💻 computer science

A Framework for Customer Needs Analysis: Integrating Dynamic Sentiment Lexicon Construction with Kano Model

This paper proposes a data-driven framework that integrates guided BERTopic, a dynamic domain-adapted sentiment lexicon, and the Kano model to objectively analyze online reviews and prioritize product improvements, demonstrated through an evaluation of 8,970 new energy vehicle reviews.

Original authors: Qiong He, Yijia Li, Zhenwei Yang, Shuai Wang, Jian Zhang

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Qiong He, Yijia Li, Zhenwei Yang, Shuai Wang, Jian Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve the mystery of what people really want from a product. In the world of business, this is called "customer needs analysis." For a long time, companies tried to solve this by handing out paper surveys or asking people to fill out forms. But that's like trying to guess what a whole city is thinking by asking just five people on a street corner—it's slow, often biased, and people might not even tell the truth.

Enter the internet age, where millions of people leave behind a digital trail of opinions in the form of online reviews. This is a goldmine of data, but it's messy. To make sense of it, scientists use two main tools. First, they use "sentiment analysis," which is like a mood ring for text; it tries to figure out if a word or sentence is happy, sad, angry, or excited. Second, they use the "Kano model," a famous framework that sorts needs into categories: some are basic (you must have them, or you're unhappy), some are linear (the more you have, the happier you are), and some are "delighters" (unexpected bonuses that make you jump for joy). The problem? Traditional mood rings are "static"—they use a fixed dictionary of words that can't handle slang, sarcasm, or context. And the Kano model usually relies on those slow, subjective surveys.

This paper proposes a clever new way to combine these tools. The researchers built a "dynamic" mood ring that learns from the specific language of car owners, and they used it to feed the Kano model automatically, skipping the surveys entirely. They tested this on nearly 9,000 reviews of new energy vehicles (like electric cars) to see if they could figure out exactly what drivers love, what they hate, and what features companies should fix first.

The Detective's New Toolkit

The researchers, Qiong He and her team from Beijing Information Science & Technology University, realized that old tools were failing to keep up with how people actually talk online. If you say a car is "big," a static dictionary might just see a neutral word. But if you say "big noise," that's bad. If you say "big trunk," that's good. A standard dictionary can't tell the difference without a human to explain the context.

To fix this, the team built a Dynamic Sentiment Lexicon. Think of this as a super-smart, self-updating dictionary that doesn't just memorize words; it understands the vibe of the conversation. They used a powerful AI brain (called BERT) that reads sentences the way humans do, looking at the words before and after a term to understand its true meaning. They also used a math trick called TF-IDF-PageRank (a fancy way of saying "find the most important words that appear together") to pick the best starting words for their dictionary.

Once they had this smart dictionary, they didn't stop there. They plugged it into the Kano Model. Usually, the Kano model requires a researcher to ask a customer, "If this feature existed, how would you feel?" and "If it didn't exist, how would you feel?" The team skipped the asking part. Instead, they let their AI read thousands of reviews, calculate how happy or angry people were about specific car parts, and automatically sort those parts into the Kano categories.

The Case of the 9,000 Reviews

The team put their framework to the test using 8,970 online reviews of new energy vehicles (NEVs) collected from a popular Chinese car website called Dongchedi. They focused on five major brands: Tesla, BYD, Chang'an, NIO, and Lixiang.

After cleaning up the data (removing duplicates and gibberish), their system successfully identified nine key product attributes that customers care about. These ranged from "Driving range" (how far the car goes on a charge) and "Space" to "Intelligence" (smart driving features) and "Chassis system" (the car's skeleton and ride quality).

Here is where the magic happened. The AI didn't just list the features; it classified them using the Kano model, revealing the hidden psychology of the drivers:

The "Must-Haves" (Must-be Attributes): The system found that Seats, the Chassis system, and the Driving experience are the basics. If these are bad, customers are furious. But if they are just "good," customers don't get excited; they just expect it. It's like having a roof on your house: if it leaks, you're angry, but if it's solid, you don't throw a party.
The "Linear" Features (One-dimensional Attributes): Appearance falls here. The more stylish the car looks, the happier the customer. It's a straight line: better looks equal more smiles.
The "Delighters" (Attractive Attributes): This was the big surprise. Features like Driving range, Space, Intelligence, Interior quality, and Price were classified as "delighters." This means that if a car excels in these areas, it creates a massive spike in customer satisfaction. However, if the car is just "okay" in these areas, customers aren't necessarily angry; they just don't get that extra spark of joy.

