A method for mining truck users' implicit needs based on natural language processing technology

Abstract

The present invention discloses a method for mining the implicit needs of truck users based on natural language processing technology. The method mainly includes: obtaining documents and materials about trucks, HOWNET sentiment dictionary and social media comment texts about trucks; performing validity analysis on the original comment texts to obtain valid comment texts; performing corpus preprocessing on the truck documents and materials, HOWNET sentiment dictionary and valid comment texts to construct an attribute vocabulary library and a sentiment vocabulary library; and manually annotating the valid comment texts. The present invention relates to the technical field of user demand mining, and specifically, to a method for mining the implicit needs of truck users based on natural language processing technology. The technical problem to be solved by the present invention is to provide a method for mining the implicit needs of truck users based on natural language processing technology, which facilitates the mining of the implicit needs of truck users.

Description

Technical Field

[0001] The present invention relates to the technical field of user demand mining, and in particular to a method for mining the implicit demands of truck users based on natural language processing technology. Background Art

[0002] In the process of product improvement and design, timely understanding of user needs is crucial. Traditional methods, mainly using questionnaires, user interviews, and eye-tracking experiments, have the disadvantages of time delays, long cycles, and high costs. Thanks to the rapid development of the e-commerce industry and social networks, online reviews have become a vehicle for expressing user needs and an effective way to identify product defects. By using data mining tools to extract product features that users are interested in from massive amounts of online reviews, and using natural language processing technology to mine and analyze text, we can obtain potential customers' opinions, textual tendencies, and emotional states. Online reviews have become an important basis for users to choose and use products, and reviews implicitly reflect users' potential needs for products. However, most research on potential user needs analysis based on online reviews is still in its infancy. How to more comprehensively and accurately obtain users' explicit and implicit needs remains an important issue that requires in-depth research. An efficient and accurate product demand mining method is urgently needed. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for mining the implicit needs of truck users based on natural language processing technology, which facilitates the mining of the implicit needs of truck users.

[0004] The present invention adopts the following technical solutions to achieve the invention objectives:

[0005] A method for mining hidden needs of truck users based on natural language processing technology is characterized by comprising the following steps:

[0006] Step 1: Based on the Windows 10 64-bit operating system and the Ubuntu 18.0.4 operating system, use PyCharm and Python 3.6 to build an experimental platform and create folders such as "Truck Information", "Sentiment Dictionary", "Online Reviews", and "Effective Review Text";

[0007] Step 2: Data collection: Collect truck literature and information, sentiment dictionary and comment text, and store them in corresponding folders;

[0008] Step 3: Analyze the validity of the comment text: separate invalid comments and save valid comments in the "Valid Comment Text" folder;

[0009] Step 4: Corpus preprocessing: Use the NLPIR-ICICLAS Chinese lexical analysis system to preprocess the data in the two folders "Truck Information" and "Effective Comment Text";

[0010] Step 5: Lexicon Construction: This is divided into attribute lexicon and sentiment lexicon. The contents in the "Truck Information" folder are segmented by the NLPIR-ICICLAS Chinese lexical analysis system to obtain nouns or noun phrases as the initial attribute lexicon; the contents in the "Sentiment Dictionary" folder are segmented by the NLPIR-ICICLAS Chinese lexical analysis system to obtain nouns or noun phrases as candidate attribute lexicon, and adjectives or adjective phrases as candidate sentiment lexicon. Finally, filter and remove duplicates to obtain the final attribute lexicon and final sentiment lexicon.

[0011] Step 6: Classify the text of valid comments, manually annotate them and then use the text classification model to classify them;

[0012] Step 7: Extracting attribute words and sentiment words: Specify the criteria for distinguishing explicit and implicit sentence patterns. If the comment text contains both attribute words and sentiment words, it is considered an explicit sentence pattern; if the comment text only contains attribute words, it is considered an implicit sentence pattern.

[0013] Step 8: Sentiment quantification: Use the SO-PMI algorithm and the HOWNET sentiment polarity quantification standard to analyze and quantify the above results, and obtain the <attribute vocabulary, sentiment mean> set;

[0014] Step 9: Demand Sorting: Build a truck demand sorting model by combining the KANO model and the DEMATEL analysis method. Use the KANO model to find the relationship between demand and user satisfaction and classify demand attributes. Use the DEMATEL analysis method to quantify the impact relationship between demands and ultimately provide a clear sorting strategy.

[0015] Step 10: Use the questionnaire survey method to conduct an online survey on the satisfaction of some truck users with various design elements of truck products, and compare and analyze the results with the above demand ranking.

[0016] As a further limitation of this technical solution, the data collection in step 2 is specifically as follows:

[0017] Use the NLPIR-ICICLAS Chinese word segmentation system and the Octopus web crawler tool to obtain literature and information about trucks, such as truck market reports, truck news, truck advertisements, etc., and save them to "Truck Information"

[0018] Under the folder;

[0019] Use literature search method to obtain multiple dictionaries and save them in the folder of "Emotional Dictionary";

[0020] The text of social media comments about trucks was obtained using manual retrieval and the Octopus web crawler tool and saved in the "Online Reviews" folder.

[0021] As a further limitation of this technical solution, in step three:

[0022] Definition of a valid comment:

[0023] (1) Valid reviews must be related to actual product information and not include irrelevant noisy information. In addition, reviews should also include information such as user experience, not just simple product introduction information.

