Online user comment-based app-based online car-hailing service quality evaluation method and system

Abstract

The application discloses a kind of online user comment-based APP online car-hailing service quality evaluation method and system, the method includes: the pre-processing comment text of online user comment data is obtained;Online car-hailing field library is constructed;Extract service quality candidate feature word and identify potential service theme, obtain service quality feature word set;Establish the mapping relationship between service quality feature word and service quality evaluation index;Service quality index classification is carried out to comment text, and service quality index label is determined;Based on field sentiment word recognition and sentiment correction, calculate each index sentiment performance score;Calculate each index importance weight;Importance-performance analysis matrix is constructed, and service quality short board index and its improvement priority are output.The present application can improve the granularity and accuracy of online car-hailing service quality evaluation, and can automatically diagnose service short board, output improvement priority.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation service evaluation and text data processing technology, specifically relating to a method and system for evaluating the quality of ride-hailing services based on online user reviews via an app. Background Technology

[0002] With the development of mobile internet and the platform economy, ride-hailing has become an important form of urban transportation service. Ride-hailing services combine online platform dispatch with offline transportation services. Users express their experiences regarding vehicle environment, driver behavior, platform pricing, location accuracy, order response, customer service, and payment security through app reviews, app store ratings, and platform feedback channels. Therefore, online user reviews of apps contain a wealth of textual information reflecting the quality of ride-hailing services.

[0003] Current methods for evaluating ride-hailing service quality largely rely on questionnaires, interviews, star ratings, or platform complaint data. While these methods can reflect user satisfaction to some extent, they suffer from problems such as data acquisition delays, coarse-grained evaluations, and high manual processing costs, making it difficult to identify specific service quality issues encountered by users during actual trips in a timely and detailed manner. Furthermore, online user reviews on apps are typically unstructured, colloquial, emotional, involve multiple co-occurrences of indicators, and are semantically ambiguous. Reviews may also contain expressions characteristic of the ride-hailing industry, such as "price gouging," "unauthorized charges," "detours," "location drift," and "long wait times." Using only general keyword matching or general sentiment analysis methods can easily lead to inaccurate identification, biased sentiment quantification, and unclear identification of service shortcomings.

[0004] Therefore, there is an urgent need to provide a technical solution that can perform domain-specific word recognition, service quality indicator mapping, indicator-level sentiment quantification, and service shortcoming diagnosis for online review texts in the ride-hailing field, in order to solve the problems of insufficient real-time performance, coarse evaluation granularity, weak domain adaptability, and difficulty in automatically determining improvement priorities in existing service quality evaluation methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for evaluating the service quality of ride-hailing services based on online user reviews on an app. By constructing a thesaurus and index mapping relationship in the ride-hailing domain, it performs fine-grained index-level sentiment quantification on massive user reviews and automatically diagnoses service shortcomings and their improvement priorities by combining the importance of the indicators. This solves the problems of existing methods having coarse evaluation granularity, weak domain adaptability, and inability to automatically guide the direction of improvement.

[0006] To address the aforementioned technical problems, this invention provides a method for evaluating the quality of ride-hailing services based on online user reviews via an app, comprising: S1. Obtain online user comment data from the ride-hailing app, and preprocess the online user comment data to obtain preprocessed comment text; S2. Construct or update the ride-hailing domain lexicon based on the preprocessed comment text. The ride-hailing domain lexicon includes domain service terms, indicator feature terms, and domain sentiment terms. S3. Based on the preprocessed comment text and the ride-hailing domain thesaurus, extract service quality candidate feature words and identify potential service topics to obtain a service quality feature word set; S4. Establish the mapping relationship between the service quality feature word set and the ride-hailing service quality evaluation index; S5. Based on the mapping relationship, classify the preprocessed comment text according to service quality indicators, and determine at least one service quality indicator label corresponding to the preprocessed comment text; S6. Based on the domain sentiment words in the ride-hailing domain lexicon, identify sentiment words in the comment text, and correct the sentiment polarity and sentiment intensity of the sentiment words to obtain the sentiment score of the comment text under the corresponding service quality indicator. Then, based on the sentiment scores of multiple comment texts under the same service quality indicator, calculate the sentiment performance score of each service quality indicator. S7. Calculate the importance weight of each service quality indicator; S8. Construct an importance-performance analysis matrix using the importance weights of each service quality indicator as importance values ​​and the emotional performance scores of each service quality indicator as performance values. Based on the position of each service quality indicator in the importance-performance analysis matrix, output the service quality weakness indicators and their improvement priorities. According to the above scheme, the preprocessing in S1 includes text cleaning, Chinese word segmentation, and stop word removal. According to the above scheme, S3 includes: extracting service quality candidate feature words from the preprocessed comment text using word frequency statistics, TF-IDF algorithm and TextRank algorithm, identifying potential service topics using LDA topic model, and obtaining a service quality feature word set based on the service quality candidate feature words and the topic words corresponding to the potential service topics. According to the above scheme, S5 includes: when the preprocessed comment text contains explicit service quality feature words, the preprocessed comment text is labeled with the corresponding service quality indicator; when the preprocessed comment text has multiple indicator co-occurrence, semantic ambiguity, polysemy, or feature word mismatch, a semantic classification model is used to perform semantic parsing on the preprocessed comment text and output at least one service quality indicator label. According to the above scheme, the semantic classification model is a text classification model trained based on manually labeled ride-hailing review samples. According to the above scheme, the emotional polarity and emotional intensity of the emotional words in S6 are modified by combining the rules of negation words, the rules of degree adverbs and the preset text window rules. According to the above scheme, step S7 includes: calculating the importance value of each service quality indicator based on the TF-IDF value of the service quality feature words corresponding to each service quality indicator in the preprocessed comment text, and normalizing the importance value to obtain the importance weight of each service quality indicator. According to the above scheme, S8 includes: when the importance weight of the service quality indicator is greater than the importance threshold and the emotional performance score is less than the performance threshold, the service quality indicator is determined as the service quality shortcoming indicator. This invention also provides a ride-hailing service quality evaluation system based on online user reviews via an app, used to execute the ride-hailing service quality evaluation method based on online user reviews via an app as described above, the system comprising: The data acquisition module is used to acquire online user comment data of the ride-hailing APP; the text preprocessing module is used to preprocess the online user comment data to obtain preprocessed comment text. The domain terminology construction module is used to construct or update the ride-hailing domain terminology based on the preprocessed comment text. The ride-hailing domain terminology includes domain service terms, indicator feature terms, and domain sentiment terms. The feature extraction and topic recognition module is used to extract service quality candidate feature words and identify potential service topics based on the preprocessed comment text and the ride-hailing domain lexicon, thereby obtaining a service quality feature word set. The indicator mapping module is used to establish the mapping relationship between the service quality feature word set and the ride-hailing service quality evaluation indicators; The comment classification module is used to classify the preprocessed comment text according to the mapping relationship and determine at least one service quality indicator label corresponding to the preprocessed comment text. The sentiment quantification module is used to identify sentiment words in comment texts based on the domain sentiment words in the ride-hailing domain lexicon, and to correct the sentiment polarity and sentiment intensity of the sentiment words to obtain the sentiment score of the comment text under the corresponding service quality indicator. Then, based on the sentiment scores of multiple comment texts under the same service quality indicator, the sentiment performance score of each service quality indicator is calculated. The indicator weighting module is used to calculate the importance weight of each service quality indicator; The diagnostic output module is used to construct an importance-performance analysis matrix using the importance weight of each service quality indicator as the importance value and the emotional performance score of each service quality indicator as the performance value, and output service quality weakness indicators and their improvement priorities according to the position of each service quality indicator in the importance-performance analysis matrix. According to the above scheme, the comment classification module is used to: when the preprocessed comment text contains explicit service quality feature words, label the preprocessed comment text to the corresponding service quality indicator; when the preprocessed comment text has multiple indicator co-occurrence, semantic ambiguity, polysemy, or feature word mismatch, use a semantic classification model to perform semantic parsing on the preprocessed comment text and output at least one service quality indicator label.

