocking air intakes, reducing thermal throttling while resolving usability issues. These data-driven adjustments, as shown in Table 6, Fig. 7, and Fig. 8, will enable Mechrevo Aurora X to resolve pain points, amplify strengths, and deliver a cohesive customer experience aligned with interconnected CR, thereby solidifying its market position.