[0024] (2) Effective reviews should include more product feature dimensions and be more detailed;

[0025] (3) In addition, effective comments are universal in sentence structure and content, not sentences with some rare words and special structures;

[0026] Definition of invalid comments:

[0027] (1) Invalid reviews are characterized by a low relevance to truck products and do not include detailed descriptions of specific design elements of truck products, such as reviews made only on e-commerce platforms;

[0028] (2) Some reviews are obviously inconsistent with the actual product features and are also defined as invalid reviews;

[0029] Clean up invalid comments to reduce the interference of noise data on experimental results:

[0030] (1) Delete spam comments. Comments that only contain a few words, a string of characters, or are entirely punctuation marks have no information value and need to be deleted.

[0031] (2) Deletion of duplicate comments. Duplicate comments increase the workload and affect the results of sentiment analysis and user attention, so they need to be deleted.

[0032] (3) Special symbol processing, including emoticons, special symbols or garbled characters, needs to be deleted.

[0033] As a further limitation of this technical solution, the specific process of step 4 is as follows:

[0034] Step 41: Data cleaning: Find the content of interest in the corpus and delete the content that is not interesting or considered as noise;

[0035] Step 42: Word segmentation: Process all text data into the smallest unit of granularity - words;

[0036] Step 43: Part-of-speech tagging: tag each word with a part-of-speech tag, such as adjective, verb, noun, etc.

[0037] Step 44: Remove stop words. Stop words generally refer to words that do not contribute to the text features, such as punctuation marks, tone, personal names, etc.

[0038] As a further limitation of this technical solution, the specific process of step six is as follows:

[0039] Step 61: To prevent data overfitting during text classification, we first classify truck product design elements. This involves classifying valid review texts based on the ten first-level functional attributes in the Truck Platform Product Feature Catalog.

[0040] Step 62: Manually label the valid comment texts, that is, manually determine which category the valid comment texts belong to among the ten categories of first-level functional attribute labels sentence by sentence, mark "1" in the label box of the corresponding category, and mark "0" in the label boxes of the remaining categories. This dataset is used as the training set;

[0041] Step 63: Use machine learning models and deep learning models to learn the training set, and then perform attribute classification on valid review texts that have not been manually annotated;

[0042] Step 64: Use the K-fold cross-validation method to select the text classification results with the highest accuracy for subsequent research.

[0043] As a further limitation of this technical solution, the specific process of step eight is as follows:

[0044] Sentiment polarity analysis uses the SO-PMI algorithm, also known as the sentiment tendency point mutual information algorithm. The overall idea of the algorithm is very simple: it determines the probability of the word P (word) to be judged and the base word P (base) appearing at the same time. If the probability of appearing with positive words is higher, then the word is judged as positive. If the probability of appearing with negative words is higher, then the word is judged as negative. If the probability of appearing with positive and negative words is the same, then the word is judged as neutral.

[0045] The algorithm formula is as follows:

[0046] (1)

[0047] Where: num(pos) refers to the total number of positive benchmark words;

[0048] Similarly, num(neg) refers to the total number of negative benchmark words;

[0049] POS i refers to positive benchmark words;

[0050] neg i Refers to negative benchmark words;

[0051] PMI(word, pos i ) refers to the point-wise mutual information between a word and a positive benchmark word;

[0052] PMI(word, neg i ) refers to the point-wise mutual information between a word and a negative benchmark word;

[0053] The formula produces the following results:

[0054] SO-PMI>0, the words are judged as positive words;

[0055] SO-PMI=0, the word is judged as a neutral word;

[0056] If SO-PMI<0, the word is judged as a negative word.

[0057] After the sentiment polarity analysis is completed, the sentiment mean is calculated according to the HOWNET sentiment polarity quantification standard table.

[0058] As a further limitation of this technical solution, in step nine, the KANO model is used to classify the attributes and identify the priorities of the requirements. The implementation steps are as follows:

[0059] Step 9-11: Questionnaire preparation: Design a questionnaire based on the set of <attribute vocabulary, satisfaction, and user attention> to fully understand user feedback on truck needs;

[0060] Step 912: Data processing. To more intuitively prioritize the needs, the Better-Worse satisfaction coefficient calculation formula is used to calculate the satisfaction coefficient and dissatisfaction coefficient of each need. When calculating, indifferent needs and reverse needs are removed. The specific calculation formulas are shown in Formula (2) and Formula (3).

[0061] The product meets this requirement, and the Better coefficient is:

[0062] Better / SI = (A+O) / (A+O+M+I) (2)

[0063] The product does not meet this requirement, and the Worse coefficient is:

[0064] Worse / DSI =-(O+I) / (A+O+M+I)(3)

[0065] Where: A——Number of charm requirement options;

[0066] O——the expected number of demand options;

[0067] M - the number of required options;

[0068] I – number of indifferent demand options;

[0069] Based on the Better and Worse coefficient values, a demand quartile chart is drawn. The sorting principle is essential demand M > desired demand O > attractive demand A > indifferent demand I. When multiple demand items belong to the same attribute, the internal sorting needs to be calculated by importance. The weight calculation formula for the i-th demand item is:

[0070] (4)

[0071] Where: w i ——The initial weight of the demand for product i;

[0072] k i ——The adjustment coefficients of essential attributes, desired attributes, attractive attributes, and indifference attributes are 1, 2, 4, and 0 respectively.

[0073] According to w i With k i The value can be used to calculate the weighted weight w' of each demand item i , thereby determining the order of demand items, and combining the classification principles to determine the priority order of final product demand.