[0007] Beneficial effects This invention effectively addresses the problem of weak domain adaptability of general sentiment analysis methods in the ride-hailing scenario by constructing a ride-hailing-specific thesaurus containing domain service terms, indicator feature terms, and domain sentiment terms. Building upon this, it employs a combination of feature word matching and semantic classification models to classify service quality indicators in review texts. This approach accurately handles complex situations such as a single review involving multiple service issues and semantic ambiguity, significantly improving the accuracy of fine-grained indicator identification. Furthermore, it quantifies indicator-level sentiment based on domain sentiment terms and integrates rules for negation words and degree adverbs, enhancing the recognition accuracy of ride-hailing-specific emotional expressions. Finally, by combining the sentiment performance scores of each indicator with the importance weights calculated based on the discriminative power of feature words in the review text, an importance-performance analysis matrix is ​​constructed. This matrix automatically identifies pain points with high user attention but low satisfaction, directly outputting service quality shortcomings and their clear improvement priorities, providing precise decision-making support for platform operations. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the overall process of evaluating ride-hailing service quality based on online user reviews provided in this embodiment of the invention. Figure 2 This is a structural diagram of a ride-hailing service quality evaluation system based on online user reviews provided in an embodiment of the present invention. Figure 3 A flowchart illustrating the classification of comment text service quality indicators provided in this embodiment of the invention; Figure 4 A flowchart for indicator-level sentiment quantification provided in this embodiment of the invention; Figure 5 A schematic diagram of the importance-performance analysis matrix provided for embodiments of the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0010] Example 1 This embodiment provides a method for evaluating the quality of ride-hailing services based on online user reviews on an app. The overall process of the method is as follows: Figure 1 As shown, it includes the following steps: S1: Obtain online user review data from the ride-hailing app, and preprocess the online user review data to obtain preprocessed review text.

[0011] First, user review data is obtained from ride-hailing apps, app stores, platform rating sections, or third-party review platforms. This user review data includes at least the review text, and may also include information such as review time, user rating, platform name, service type, review source, and service city.

[0012] In one implementation, user review data can be obtained through web crawlers, public platform interfaces, app store review interfaces, or the platform's internal database, and the obtained data can be stored as a structured data table. The structured data table may include fields such as review number, review text, review time, user rating, platform name, review source, service type, and user anonymity identifier.

[0013] In another implementation, the system can automatically acquire new comment data according to a preset period. This preset period can be daily, weekly, or monthly, or it can be set according to the platform's service quality monitoring needs. By periodically updating comment data, the system can continuously reflect changes in the quality of ride-hailing services.

[0014] After obtaining online user comment data, the data is preprocessed to obtain preprocessed comment text. The preprocessing includes text cleaning, Chinese word segmentation, and stop word removal. Further, the preprocessing also includes one or more of the following: duplicate comment deletion, invalid comment filtering, advertising comment filtering, abnormal character cleaning, text normalization, fake comment filtering, and new word discovery.

[0015] Specifically, the duplicate comment deletion is configured to delete completely duplicated or highly similar comment texts; the invalid comment filtering is configured to delete blank comments, garbled comments, and meaningless character comments; the advertisement comment filtering is configured to delete comments containing promotion links, two-dimensional codes, marketing words or irrelevant commodity information; and the abnormal character cleaning is configured to remove web page tags, abnormal symbols, invalid punctuation marks and abnormal emoticons.

[0016] In one embodiment, for the comment text "Great great great great great!!!!", the system may compress consecutive repeated characters and repeated punctuation marks to obtain the normalized text "Great". For comments containing web links, advertisement content or abnormal symbols, the system performs deletion or marking according to preset filtering rules.