The "Opportunity" Score: Who to Fix First?

Knowing what people want is great, but companies have limited budgets. They can't fix everything at once. So, the researchers added a final step: an Opportunity Algorithm. This acts like a priority list. It looks at two things:

Importance: How often do people talk about this feature?
Satisfaction: How happy are people with it right now?

The algorithm calculates a "chance score" to tell manufacturers where to focus. The results were clear:

Top Priority: Space and Driving range were the winners. They are already "delighters" (people love them when they are good), but there is still room to make them even better to win more customers.
Urgent Fix: Driving experience was a "Must-have" that customers were surprisingly unhappy with. Even though it's a basic requirement, the low satisfaction score meant it was a ticking time bomb for the brands.
Low Priority: Intelligence and Appearance had lower opportunity scores, suggesting that while they are important, the current gap between what customers want and what they have isn't as critical as the driving experience or space issues.

The Verdict

The study suggests that by using this dynamic, AI-driven approach, companies can move away from guessing and subjective surveys. Instead, they can listen to the "voice of the customer" directly from the chaos of online reviews. The framework successfully turned 8,970 messy reviews into a clear, actionable roadmap: fix the driving experience to stop the bleeding, and supercharge the space and battery range to create the next big hit.

While the authors note that their method is currently tested only on electric cars and relies on a single website's data, they argue that this "dynamic lexicon" approach offers a repeatable, objective way to understand what customers truly need, turning the noise of the internet into a clear signal for product improvement.

Technical Summary: A Framework for Customer Needs Analysis Integrating Dynamic Sentiment Lexicon Construction with Kano Model

Problem Statement
Accurate customer needs analysis is fundamental to effective product design and iterative optimization. While online reviews offer a rich, low-cost source of authentic user feedback, traditional analytical methods face two significant limitations:

Static Sentiment Lexicons: Conventional sentiment dictionaries struggle to capture context-dependent and domain-specific emotional expressions found in User-Generated Content (UGC). They often fail to adapt to the rapid evolution of language and the nuanced semantic shifts of words in specific industries.
Subjectivity in the Kano Model: The traditional application of the Kano model relies heavily on manual questionnaires and surveys. This approach introduces high subjectivity, suffers from small sample bias, and lacks scalability, limiting its objectivity in identifying complex customer psychology.

Methodology
The authors propose an integrated, data-driven framework that automates attribute extraction, sentiment quantification, needs classification, and priority ranking. The process consists of four sequential stages:

Data Collection and Preprocessing:
- Reviews are crawled from e-commerce platforms (specifically Dongchedi for New Energy Vehicles).
- Texts undergo preprocessing, including the removal of non-Chinese characters, punctuation, and stop words, followed by segmentation and part-of-speech tagging using the Jieba toolkit.
Product Attribute Extraction (Guided BERTopic):
- The study employs Guided BERTopic, a topic modeling approach that integrates domain-specific knowledge with data-driven clustering.
- Seed Word Guidance: Domain experts provide seed words to guide the topic generation, ensuring relevance and reducing noise.
- Modeling Steps:
  - Document Embedding: Text is mapped to high-dimensional vectors using paraphrase-multilingual-MiniLM-L12-v2.
  - Seed Guidance: Cosine similarity between document embeddings and the centroid of seed word embeddings enhances the relevance of specific topics.
  - Dimensionality Reduction: UMAP reduces dimensionality while preserving local and global data structures.
  - Clustering: HDBSCAN groups documents into clusters of arbitrary shapes, isolating noise.
  - Keyword Extraction: An improved c-TF-IDF algorithm, enhanced with a weight factor for seed words, generates topic keyword distributions.
Dynamic Sentiment Lexicon Construction:
- To overcome the rigidity of static dictionaries, the authors construct a domain-adapted lexicon using a hybrid of statistical and deep learning methods:
  - Seed Word Selection: A TF-IDF-PageRank heterogeneous feature fusion model identifies high-discriminative sentiment seed words. TF-IDF captures local statistical significance, while PageRank captures global structural correlations. The final score is a linear combination of both ( $S(w_i) = \alpha \cdot TF\text{-}IDF + (1-\alpha) \cdot RP$ ).
  - Contextual Representation: The BERT model is used to generate dynamic, context-aware word vectors, overcoming the static nature of Word2Vec or GloVe.
  - Dynamic Expansion: Candidate words are expanded into the lexicon based on cosine similarity (threshold 0.75) with seed vectors, allowing the dictionary to adapt to new emotional expressions.
- Contextual Polarity: The framework accounts for context sensitivity (e.g., "large" is positive for space but negative for noise) by mapping sentiment words to specific attribute tuples.
Customer Needs Classification and Prioritization:
- Kano Model Integration: Instead of surveys, the model uses regression analysis on review data. A formula ( $y = \alpha + \sum(\beta_j^{pos}X_j^{pos} + \beta_j^{neg}X_j^{neg})$ ) maps the contribution of positive and negative sentiments to customer satisfaction ratings.
- Decision Tree: Attributes are classified into five categories (Attractive, One-dimensional, Must-be, Reverse, Indifferent) based on estimated coefficients ( $\beta$ ) and thresholds ( $\lambda, \delta$ ).
- Opportunity Algorithm: To rank improvement priorities, an "Opportunity Score" is calculated combining Importance (frequency of attribute mentions) and Satisfaction (aggregated sentiment scores).
- Formula: $Opportunity = Importance + \max(Importance - Satisfaction, 0)$ .

Key Results
The framework was validated using 8,970 online reviews of New Energy Vehicles (NEVs) from five major brands (Tesla, BYD, Chang'an, NIO, Lixiang).

Attribute Extraction: Nine core product attributes were identified: Driving Range, Space, Appearance, Intelligence, Interior, Price, Seats, Chassis System, and Driving Experience.
Lexicon Construction: A dynamic domain sentiment lexicon was built containing 736 words (512 positive, 224 negative), successfully capturing domain-specific nuances.
Kano Classification:
- Attractive Attributes: Driving range, Space, Intelligence, Interior, and Price. Improvements here significantly boost satisfaction.
- One-dimensional Attributes: Appearance. Satisfaction is linearly correlated with performance.
- Must-be Attributes: Seats, Chassis System, and Driving Experience. These are basic expectations; poor performance causes dissatisfaction, but excellent performance yields diminishing returns.
Prioritization: The Opportunity algorithm ranked Space, Driving Range, and Driving Experience as the top three priorities for improvement. Notably, while "Driving Experience" is a basic need, it currently has low satisfaction scores, indicating a critical area for optimization.

Significance and Claims
The paper claims to provide a repeatable and transferable solution for intelligent product optimization by addressing the limitations of traditional methods:

Objectivity: By replacing subjective questionnaires with data-driven analysis of online reviews, the framework eliminates sample bias and human subjectivity.
Adaptability: The dynamic sentiment lexicon construction solves the problem of static dictionaries failing to capture context-sensitive and evolving emotional expressions in UGC.
Actionability: The integration of the Kano model with an opportunity algorithm transforms raw user feedback into quantifiable, ranked improvement priorities.

The authors conclude that this data-driven approach enhances the efficiency and accuracy of product development, enabling manufacturers to adapt swiftly to market dynamics. They acknowledge limitations, including the current application to a single domain (NEVs), the lack of specific fake review detection, and the use of single-platform data, suggesting these as areas for future research.

The Detective's New Toolkit

The Case of the 9,000 Reviews

The "Opportunity" Score: Who to Fix First?

The Verdict

Technical Summary: A Framework for Customer Needs Analysis Integrating Dynamic Sentiment Lexicon Construction with Kano Model

More like this