[0074] As a further limitation of this technical solution, in step nine, the DEMATEL analysis method is combined with an expert group to further explore the impact relationship between requirements using expert opinions, and transform it into a more objective product demand ranking method. The brief calculation steps are as follows:

[0075] Step 921: By scoring the impact of each demand pairwise, the causal relationship between the demands is quantified to obtain the direct impact matrix A, where a ij Express the degree of influence of demand i on demand j;

[0076] Step 922: Obtain the normalized direct influence matrix N through equations (5) and (6):

[0077] (5)

[0078] (6)

[0079] Step 923: Solve the comprehensive influence matrix T:

[0080] T=N(IN)-1 (7)

[0081] Where: I——unit matrix;

[0082] Step 924: setting the threshold and outputting the coordinate information of the causal relationship graph;

[0083] Step 925: Divide the impact and cause-effect diagram into four quadrants and determine the demand categories.

[0084] Compared with the prior art, the advantages and positive effects of the present invention are:

[0085] This invention uses online truck user reviews as a data source, employing natural language processing to conduct demand discovery research. It then utilizes text classification to make implicit needs explicit, prioritizing them based on user interest and satisfaction. Compared to psychology-based modeling, natural language processing can quickly and accurately capture user needs, improving accuracy while also saving companies money. It can also provide real-time insights and suggestions for product design or improvement, making it particularly practical in the era of big data.

[0086] 2. This method proposes a new method for mining user implicit needs. Based on current research, this method uses multiple text classification models and utilizes the K-fold cross-validation method to select the result with the highest accuracy for mining user implicit needs, thereby improving the accuracy.

[0087] 3. Considering that some product attributes of trucks are affected by environmental factors to varying degrees, which in turn affects users’ satisfaction with these product attributes, such as appearance, tires, and brake systems, the former is not affected by natural conditions, while the latter two are affected to varying degrees. Therefore, an environmental factor influence coefficient is proposed to minimize the impact of environmental factors on users’ subjective evaluation of truck attributes. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is the main technical route of the present invention.

[0089] Figure 2 This is a schematic diagram of the process of crawling social media comment text using the web crawler technology of the present invention.

[0090] Figure 3 This is a schematic diagram of the code for crawling test drive reviews on e-commerce websites using the web crawler technology of the present invention.

[0091] Figure 4 This is a schematic diagram of corpus preprocessing using the NLPIR-ICICLAS Chinese lexical analysis system of the present invention.

[0092] Figure 5 The diagram is a flowchart of the vocabulary building process of the present invention.

[0093] Figure 6 This is a schematic diagram of manually labeling valid comment texts according to the present invention.

[0094] Figure 7 This is a schematic diagram of reading manually annotated text corpus data according to the present invention.

[0095] Figure 8 Schematic diagram of text data cleaning according to the present invention.

[0096] Figure 9 This is a schematic diagram of word order construction of the present invention.

[0097] Figure 10 It is a schematic diagram for converting the text of the present invention into digital coding.

[0098] Figure 11 This is a schematic diagram of the detailed process of the three machine learning methods of the present invention: Naive Bayes, Random Forest and Support Vector Machine.

[0099] Figure 12 Schematic diagram of the operation process of the Billet model of the present invention.

[0100] Figure 13 Schematic diagram of the operation process of the Attention_BiLSTM model of the present invention.

[0101] Figure 14 Schematic diagram of the KANO model of the present invention.

[0102] Figure 15 This is the four-quadrant impact causal relationship diagram of the present invention.

[0103] Figure 16 This is the truck demand prioritization model of the present invention. DETAILED DESCRIPTION

[0104] A specific embodiment of the present invention is described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.

[0105] like Figures 1-16 As shown, the present invention includes the following steps:

[0106] Step 1: Using PyCharm and Python 3.6, we built an experimental platform based on Windows 10 64-bit and Ubuntu 18.0.4. We created folders called "Truck Information," "Sentiment Dictionary," "Online Reviews," and "Effective Review Text." We used the NLPIR-ICICLAS Chinese word segmentation system to conduct text mining and demand analysis experiments. We also used the Octopus web crawler to crawl online reviews and product community posts as the review corpus and experimental dataset.

[0107] Step 2: Data collection: Collect the literature and information, sentiment dictionary and comment text of the truck and store them in the corresponding folder.

[0108] Using the NLPIR-ICICLAS Chinese word segmentation system and the Octopus web crawler, we retrieved literature and information about trucks, such as a catalog of truck product characteristics, a truck market report from January 1, 2019, to December 31, 2020, truck news, and truck advertisements. These documents were saved in the "Truck Information" folder. We also used literature retrieval to retrieve various dictionaries, which were saved in the "Sentiment Dictionary" folder.

[0109] Manual retrieval and Octopus web crawler tools were used to obtain social media comment texts about trucks from January 1, 2019 to December 31, 2021, including test drive reviews on the "Truck Home" forum, e-commerce websites, and reviews on short video software, totaling 4,463 comment data, which were saved in the "Online Reviews" folder.

[0110] Step 3: Analyze the validity of the comment text: separate invalid comments and save valid comments in the "Valid Comment Text" folder.

[0111] The effectiveness of online reviews refers to whether the review recipient can obtain effective information from the review to assist in making decisions. For users, the main function of effective reviews is to eliminate uncertainty. From the perspective of product design or user demand mining research, it is expected to obtain users' detailed evaluation of product design elements from the review data.

[0112] Online reviews contain a lot of useless information. To ensure that the collected text is valuable for research, starting from the perspective of truck products and integrating the research results of other scholars on the factors affecting the effectiveness of reviews, we define effective reviews based on the degree of relevance of the review content to truck product characteristics, the specificity of the review content, and the readability of the review content:

[0113] (1) Valid reviews must be related to actual product information and not include irrelevant noisy information; in addition, reviews should also include information such as user experience, not just simple product introduction information.