[0017] Subsequently, Chinese word segmentation processing is performed on the cleaned comment text, and stop words are removed. The stop words include words such as "de", "le", "shi", "a", "zhe ge", "na ge" and other words that have no obvious distinguishing effect on service quality evaluation.

[0018] In one embodiment, the system further performs fake comment filtering. The fake comment filtering may be judged according to one or more indicators of comment text similarity, comment posting time, abnormal scoring degree, comment length and user behavior features. For example, when a plurality of comment texts appear intensively in a short time and have high text similarity, the system marks them as suspected abnormal comments, and deletes them or reduces their influence in the subsequent evaluation process. The suspected fake comments are deleted, marked or subjected to weight reduction processing.

[0019] S2: Construct or update a ride-hailing field lexicon based on the preprocessed comment text, wherein the ride-hailing field lexicon includes field service words, index feature words and field sentiment words.

[0020] The ride-hailing field lexicon is constructed based on the preprocessed comment text, in combination with word frequency statistics, field new word discovery and manual verification. Specifically, word frequency statistics and field new word discovery are performed on the preprocessed comment text to obtain candidate field words; the ride-hailing field words are determined according to the occurrence frequency, context collocation stability, semantic completeness of the candidate field words and a manual verification result; and the ride-hailing field lexicon is constructed or updated based on the ride-hailing field words. The system may identify new words with ride-hailing industry features according to the occurrence frequency, context collocation stability, semantic completeness of the candidate words and the manual verification result, and add the new words to the ride-hailing field lexicon.

[0021] The terms mentioned in the ride-hailing field include, but are not limited to: location drift, inaccurate location, cancellation fee, suspected differential pricing, price discrimination against loyal customers, taking a longer route, detour, failed order dispatch, slow order dispatch, long wait time, inability to hail a ride, successful carpooling, opaque pricing, fixed price, password-free payment, customer service evasion, unresponsive customer service, poor driver attitude, and unpleasant odor in the vehicle.

[0022] Among them, domain service terms are used to describe the ride-hailing service process, such as "order dispatch," "location," "pricing," "customer service," "driver," "vehicle," and "payment"; indicator feature terms are used to correspond to service quality evaluation indicators, such as "inaccurate location" corresponding to location accuracy and "long waiting time" corresponding to waiting time; domain sentiment terms are used to express users' emotional inclination towards specific service processes, such as "punctual," "convenient," "detour," "arbitrary charges," and "customer service shirking responsibility."

[0023] By building a thesaurus for the ride-hailing industry, the system's ability to recognize industry expressions, colloquial expressions, and emerging internet slang in ride-hailing reviews can be improved.

[0024] S3: Based on the preprocessed comment text and the ride-hailing domain thesaurus, extract service quality candidate feature words and identify potential service topics to obtain a service quality feature word set.

[0025] Based on preprocessed review text and a ride-hailing domain thesaurus, the system first uses a keyword extraction method to extract service quality candidate feature words, then uses a topic recognition model to identify potential service topics in the review text, and forms a service quality feature word set according to the service quality candidate feature words and the topic words corresponding to the potential service topics.

[0026] Specifically, the system employs word frequency statistics, the TF-IDF algorithm, and the TextRank algorithm to extract candidate feature words for service quality from preprocessed review texts, and uses the LDA topic model to identify potential service topics in the review texts. Word frequency statistics are used to extract frequently used words by users, the TF-IDF algorithm is used to filter keywords with textual discriminative power, the TextRank algorithm is used to extract keywords with semantic relevance, and the LDA topic model is used to supplement the identification of potential service topics that are difficult to discover using keywords alone.

[0027] More specifically, the system first obtains high-frequency words from the comment corpus through word frequency statistics, selecting words with high frequency and service quality evaluation significance as candidate feature words. Second, the system uses the TF-IDF algorithm to filter keywords that reflect user focus and differentiate different comment topics. Third, the system uses the TextRank algorithm to analyze the co-occurrence relationships between words, extracting keywords with high semantic relevance in the comment text. Finally, the system uses the LDA topic model to identify potential service topics in the comment text, supplementing service quality issues that are difficult to detect using only high-frequency words.

[0028] In one implementation, the system can identify potential theme clusters such as service perception, platform functions, driver evaluation, software experience, and service expectations, and further organize the keywords under each theme into a set of service quality feature words.

[0029] The system merges, deduplicates, and filters the candidate words obtained from word frequency statistics, TF-IDF, TextRank, and LDA topic models (i.e., the high-frequency candidate words, keywords, and topic words) to obtain the final service quality feature word set.

[0030] S4: Establish the mapping relationship between the set of service quality feature words and the evaluation indicators of ride-hailing service quality.

[0031] A service quality evaluation index system for ride-hailing services is constructed. This index system is used to map service quality feature words and is subsequently used for review classification, sentiment quantification, index weighting, and service quality diagnosis.

[0032] In a preferred embodiment, the service quality evaluation index system includes primary indicators and secondary indicators. The primary indicators include tangibility, reliability, security, responsiveness, empathy, and economy.

[0033] Among them, tangibility is used to characterize user-perceptible service elements such as APP interface, vehicle environment, and vehicle facilities; reliability is used to characterize the stability of services such as platform dispatching, location, trip tracking, and order fulfillment; security is used to characterize security guarantees such as driving safety, payment security, and information security; responsiveness is used to characterize service feedback capabilities such as customer service response, complaint handling, and order response; empathy is used to characterize the platform or driver's understanding and care for users' personalized needs; and economy is used to characterize price-related issues such as pricing standards, reasonableness of fees, price transparency, and suspected differential pricing.