[0114] (2) Effective reviews should include more product feature dimensions and be more detailed.

[0115] (3) In addition, effective comments are universal in sentence structure and content, and are not sentences with some obscure words and special structures.

[0116] The characteristics of invalid reviews are opposite to those of valid reviews. Invalid reviews cannot provide companies with real user experiences, and will increase the difficulty of user demand mining, and are likely to interfere with the results of demand mining.

[0117] Definition of invalid comments:

[0118] (1) Invalid reviews are characterized by a low degree of relevance to truck products and do not include detailed descriptions of specific design elements of truck products, such as reviews made only on e-commerce platforms.

[0119] (2) Some reviews are obviously inconsistent with the actual product features and are also defined as invalid reviews. Invalid reviews are cleaned up to reduce the interference of noise data on the experimental results.

[0120] (1) Delete spam comments. Comments that only contain a few words, a string of characters, or are entirely punctuation marks have no information value and need to be deleted.

[0121] (2) Deletion of duplicate comments. Duplicate comments increase the workload and affect the results of sentiment analysis and user attention, so they need to be deleted.

[0122] (3) Special symbol processing, including emoticons, special symbols or garbled characters, needs to be deleted.

[0123] The above criteria serve as the standard for the validity of manually annotated comment texts. After the validity analysis of the comment texts, 2713 valid comments were obtained and saved in the "Valid Comment Text" folder.

[0124] Step 4: Corpus preprocessing: Use the NLPIR-ICICLAS Chinese lexical analysis system to preprocess the data in the two folders "Truck Information" and "Effective Comment Text".

[0125] Step 4-1: Data cleaning: Find the content of interest in the corpus and remove the content that is of no interest or considered noise; this includes extracting the title, abstract, body and other information from the original text, and removing advertisements, tags, HTML, JS code and comments from the crawled web page content.

[0126] Step 42: Word segmentation: Chinese corpus data is generally short or long text. When conducting text mining analysis, all text data needs to be processed into the smallest unit granularity - words.

[0127] Step 43: Part-of-speech tagging: Label each word with a part of speech, such as adjective, verb, or noun. This allows the text to incorporate more useful linguistic information during subsequent processing. The ICTCLAS Chinese part-of-speech tagging set is used, as shown in Table 1.

[0128] Table 1 ICTCLAS Chinese part-of-speech tagging set

[0129]

[0130]

[0131] Step 44: Remove stop words. Stop words generally refer to words that do not contribute to the text features, such as punctuation marks, tone, personal names, etc.

[0132] Step 5: Lexicon Construction: This is divided into an attribute lexicon and a sentiment lexicon. The contents of the "Truck Information" folder, after being segmented by the NLPIR-ICICLAS Chinese lexical analysis system, yield nouns or noun phrases as the initial attribute lexicon; the contents of the "Sentiment Dictionary" folder, after being segmented by the NLPIR-ICICLAS Chinese lexical analysis system, yield nouns or noun phrases as candidate attribute lexicons, and adjectives or adjective phrases as candidate sentiment lexicons. Finally, the final attribute lexicon and sentiment lexicon are filtered and deduplicated to create the final lexicon.

[0133] Step 6: Perform text classification on valid comments, and use the text classification model to classify them after manual annotation.

[0134] Step 61: To prevent data overfitting during text classification, we first classify the truck product design elements, that is, classify the valid review texts according to the ten first-level functional attributes in the "Truck Platform Product Feature Catalog".

[0135] A focus group was established, consisting of 12 industrial design graduate students, 6 male and 6 female, aged between 24 and 28. To prevent data overfitting during text classification, the design elements of truck products were first classified.

[0136] The product feature catalog is a comprehensive language system used to accurately describe "what the product is," encompassing descriptions from broad indicators to specific product features. Typically, indicators derived from product definition research are refined to the third level, which then leads to detailed development targets and then to the fourth level. The first level catalog represents the customer domain: the primary, customer-related vehicle features that directly influence purchasing decisions. The second level catalog represents the functional domain: the secondary or functional vehicle features necessary to achieve the first level features. The third level catalog represents the physical domain: the physical prerequisites for achieving the functional domain features, with a large number of detailed target values. Levels four and above provide even more detailed indicator descriptions as needed.

[0137] The truck platform's product feature catalog currently contains 1,250 items. This catalog is universal and applicable to all vehicle models. The first-level feature catalog (customer domain) contains 10 items, the second-level feature catalog (functional domain) contains 54 items, and the third-level feature catalog (physical domain) contains 265 items. Some features are broken down to the fourth and fifth levels of technical indicators. Specific projects will select representative items from this catalog based on their specific characteristics and develop dedicated feature catalogs.

[0138] Step 62: Manually label the valid comment texts, that is, manually determine which category of the ten first-level functional attribute labels the valid comment texts belong to sentence by sentence, mark "1" in the label box of the corresponding category, and mark "0" in the label boxes of the remaining categories. This dataset is used as the training set.

[0139] The nouns or noun phrases in the first-level functional attributes, second-level functional attributes, and final attribute vocabulary in the Truck Platform Product Feature Catalog were hierarchically divided. The division process is shown in Table 2. The classification results of the first-level functional attributes and final attribute vocabulary were finally retained, as shown in Table 3. Then, the valid review texts were manually annotated by the same focus group, totaling 275 items, as shown in Table 3. Figure 6 shown.