[0034] In one implementation, the secondary indicators include, but are not limited to: APP user interface friendliness, APP system stability, vehicle cleanliness, vehicle comfort, ride-hailing success rate, waiting time, location accuracy, route rationality, driver safety, payment security, customer service responsiveness, complaint response, driver attitude, pricing standard clarity, fare reasonableness, and suspected differential pricing. It should be noted that the above primary and secondary indicators are only preferred implementations. This invention does not limit the specific number of indicators; the system can add, delete, or adjust the indicator system according to different platforms, cities, service types, or evaluation needs.

[0035] Next, a mapping relationship is established between service quality feature words and service quality evaluation indicators. This mapping relationship is used to map feature words in the review text to specific service quality evaluation indicators. In one implementation, the mapping relationship can be stored in the form of a dictionary table, rule table, database mapping table, or vector index library. A mapping table is established containing feature words, indicator numbers, indicator names, their associated service quality dimensions, and matching weights. When a service quality feature word corresponds to multiple service quality evaluation indicators, multiple candidate service quality evaluation indicators are retained, and the final service quality indicator label is determined based on the contextual semantics of the review text.

[0036] For example, the following mapping relationships can be established: "Inaccurate positioning," "positioning drift," and "navigation error" correspond to positioning accuracy; "Long wait for a ride," "Slow order dispatch," and "No one accepting the order for a long time" correspond to waiting time or ride-hailing success rate; "Customer service shirking responsibility," "Customer service not responding," and "Unsuccessful complaints" correspond to customer service responsiveness or complaint responsiveness; "Detours," "Long detours," and "Unreasonable routes" correspond to the rationality of the driving route; "Arbitrary charges," "Opaque pricing," and "Unreasonable fixed price" correspond to the clarity of pricing standards or the rationality of fees; "Price discrimination against loyal customers," "Higher prices for long-term users," and "Different prices for the same route" correspond to suspected differential pricing situations; "Poor driver attitude," "Inappropriate driver communication," and "Poor service attitude" correspond to driver attitude; "Unpleasant odors in the car," "Dirty car," and "Uncomfortable seats" correspond to vehicle cleanliness or vehicle comfort.

[0037] When a feature word may correspond to multiple service quality indicators, the system retains multiple candidate indicators and determines the final indicator label based on the context during the subsequent semantic classification process.

[0038] S5: Based on the mapping relationship, classify the preprocessed comment text according to service quality indicators, and determine at least one service quality indicator label corresponding to the preprocessed comment text.

[0039] Based on the service quality feature words and mapping relationships, the preprocessed comment texts are classified according to service quality indicators, and one or more service quality indicator tags corresponding to the comment text are determined. The service quality indicator classification process for comment texts is as follows: Figure 3 As shown.

[0040] In one implementation, comment classification includes two levels of processing: the first level is preliminary matching based on a feature lexicon, and the second level is classification correction based on a semantic classification model.

[0041] First, the preprocessed comment text is matched against the service quality feature word set. When the preprocessed comment text contains explicit service quality feature words, it is labeled with the corresponding service quality indicator. For example, if the comment text "The location is always inaccurate, the driver can't find me" contains "inaccurate location," the system will label it with the "location accuracy" indicator. Similarly, if the comment text "Customer service doesn't reply for half a day, and no one handles my complaints" contains expressions like "customer service doesn't reply" and "complaint," the system will label it with the "customer service responsiveness" and "complaint responsiveness" indicators.

[0042] Secondly, when the preprocessed review text exhibits multiple co-occurrence of indicators, semantic ambiguity, polysemy, or missed feature words, the system employs a semantic classification model to perform semantic parsing on the preprocessed review text, outputting one or more service quality indicator labels. The semantic classification model is a text classification model trained based on manually labeled ride-hailing review samples. The input to the semantic classification model includes the preprocessed review text, candidate service quality indicators, candidate service quality feature words, and contextual semantic information. The output includes one or more service quality indicator labels corresponding to the review text, as well as the label probability value or confidence score corresponding to each service quality indicator label.

[0043] Specifically, the system uses manually labeled ride-hailing review samples as a training set, which includes review texts and their corresponding service quality index labels. The semantic classification model is trained for domain adaptation using the training set, enabling the semantic classification model to identify key semantic units in the review texts and output corresponding service quality index labels for review texts with multiple co-occurrence of indicators, semantic ambiguity, polysemy, or mismatched feature words.

[0044] For example, a review text like "The driver had a bad attitude, took a long detour, and the software's location was inaccurate" involves driver attitude, route rationality, and location accuracy. The system outputs multiple service quality indicator labels through a semantic classification model. Another example is a review text like "The driver's service attitude was good, but the app was too laggy and kept failing to open," which involves both driver attitude and app system stability. The system identifies the corresponding indicators based on the context.

[0045] Using the above methods, the system can handle issues such as a single comment involving multiple service quality indicators, the same word referring to different indicators in different contexts, and the difficulty in recognizing new industry terms and colloquial expressions.

[0046] S6: Identify sentiment words in the comment text based on the domain sentiment words in the ride-hailing domain lexicon, and correct the sentiment polarity and sentiment intensity of the sentiment words to obtain the sentiment score of the comment text under the corresponding service quality indicator. Then, calculate the sentiment performance score of each service quality indicator based on the sentiment scores of multiple comment texts under the same service quality indicator.

[0047] First, a sentiment lexicon for the ride-hailing industry is constructed. Based on a general sentiment lexicon and online ride-hailing review corpus, a sentiment lexicon for the ride-hailing industry is constructed. This sentiment lexicon is a sentiment quantification sub-lexicon of domain sentiment words in the ride-hailing industry lexicon, used to identify sentiment words in review texts and provide a basis for sentiment polarity and intensity for sentiment quantification. In this embodiment, the process of constructing the ride-hailing industry sentiment lexicon, calculating the sentiment scores of corresponding indicators for review texts, and calculating the sentiment performance scores of each service quality indicator is as follows: Figure 4 As shown.