[0140] Table 2 Classification process of truck product design elements

[0141]

[0142] Table 3 Truck product design elements stratification results

[0143]

[0144]

[0145] Step 63: Manually label the valid comment texts, that is, manually determine which category of the ten first-level functional attribute labels the valid comment texts belong to sentence by sentence, mark "1" in the label box of the corresponding category, and mark "0" in the label boxes of the remaining categories. This dataset is used as the training set.

[0146] There are a total of six text classification models used in this study, including three machine learning models - Naive Bayes, Random Forest and Support Vector Machines (SVM), and three deep learning models - LSTM algorithm (Long Short Term Memory, long short-term memory network), BiLSTM algorithm (Bi-directional Long Short Term Memory, bidirectional long short-term memory network) and Attention_BiLSTM (bidirectional long short-term memory network based on attention mechanism).

[0147] Naive Bayes is a classification method based on Bayes' theorem and the conditional independence assumption. Given a training dataset, it first learns the joint probability distribution of inputs and outputs based on the conditional independence assumption. Then, based on this model, it applies Bayes' theorem to find the output with the highest posterior probability for a given input. Naive Bayes is a classification algorithm based on probability theory.

[0148] The Random Forest algorithm is a supervised learning algorithm and an ensemble learning algorithm based on decision trees. First, m samples are randomly sampled with replacement from the original dataset to generate m training sets. Then, m decision tree models are trained on each of the m training sets. For each decision tree model, assuming the number of training sample features is n, the best feature is selected for each split based on information gain / information gain-BiGini coefficient. Finally, the resulting multiple decision trees are combined to form a random forest. For classification problems, the final classification result is determined by voting among the multiple tree classifiers. For regression problems, the final prediction result is determined by the mean of the predictions from the multiple trees.

[0149] Support Vector Machines (SVM) are a binary classification model. Its basic model is a linear classifier defined by maximizing the margin in feature space, which distinguishes it from a perceptron. SVM also incorporates kernel techniques, making it a de facto nonlinear classifier. The SVM learning strategy is to maximize the margin, which can be formalized as a convex quadratic programming problem, equivalent to minimizing a regularized hinge loss function. The SVM learning algorithm is an optimization algorithm for solving convex quadratic programming. The core of the SVM algorithm is to find the geometric distance and margin to handle linearly separable problems.

[0150] The LSTM algorithm (Long Short Term Memory) is an important and currently widely used time series algorithm. It is a special type of RNN (Recurrent Neural Network) capable of learning long-term dependencies. It is primarily designed to address the problems of vanishing and exploding gradients during long sequence training. Simply put, compared to ordinary RNNs, LSTMs perform better on longer sequences.

[0151] The Bi-directional Long Short Term Memory (BiLSTM) algorithm consists of two LSTMs: one processes the input sequence in the forward direction and the other processes it in the reverse direction. The outputs of the two LSTMs are then concatenated. The final BiLSTM output is obtained only after all time steps have been calculated. The forward LSTM generates a result vector after m time steps, while the backward LSTM generates another result after n time steps. These two result vectors are concatenated to form the final BiLSTM output.

[0152] Attention_BiLSTM (Attention-based Bidirectional Long Short-Term Memory Networks) was first proposed in the 2015 paper "Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification" published by the Chinese Academy of Sciences. It is mainly used for relation classification. It can use BiLSTM and the Attention mechanism, which can automatically focus on words that have a decisive influence on classification, thereby capturing the most important semantic information in a sentence without using additional knowledge and natural language processing technology systems.

[0153] First, import and read the manually annotated text corpus data, such as Figure 7 Then the text data is cleaned, as shown in Figure 8 As shown, construct the word order as Figure 9 As shown. The cleaning includes extracting valid characters using regular expressions, and then using the open source Jieba word segmentation tool for word segmentation. Stop words are used to remove useless characters and words that are not useful for the research content. After filtering stop words, the word length after word segmentation is also filtered. The length is less than or equal to 1, otherwise it is retained. For tags, the ID number is used to identify them, that is, the current tag is deduplicated and then numbered, and the text is converted into a digital code, such as Figure 10 shown.

[0154] After processing, the Word2vec word vector tool is used to train the model, and then the training set and test set are divided, and a sentiment classification model based on deep learning is built.

[0155] Here is the process of machine learning:

[0156] 1. Perform word segmentation and remove stop words;

[0157] 2. Obtain a text representation model by training TF-IDF;

[0158] 3. Input the features into the corresponding model for training and evaluation indicators.

[0159] The detailed process of three machine learning methods, Naive Bayes, Random Forest and Support Vector Machines (SVM) is as follows: Figure 11 shown.

[0160] Here is the process of deep learning:

[0161] 1. For the data after word segmentation, build a word order dictionary and obtain the word: sequence number;

[0162] 2. Train the embedding of each word, obtained through Word2vec training, and construct it with the word order to obtain the sequence number: embedding;

[0163] 3. Convert the data into ID numbers and randomly divide them into training sets and test sets;

[0164] 4. Build LSTM, BiLSTM, and Attention_BiLSTM, set the corresponding parameters, and perform training.

[0165] The operation process of the BiLSTM algorithm (Bi-directional Long Short Term Memory) is as follows Figure 12 As shown, the operation process of Attention_BiLSTM (bidirectional long short-term memory network based on attention mechanism) is as follows Figure 13 shown.

[0166] Step 64: Use the K-fold cross-validation method to select the text classification results with the highest accuracy for subsequent research.

[0167] The basic idea of cross-validation is to divide the original data set into groups, using one part as the training set and the other as the validation set (or test set). The classifier is first trained on the training set, and then the trained model is tested on the validation set, using this as a performance metric for evaluating the classifier. This study used K-fold cross-validation to select the text classification results with the highest accuracy for subsequent research.

[0168] The sample data was divided into 10 parts and trained and tested using the K-fold cross-validation method. Prediction index values were calculated based on each data point, and the average of these 10 results was taken to determine the final prediction result. Precision (P), recall (R), and F1 score are commonly used metrics for model evaluation. Precision is the percentage of correct predictions across all samples, while recall refers to the percentage of correct similarity predictions within a dataset. The F1 score is a comprehensive metric with a maximum value of 1 and a minimum value of 0, representing the harmonic mean of precision and recall. A higher F1 score indicates a better text similarity model. The average evaluation results of the six models are shown in Table 4, showing that the LSTM model achieved the highest accuracy.

[0169] Table 4 Average evaluation results of 6 text classification models

[0170]

[0171] The results of the LSTM model are exported, and some of the results are shown in Table 5.

[0172] Table 5 Classification results of LST model (partial)

[0173]

[0174]

[0175]

[0176] Step 7: Extraction of attribute words and sentiment words: Specify the criteria for distinguishing explicit and implicit sentence patterns. If the comment text contains both attribute words and sentiment words, it is determined to be an explicit sentence pattern; if the comment text only contains attribute words, it is determined to be an implicit sentence pattern.

[0177] Step 8: Sentiment quantification: Use the SO-PMI algorithm and the HOWNET sentiment polarity quantification standard to analyze and quantify the above results, and obtain the <attribute vocabulary, sentiment mean> set.

[0178] Sentiment polarity analysis uses the SO-PMI algorithm, also known as the sentiment tendency point mutual information algorithm. The overall idea of the algorithm is very simple: it determines the probability of the word P (word) to be judged and the base word P (base) appearing at the same time. If the probability of appearing with positive words is higher, then the word is judged as positive. If the probability of appearing with negative words is higher, then the word is judged as negative. If the probability of appearing with positive and negative words is the same, then the word is judged as neutral.

[0179] The algorithm formula is as follows:

[0180] (1)

[0181] Where: num(pos) refers to the total number of positive benchmark words;

[0182] Similarly, num(neg) refers to the total number of negative benchmark words;

[0183] POS i refers to positive benchmark words;

[0184] neg i Refers to negative benchmark words;

[0185] PMI(word,pos i ) refers to the point-wise mutual information between a word and a positive benchmark word;

[0186] PMI (word, neg i ) refers to the point-wise mutual information between a word and a negative benchmark word;

[0187] The formula produces the following results:

[0188] SO-PMI>0, the words are judged as positive words;

[0189] SO-PMI=0, the word is judged as a neutral word;

[0190] If SO-PMI<0, the word is judged as a negative word.

[0191] After the sentiment polarity analysis is completed, the sentiment mean is calculated according to the HOWNET sentiment polarity quantification standard table.

[0192] Table 6 HOWNET sentiment polarity quantification standard table

[0193]

[0194] Considering the impact of natural conditions on some functional attributes, we used focus groups to develop an environmental factor impact coefficient. All truck attributes were categorized into three categories: unaffected, slightly affected, and significantly affected. Scientifically selected values were then combined with the sentiment mean to calculate satisfaction, resulting in the set "attribute vocabulary, satisfaction level." We then used word frequency statistics to determine user attention based on the frequency of occurrence of each attribute vocabulary. This resulted in the set "attribute vocabulary, satisfaction level, user attention level."

[0195] Step 9: Demand sorting: A truck demand sorting model is constructed by combining the KANO model and the DEMATEL analysis method. The specific process is shown in Table 7. The KANO model is used to find the relationship between demand and user satisfaction and classify demand attributes. The DEMATEL analysis method is used to quantify the impact relationship between demands and finally provide a clear sorting strategy.

[0196] Table 7 Construction of truck demand sorting model

[0197]

[0198] The KANO model is used to classify and prioritize requirements. The implementation steps are as follows:

[0199] Step 911: Questionnaire preparation: Design a questionnaire based on the set of <attribute vocabulary, satisfaction, user attention> to fully understand the user's feedback on truck needs.

[0200] Step 912: Data Processing. To more intuitively prioritize the needs, use the Better-Worse satisfaction coefficient formula to calculate the satisfaction and dissatisfaction coefficients for each need. When calculating, remove indifferent needs and reverse needs. The specific calculation formulas are shown in Formulas (2) and (3).

[0201] The product provides this requirement, and the Better coefficient is:

[0202] Better / SI=(A+O) / (A+O+M+I)(2)

[0203] The product does not meet this requirement, and the Worse coefficient is:

[0204] Worse / DSI=-(O+I) / (A+O+M+I)(3)

[0205] Where: A——Number of charm requirement options;

[0206] O——the expected number of demand options;

[0207] M - the number of required options;

[0208] I – the number of indifferent demand options.

[0209] Based on the Better and Worse coefficient values, a demand quartile chart is drawn. The sorting principle is essential demand M > desired demand O > attractive demand A > indifferent demand I. When multiple demand items belong to the same attribute, the internal sorting requires importance calculation. The weight calculation formula for the i-th demand item is:

[0210] (4)

[0211] Where: w i ——The initial weight of the demand for product i;

[0212] k i ——The adjustment coefficients of essential attributes, desired attributes, attractive attributes, and indifference attributes are 1, 2, 4, and 0 respectively.

[0213] According to w i With k i The value can be used to calculate the weighted weight w' of each demand item i , thereby determining the order of demand items, and combining the classification principles to determine the priority order of final product demand.