[0048] In one implementation, the ride-hailing domain sentiment dictionary includes positive sentiment words, negative sentiment words, and neutral sentiment words. Positive sentiment words include, but are not limited to: punctual, convenient, fast, safe, clean, comfortable, good attitude, reasonable price, and quick order acceptance. Negative sentiment words or negative domain expressions include, but are not limited to: taking detours, arbitrary charges, overcharging loyal customers, inaccurate location, long wait times, poor attitude, unhelpful customer service, dirty vehicle, and unsafety. Neutral sentiment words include, but are not limited to: average, ordinary, okay, and normal.

[0049] Each sentiment word is assigned a sentiment polarity and sentiment intensity. The sentiment polarity includes positive, negative, and neutral; the sentiment intensity can be represented by a rating system or a score. Positive sentiment words are used to express user satisfaction or approval, negative sentiment words are used to express user dissatisfaction or complaints, and neutral sentiment words are used to express weak user attitude or indistinct evaluation.

[0050] In one implementation, the data structure of the sentiment dictionary may include sentiment words, sentiment polarity, sentiment intensity, applicable indicators, associated dimensions, and a set of synonyms. For example, "on time" is a positive sentiment word, with the applicable indicator being waiting time; "taking a detour" is a negative sentiment word, with the applicable indicator being the rationality of the driving route; "unauthorized charges" is a negative sentiment word, with the applicable indicator being the rationality of the charges; "location drift" is a negative sentiment word, with the applicable indicator being location accuracy; "customer service shirking responsibility" is a negative sentiment word, with the applicable indicator being customer service responsiveness; and "clean vehicle interior" is a positive sentiment word, with the applicable indicator being vehicle cleanliness.

[0051] Then, sentiment words in the comment text are identified based on the ride-hailing domain sentiment lexicon. Specifically, sentiment words in the comment text are identified based on the service quality indicator tags corresponding to the comment text; and the initial sentiment polarity and initial sentiment intensity of the sentiment words are determined according to the domain sentiment words in the ride-hailing domain lexicon.

[0052] Next, the emotional polarity and intensity of the emotional words are corrected, specifically by combining negation word rules, degree adverb rules, and preset text window rules. Negation words and degree adverbs are identified within the preset text window; the initial emotional polarity and intensity are corrected according to the negation word rules and degree adverb rules to obtain a corrected emotional score; and the corrected emotional score is associated with the service quality indicator corresponding to the service quality indicator label.

[0053] The negative words include, but are not limited to, “no,” “not,” “without,” “not,” “not,” “never,” “not,” “without,” “not quite,” and “not really.” When a negative word appears within a preset text window before or after a sentiment word, the system reverses or weakens the sentiment polarity of the sentiment word. Specifically, the negative word rules include: when there are an odd number of negative words within the preset text window, the sentiment polarity of the sentiment word is changed; when there are an even number of negative words within the preset text window, the sentiment polarity of the sentiment word is maintained. For example, in the comment text “The driver is impolite,” “polite” is a positive sentiment word, but if the negative word “no” appears before it, the system will correct the sentiment tendency of the driver's attitude index corresponding to the comment to negative.

[0054] In one implementation, the preset window range is a preset number of word segments before and after the sentiment word, which can be 3, 5, or 7; preferably, the preset window range is 5 word segments before and after the sentiment word. The degree adverbs include, but are not limited to, "extremely," "very," "especially," "quite," "quite," "relatively," "somewhat," and "slightly." Corresponding weight levels are set for different degree adverbs. When a degree adverb appears within the preset text window range before and after the sentiment word, the system adjusts the sentiment intensity according to the degree adverb level. Specifically, the degree adverb rules include: amplifying or reducing the sentiment intensity of the sentiment word according to the degree weight corresponding to the degree adverb. For example, in the comment text "The driver's attitude was very bad," "bad" is a negative sentiment word, and "very" is an amplifying degree adverb; the system increases the intensity of this negative sentiment.

[0055] In one implementation, the system comprehensively determines the sentiment score of the comment text for the corresponding service quality indicator based on the polarity of sentiment words, the basic strength of sentiment words, the correction results of degree adverbs, and the correction results of negation words.

[0056] When a single review involves multiple service quality indicators, the system calculates the sentiment score for each indicator separately. For example, a review stating "The driver had a bad attitude, took a long detour, and the software's location was inaccurate" involves driver attitude, route rationality, and location accuracy; the system calculates the sentiment score for each of these three indicators separately.

[0057] Finally, based on the indicator tags and sentiment scores corresponding to the comment text, the sentiment performance score for each service quality indicator is calculated. For any service quality indicator, the system aggregates the sentiment scores of all comments categorized under that indicator and calculates the sentiment performance score for that service quality indicator based on one or more factors, including comment quantity, comment source, user rating, comment time, and comment validity.

[0058] In one implementation, the system can calculate the sentiment performance score of the indicator by averaging, that is, by summing up the sentiment scores of all comments under the same indicator and taking the average value.

[0059] In another implementation, the system can use a weighted approach to calculate the sentiment performance score. For example, comments that are more recent, have lower user ratings, higher number of likes, or higher validity can be given higher weights; comments that are suspected of being abnormal or duplicates can be given lower weights or excluded from the calculation.

[0060] Through the above methods, the system can obtain the user satisfaction level or performance corresponding to each service quality indicator.

[0061] S7: Calculate the importance weight of each service quality indicator.

[0062] Based on the TF-IDF values ​​of the service quality feature words corresponding to each service quality indicator in the comment corpus, the importance weight of each service quality indicator is calculated. Specifically, based on the TF-IDF values ​​of the service quality feature words corresponding to each service quality indicator in the preprocessed comment text, the importance value of each service quality indicator is calculated, and the importance value is normalized to obtain the importance weight of each service quality indicator.

[0063] In one implementation, the system calculates the indicator weights based on TF-IDF results. For a given service quality indicator, the system uses all its corresponding feature words as a feature word set and combines the TF-IDF values ​​of the feature words in this set to obtain the importance value of the indicator. Subsequently, the system normalizes the importance values ​​of all indicators to obtain the importance weight of each service quality indicator.

[0064] In another implementation, the system can further adjust the importance value calculated from the TF-IDF value by incorporating factors such as the frequency of indicator occurrence, the number of comments, the concentration of negative comments, user rating differences, and time decay factors to obtain a comprehensive importance weight. Specifically, the system obtains the TF-IDF values ​​of the service quality feature words corresponding to each service quality indicator in the comment corpus; calculates the importance value of each service quality indicator based on the TF-IDF values; adjusts the importance value based on at least one of the following factors: the number of comments corresponding to each service quality indicator, the frequency of occurrence of relevant feature words, the concentration of negative comments, user rating differences, and comment time decay factors; and normalizes the adjusted importance value to obtain the importance weight of each service quality indicator.

[0065] For example, if there are many comments related to "location accuracy" and the relevant feature words have high distinguishability in the comment corpus, the system will identify this indicator as one that users pay close attention to; if "suspected differential pricing" occurs relatively infrequently but has significant negative sentiment and high distinguishability, the system can also increase the importance weight of this indicator.

[0066] S8: Construct an importance-performance analysis matrix using the importance weights of each service quality indicator as importance values ​​and the emotional performance scores of each service quality indicator as performance values. Based on the position of each service quality indicator in the importance-performance analysis matrix, output the service quality weakness indicators and their improvement priorities.

[0067] An importance-performance analysis matrix is ​​constructed using the importance weights of each service quality indicator as importance values ​​and the affective performance scores of each service quality indicator as performance values. Based on the position of each service quality indicator in the matrix, service quality weakness indicators and their improvement priorities are determined. The importance-performance analysis matrix is ​​as follows: Figure 5 As shown.

[0068] In one implementation, the system divides service quality indicators into four zones: a strength maintenance zone, a continuous maintenance zone, a low priority zone, and a key repair zone. The strength maintenance zone represents indicators with high user attention and good performance; the continuous maintenance zone represents indicators with relatively low user attention but good performance; the low priority zone represents indicators with low user attention and low performance; and the key repair zone represents indicators with high user attention but low performance.

[0069] The importance threshold of the importance-performance analysis matrix can be the average, median, or preset threshold of the importance weights of all service quality indicators; the performance threshold of the importance-performance analysis matrix can be the average, median, or preset threshold of the affective performance scores of all service quality indicators. When the importance weight of a service quality indicator is higher than the importance threshold and the affective performance score is lower than the performance threshold, the service quality indicator is identified as a key area for improvement, and the key area for improvement indicator is output as a service quality weakness indicator.

[0070] Indicators in the key improvement areas are prioritized for improvement. The system calculates an improvement priority score based on the importance weight of the indicators in the key improvement areas, the emotional performance score, the number of negative comments, and the intensity of negative sentiment. The indicators in the key improvement areas are then sorted from high to low according to the improvement priority scores, and the system outputs the service quality shortcomings indicators and their improvement priorities.

[0071] In one implementation, the system can output the following priority improvement indicators: clarity of pricing standards, suspected differential pricing, location accuracy, success rate of hailing a ride, user-friendliness of the APP interface, waiting time, customer service responsiveness, and driver attitude.

[0072] Furthermore, after step S8, the method of this embodiment also includes S9: generating a ride-hailing service quality evaluation result.

[0073] Based on the importance weights and affective performance scores of each service quality indicator, the scores for each service quality indicator, the scores for each service quality dimension, and the overall service quality evaluation result are calculated. Specifically, the service quality indicator scores are calculated based on the weights and affective performance scores of each service quality indicator; the service quality dimension scores are calculated based on the scores of multiple service quality indicators under the same service quality dimension; and the overall service quality evaluation result is generated based on the scores of multiple service quality dimensions.

[0074] In one implementation, the system can use weighted summation, weighted averaging, normalized weighted calculation, or a combination thereof to synthesize the importance weights and affective performance scores of a service quality indicator, thereby obtaining an evaluation score for that indicator. For a primary service quality dimension, the system generates an evaluation score for that dimension based on the evaluation scores of each secondary indicator under that dimension. For the comprehensive service quality evaluation result, the system generates an overall evaluation result based on the scores of each primary dimension.

[0075] The system can output the following evaluation results: scores for each service quality dimension; scores for each service quality indicator; number of comments for each indicator; proportion of positive comments for each indicator; proportion of negative comments for each indicator; overall service quality evaluation result; and service quality trend.

[0076] In one implementation, the system can output service quality evaluation results according to different time periods. For example, evaluation results can be generated weekly, monthly, or quarterly to observe the changing trends of ride-hailing service quality.

[0077] Furthermore, after step S9, the method of this embodiment also includes S10: generating a ride-hailing service quality evaluation report.

[0078] Based on the service quality evaluation results and service quality deficiency diagnosis results, the system generates a ride-hailing service quality evaluation report, which is then stored, displayed, or exported as a structured output. The evaluation report includes the data collection time range, the number of valid reviews, scores for each service quality dimension, scores for each service quality indicator, emotional performance score, indicator weights, importance-performance analysis matrix results, service quality deficiency indicators, and improvement priority ranking.

[0079] In one implementation, the evaluation report can be output in the form of a table, bar chart, line chart, radar chart, importance-performance analysis matrix, or text summary.