[0214] Combining the DEMATEL analysis method, an expert panel was established to further explore the impact relationship between requirements using expert opinions, and transform it into a more objective product demand ranking method. The brief calculation steps are as follows:

[0215] Step 921: By scoring the impact of each demand pairwise, the causal relationship between the demands is quantified to obtain the direct impact matrix A, where a ij Describe the degree of influence of demand i on demand j.

[0216] Step 922: Obtain the normalized direct influence matrix N through equations (5) and (6):

[0217] (5)

[0218] (6)

[0219] Step 923: Solve the comprehensive influence matrix T:

[0220] T=N(IN) -1 (7)

[0221] Where: I——unit matrix.

[0222] Step 924: Set the threshold and output the coordinate information of the impact causal relationship diagram. Impact causal relationship diagram

[0223] It is constructed based on the comprehensive influence matrix T information, and the coordinate information is shown in Table 8.

[0224] Table 8 INRM coordinate information definition

[0225]

[0226] Step 925: Divide the impact and cause-effect diagram into four quadrants and determine the demand categories.

[0227] By calculating the average value of the horizontal vector centrality D+R and the vertical vector causality DR, INRM is divided into four quadrants, such as Figure 15 In the four-quadrant causal relationship diagram, each quadrant represents different meanings and characteristics. The position of a specific requirement in the diagram can be used to determine the category of the requirement to which it belongs.

[0228] The KANO model for dividing user needs is used as the primary criterion, and the DEMATEL analysis method for expert decision-making is used as a secondary criterion for adjustment and supplementation to construct a truck demand ranking hierarchical model, including the target layer, KANO criterion layer, DEMATEL criterion layer and demand ranking table. Among them, the target layer is the truck demand priority ranking; the KANO criterion layer makes a preliminary priority division of product needs, and these needs will be arranged in the order of priority M>O>A>I; the DEMATEL criterion layer identifies key needs and their mutual influence through INRM, and further sorts them in combination with the DEMATEL first quadrant>second quadrant>third quadrant>fourth quadrant sorting method. The needs in the same quadrant of DEMATEL will be sorted from high to low according to the size of the centrality; after all the above needs are sorted qualitatively and quantitatively, a truck demand priority ranking table is formed to more efficiently convey the demand priority and influence relationship information. The specific process is as follows: Figure 16 shown.

[0229] Step 10: Use the questionnaire survey method to conduct an online survey on the satisfaction of some truck users with various design elements of truck products, and compare and analyze the results with the above demand ranking.

[0230] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. A method for mining hidden needs of truck users based on natural language processing technology, characterized by: The following steps are involved: Step 1: Based on the Windows 10 64-bit operating system and Ubuntu 18.0.4 operating system, use PyCharm and Python 3.6 to build an experimental platform and create folders "Truck Information", "Sentiment Dictionary", "Online Reviews", and "Effective Review Text"; Step 2: Collect the truck's literature and information, sentiment dictionary and comment text, and store them in the corresponding folder; Step 3: Separate invalid comments and save valid comments to the "Valid Comment Text" folder; Step 4: Use the NLPIR-ICICLAS Chinese lexical analysis system to preprocess the data in the two folders "Truck Information" and "Effective Comment Text"; Step 5: Divide the vocabulary into attribute vocabulary and sentiment vocabulary. The contents in the "Truck Information" folder are segmented by the NLPIR-ICICLAS Chinese lexical analysis system to obtain nouns or noun phrases as the initial attribute vocabulary; the contents in the "Sentiment Dictionary" folder are segmented by the NLPIR-ICICLAS Chinese lexical analysis system to obtain nouns or noun phrases as candidate attribute vocabulary, and adjectives or adjective phrases as candidate sentiment vocabulary. Finally, filter and remove duplicates to obtain the final attribute vocabulary and final sentiment vocabulary. Step 6: Classify the text of valid comments, manually annotate them and then use the text classification model to classify them; Step 7: Define the criteria for distinguishing explicit and implicit sentence patterns. If the comment text contains both attribute words and sentiment words, it is determined to be an explicit sentence pattern; if the comment text only contains attribute words, it is determined to be an implicit sentence pattern. Step 8: Use the SO-PMI algorithm and HOWNET sentiment polarity quantification standard to analyze and quantify the above results, and obtain the set of <attribute vocabulary, sentiment mean>; Step 9: Build a truck demand ranking model by combining the KANO model and the DEMATEL analysis method. Use the KANO model to find the relationship between demand and user satisfaction and classify demand attributes. Use the DEMATEL analysis method to quantify the impact relationship between demands and ultimately provide a clear ranking strategy. Step 10: Conduct an online survey using questionnaires to investigate truck users' satisfaction with various design elements of truck products, and compare the results with the above demand ranking results. The specific process of step six is as follows: Step 61: To prevent data overfitting during text classification, we first classify truck product design elements. This involves classifying valid review texts based on the ten first-level functional attributes in the Truck Platform Product Feature Catalog. Step 62: Manually label the valid review texts. This means manually determining which of the ten first-level functional attribute labels the valid review texts belong to sentence by sentence. Mark "1" in the label box of the corresponding category, and mark "0" in the label boxes of the remaining categories. This dataset is used as the training set. Step 63: Use machine learning models and deep learning models to learn the training set, and then perform attribute classification on valid review texts that have not been manually annotated; Step 64: Use the K-fold cross-validation method to select the text classification results with the highest accuracy for subsequent research.