[0080] In another implementation, the system sets service quality warning thresholds. When the sentiment performance score of a service quality indicator falls below a preset threshold, or when an indicator is in a critical improvement zone for multiple consecutive evaluation periods, the system generates a service quality warning. For example, when "location accuracy" is in a critical improvement zone for multiple consecutive evaluation periods, the system generates a location service quality warning; when the proportion of negative reviews for "customer service responsiveness" exceeds a preset threshold, the system generates a customer service warning; when the number of reviews related to "suspected differential pricing" increases significantly in a short period, the system generates a price fairness warning.

[0081] Example 2 This embodiment provides a ride-hailing service quality evaluation system based on online user reviews via an app. The system structure is as follows: Figure 2 As shown, the system is used to perform the method described in Embodiment 1. The system includes a data acquisition module, a text preprocessing module, a domain lexicon construction module, a feature extraction and topic recognition module, an indicator mapping module, a comment classification module, a sentiment quantification module, an indicator weighting module, an evaluation calculation module, and a diagnostic output module.

[0082] The data acquisition module is used to acquire online user review data for ride-hailing apps. The user review data includes at least the review text, and may also include one or more of the following: review time, user rating, platform name, service type, and review source.

[0083] The text preprocessing module is used to preprocess online user comment data to obtain preprocessed comment text. The preprocessing includes one or more of the following: duplicate comment deletion, invalid comment filtering, advertising comment filtering, abnormal character cleaning, fake comment filtering, text normalization, Chinese word segmentation, and stop word removal.

[0084] The domain lexicon construction module is used to build a ride-hailing domain lexicon based on preprocessed comment text. The ride-hailing domain lexicon includes domain service terms, indicator feature terms, and domain sentiment terms.

[0085] The feature extraction and topic recognition module is used to extract service quality candidate feature words from the preprocessed review text using word frequency statistics, TF-IDF algorithm and TextRank algorithm, and to perform potential topic recognition on the preprocessed review text using LDA topic model, and to obtain a service quality feature word set based on the service quality candidate feature words and the topic words corresponding to the potential service topics.

[0086] The indicator mapping module is used to construct a service quality evaluation indicator system for ride-hailing services, establish a mapping relationship between service quality evaluation indicators and service quality feature words, and map service quality feature words to service quality evaluation indicators according to the mapping relationship.

[0087] The comment classification module is used to perform preliminary classification of preprocessed comment text based on service quality feature words and mapping relationships. When the comment text has multiple co-occurrence of indicators, semantic ambiguity, polysemy, or feature word mismatch, a semantic classification model is used to perform semantic parsing on the comment text and output one or more service quality indicator labels. Specifically, the comment classification module is used to: label the preprocessed comment text with the corresponding service quality indicator when the preprocessed comment text contains explicit service quality feature words; and when the preprocessed comment text has multiple co-occurrence of indicators, semantic ambiguity, polysemy, or feature word mismatch, a semantic classification model is used to perform semantic parsing on the preprocessed comment text and output at least one service quality indicator label.

[0088] In one embodiment, the comment classification module includes a feature word matching unit and a semantic classification unit. The feature word matching unit is used to match the preprocessed comment text with the service quality feature word set, and when the preprocessed comment text contains explicit service quality feature words, it labels the preprocessed comment text with the corresponding service quality indicator. The semantic classification unit is used to perform semantic parsing on the preprocessed comment text when there are multiple indicator co-occurrences, semantic ambiguity, polysemy, or feature word mismatches, and output one or more service quality indicator labels.

[0089] The sentiment quantification module is used to identify sentiment words in comment texts based on a sentiment dictionary for the ride-hailing industry, and to modify the sentiment polarity and intensity by combining negation word rules, degree adverb rules and preset window rules, so as to obtain the sentiment score of the comment text under the corresponding service quality indicator, and to calculate the sentiment performance score of each service quality indicator based on the sentiment score of the comment text under the same service quality indicator.

[0090] In one implementation, the sentiment quantification module includes a sentiment word recognition unit, a negation word recognition unit, a degree adverb recognition unit, a sentiment correction unit, and an index aggregation unit. The sentiment word recognition unit is used to identify sentiment words in the comment text; the negation word recognition unit is used to identify negation words within a preset text window; the degree adverb recognition unit is used to identify degree adverbs within a preset text window; the sentiment correction unit is used to correct sentiment polarity and sentiment intensity according to negation word rules and degree adverb rules; and the index aggregation unit is used to aggregate the corrected sentiment scores to the corresponding service quality indicators.

[0091] The indicator weighting module is used to calculate the importance value of each service quality indicator based on the TF-IDF value of the service quality feature words corresponding to each service quality indicator in the comment corpus, and to normalize the importance value to obtain the importance weight of each service quality indicator.

[0092] The evaluation calculation module is used to calculate the scores of each service quality indicator, the scores of each service quality dimension, and the comprehensive service quality evaluation results based on the importance weights and emotional performance scores of each service quality indicator.

[0093] The diagnostic output module is used to construct an importance-performance analysis matrix by using the importance weight of each service quality indicator as the importance value and the emotional performance score of each service quality indicator as the performance value. Based on the position of each service quality indicator in the matrix, it determines the advantage maintenance zone, the continuous maintenance zone, the low priority zone, and the key repair zone, and then outputs the ride-hailing service quality shortcoming indicators and their improvement priorities.

[0094] In one implementation, the diagnostic output module includes a matrix construction unit, a region division unit, and a priority ranking unit. The matrix construction unit is used to construct an importance-performance analysis matrix using the weights of service quality indicators as importance values ​​and the emotional performance scores of service quality indicators as performance values. The region division unit is used to divide service quality indicators into advantage maintenance zones, continuous maintenance zones, low priority zones, or key improvement zones. The priority ranking unit is used to rank the service quality indicators in the key improvement zones and output the service quality weakness indicators and their improvement priorities.

[0095] In one implementation, the diagnostic output module is further used to generate a ride-hailing service quality evaluation report. The evaluation report includes the data collection time range, the number of valid reviews, service quality dimension scores, service quality indicator scores, emotional performance scores, indicator weights, service quality weakness indicators, and improvement priority ranking.