2. The method for mining truck users' implicit needs based on natural language processing technology according to claim 1 is characterized by: The data collection of step 2 is as follows: Use the NLPIR-ICICLAS Chinese word segmentation system and the Octopus web crawler tool to obtain information about trucks and save it to the "Truck Information" folder; Use literature search method to obtain multiple dictionaries and save them in the folder of "Emotional Dictionary"; The text of social media comments about trucks was obtained using manual retrieval and the Octopus web crawler tool and saved in the "Online Reviews" folder.

3. The method for mining hidden needs of truck users based on natural language processing technology according to claim 2 is characterized by: In the step three: Definition of a valid comment: (1) Effective reviews must be related to actual product information and not contain irrelevant, noisy information. In addition, reviews should also contain information about the user's experience, not just simple product introduction information. (2) Effective reviews should include product feature dimensions and be detailed; (3) In addition, effective comments are universal in sentence structure and content, not sentences with rare words and special structures; Definition of invalid comments: (1) Invalid reviews are characterized by a low degree of relevance to truck products and do not include detailed descriptions of specific design elements of truck products; (2) Some reviews are obviously inconsistent with the actual product features and are also defined as invalid reviews; Clean up invalid comments to reduce the interference of noise data on experimental results: (1) Deletion of spam comments: Comments that contain only a few words, a string of characters, or are entirely punctuation marks. Such comments have no information value and need to be deleted; (2) Deletion of duplicate comments. Duplicate comments increase the workload and affect the results of sentiment analysis and user attention, so they need to be deleted. (3) Special symbol processing, including emoticons, special symbols or garbled characters, needs to be deleted.

4. The method for mining hidden needs of truck users based on natural language processing technology according to claim 3 is characterized by: The specific process of step 4 is as follows: Step 41: Find the content of interest in the corpus and clean and delete the content that is not interesting or considered as noise; Step 42: Process all text data into the smallest unit granularity - words; Step 43: Label each word with a part of speech label; Step 44: Remove stop words. Stop words refer to words that do not contribute to the text features.

5. The method for mining hidden needs of truck users based on natural language processing technology according to claim 4 is characterized by: The specific process of step eight is as follows: Sentiment polarity analysis uses the SO-PMI algorithm, also known as the sentiment tendency point mutual information algorithm, to determine the probability of the word P (word) to be judged and the base word P (base) appearing together. If the probability of appearing with positive words is higher, then the word is judged as positive. If the probability of appearing with negative words is higher, then the word is judged as negative. If the probability of appearing with positive and negative words is equal, then the word is judged as neutral. The algorithm formula is as follows: （1） Where: num(pos) refers to the total number of positive benchmark words; Similarly, num(neg) refers to the total number of negative benchmark words; POS i refers to positive benchmark words; neg i Refers to negative benchmark words; PMI(word,pos i ) refers to the point-wise mutual information between a word and a positive benchmark word; PMI (word, neg i ) refers to the point-wise mutual information between a word and a negative benchmark word; The formula produces the following results: SO-PMI>0, the words are judged as positive words; SO-PMI=0, the word is judged as a neutral word; SO-PMI<0, the words are judged as negative words; After the sentiment polarity analysis is completed, the sentiment mean is calculated according to the HOWNET sentiment polarity quantification standard table.

6. The method for mining hidden needs of truck users based on natural language processing technology according to claim 5 is characterized by: In step nine, the KANO model is used to classify the attributes and identify the priorities of the requirements. The implementation steps are as follows: Step 9: Design a questionnaire based on the set of <attribute vocabulary, satisfaction, and user attention> to fully understand user feedback on truck needs; Step 912: Data processing. To obtain the priority ranking of needs more intuitively, the Better-Worse satisfaction coefficient calculation formula is used to calculate the satisfaction coefficient and dissatisfaction coefficient of each need. When calculating, indifferent needs and reverse needs are removed. The specific calculation formulas are as follows: (2) and (3); The product provides this requirement, and the Better coefficient is: Better / SI=(A+O) / (A+O+M+I)(2) The product does not meet this requirement, and the Worse coefficient is: Worse / DSI=-(O+I) / (A+O+M+I)(3) Where: A——Number of charm requirement options; O——the expected number of demand options; M - the number of required options; I – number of indifferent demand options; Based on the Better and Worse coefficient values, a demand quartile chart is drawn. The sorting principle is essential demand M > desired demand O > attractive demand A > indifferent demand I. When multiple demand items belong to the same attribute, the internal sorting needs to be calculated by importance. The weight calculation formula for the i-th demand item is: （4） Where: w i ——The initial weight of the demand for product i; k i ——The adjustment coefficients for essential attributes, desired attributes, attractive attributes, and indifferent attributes are 1, 2, 4, and 0, respectively; According to w i With k i The value can be used to calculate the weighted weight w′ of each demand item i , thereby determining the order of demand items, and combining the classification principles to determine the priority order of final product demand.

7. The method for mining hidden needs of truck users based on natural language processing technology according to claim 5 is characterized by: In step nine, the DEMATEL analysis method is combined with the establishment of an expert group. Expert opinions are used to further explore the impact relationship between requirements and transform it into a more objective product demand ranking method. The brief calculation steps are as follows: Step 921: By scoring the impact of each demand pairwise, the causal relationship between the demands is quantified to obtain the direct impact matrix A, where a ij Express the degree of influence of demand i on demand j; Step 922: Obtain the normalized direct influence matrix N through equations (5) and (6): （5） （6） Step 923: Solve the comprehensive influence matrix T: T=N(I-N) -1 (7) Where: I——unit matrix; Step 924: setting the threshold and outputting the coordinate information of the causal relationship graph; Step 925: Divide the impact and cause-effect diagram into four quadrants and determine the demand categories.

Images