[0096] This invention enables fully automated processing of ride-hailing service quality evaluation, directly utilizing massive amounts of online review data, significantly reducing the workload of traditional questionnaire surveys and manual review analysis. Furthermore, through continuous collection and real-time analysis of review data, it can capture user feedback more promptly than traditional delayed evaluation methods, and refine the evaluation granularity to specific service stages.

[0097] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0098] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the quality of ride-hailing services based on online user reviews on an app, characterized in that, include: S1. Obtain online user comment data from the ride-hailing app, and preprocess the online user comment data to obtain preprocessed comment text; S2. Construct or update the ride-hailing domain lexicon based on the preprocessed comment text. The ride-hailing domain lexicon includes domain service terms, indicator feature terms, and domain sentiment terms. S3. Based on the preprocessed comment text and the ride-hailing domain thesaurus, extract service quality candidate feature words and identify potential service topics to obtain a service quality feature word set; S4. Establish the mapping relationship between the service quality feature word set and the ride-hailing service quality evaluation index; S5. Based on the mapping relationship, classify the preprocessed comment text according to service quality indicators, and determine at least one service quality indicator label corresponding to the preprocessed comment text; S6. Based on the domain sentiment words in the ride-hailing domain lexicon, identify sentiment words in the comment text, and correct the sentiment polarity and sentiment intensity of the sentiment words to obtain the sentiment score of the comment text under the corresponding service quality indicator. Then, based on the sentiment scores of multiple comment texts under the same service quality indicator, calculate the sentiment performance score of each service quality indicator. S7. Calculate the importance weight of each service quality indicator; S8. Construct an importance-performance analysis matrix using the importance weights of each service quality indicator as importance values ​​and the emotional performance scores of each service quality indicator as performance values. Based on the position of each service quality indicator in the importance-performance analysis matrix, output the service quality weakness indicators and their improvement priorities.

2. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 1, is characterized in that... The preprocessing in S1 includes text cleaning, Chinese word segmentation, and stop word removal.

3. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 1, is characterized in that... S3 includes: extracting service quality candidate feature words from the preprocessed comment text using word frequency statistics, TF-IDF algorithm and TextRank algorithm, identifying potential service topics using LDA topic model, and obtaining a service quality feature word set based on the service quality candidate feature words and the topic words corresponding to the potential service topics.

4. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 1, is characterized in that... S5 includes: when the preprocessed comment text contains explicit service quality feature words, labeling the preprocessed comment text with the corresponding service quality indicator; when the preprocessed comment text has multiple indicator co-occurrence, semantic ambiguity, polysemy, or feature word mismatch, using a semantic classification model to perform semantic parsing on the preprocessed comment text and outputting at least one service quality indicator label.

5. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 4, is characterized in that... The semantic classification model is a text classification model trained based on manually labeled ride-hailing review samples.

6. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 1, is characterized in that... In step S6, the emotional polarity and intensity of the emotional words are modified by combining the rules of negation words, the rules of degree adverbs, and the preset text window rules.

7. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 1, is characterized in that... S7 includes: calculating the importance value of each service quality indicator based on the TF-IDF value of the service quality feature words corresponding to each service quality indicator in the preprocessed comment text, and normalizing the importance value to obtain the importance weight of each service quality indicator.

8. The method for evaluating ride-hailing service quality based on online user reviews via an app, as described in claim 1, is characterized in that... S8 includes: when the importance weight of a service quality indicator is greater than the importance threshold and the emotional performance score is less than the performance threshold, the service quality indicator is identified as a service quality weakness indicator.

9. A ride-hailing service quality evaluation system based on online user reviews via an app, characterized in that, The system is used to perform the ride-hailing service quality evaluation method based on online user reviews of an app as described in any one of claims 1 to 8, the system comprising: The data acquisition module is used to acquire online user comment data of the ride-hailing APP; the text preprocessing module is used to preprocess the online user comment data to obtain preprocessed comment text. The domain terminology construction module is used to construct or update the ride-hailing domain terminology based on the preprocessed comment text. The ride-hailing domain terminology includes domain service terms, indicator feature terms, and domain sentiment terms. The feature extraction and topic recognition module is used to extract service quality candidate feature words and identify potential service topics based on the preprocessed comment text and the ride-hailing domain lexicon, thereby obtaining a service quality feature word set. The indicator mapping module is used to establish the mapping relationship between the service quality feature word set and the ride-hailing service quality evaluation indicators; The comment classification module is used to classify the preprocessed comment text according to the mapping relationship and determine at least one service quality indicator label corresponding to the preprocessed comment text. The sentiment quantification module is used to identify sentiment words in comment texts based on the domain sentiment words in the ride-hailing domain lexicon, and to correct the sentiment polarity and sentiment intensity of the sentiment words to obtain the sentiment score of the comment text under the corresponding service quality indicator. Then, based on the sentiment scores of multiple comment texts under the same service quality indicator, the sentiment performance score of each service quality indicator is calculated. The indicator weighting module is used to calculate the importance weight of each service quality indicator; The diagnostic output module is used to construct an importance-performance analysis matrix using the importance weight of each service quality indicator as the importance value and the emotional performance score of each service quality indicator as the performance value, and output service quality weakness indicators and their improvement priorities according to the position of each service quality indicator in the importance-performance analysis matrix.

10. The ride-hailing service quality evaluation system based on online user reviews via an app, as described in claim 9, is characterized in that... The comment classification module is used to: when the preprocessed comment text contains explicit service quality feature words, label the preprocessed comment text with the corresponding service quality indicator; when the preprocessed comment text has multiple indicator co-occurrence, semantic ambiguity, polysemy, or feature word mismatch, use a semantic classification model to perform semantic parsing on the preprocessed comment text and output at least one service quality indicator label.

Images