User complaint text processing method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a user complaint text processing method and device, electronic equipment, a computer readable storage medium and a computer program product. The user complaint text processing method comprises the steps of determining distribution characteristics of complaint texts, wherein the distribution characteristics of the complaint texts comprise complaint question classification tags; performing topic modeling on the complaint text to obtain a topic and a topic keyword of the complaint text; determining an association rule between the subject keyword and the complaint question classification tag; and improving a product or service corresponding to the complaint question classification tag based on the subject keyword of which the association rule with the complaint question classification tag is determined. According to the user complaint text processing method, a visual relation can be established between the complaint text and the improvement of the product and the improvement of the service, so that the product or the service can be accurately, timely and efficiently improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method for processing user complaint texts, a device for processing user complaint texts, an electronic device, a computer-readable storage medium, and a computer program product. Background Art

[0002] Complaints are a kind of behavior in which users actively share information and give feedback, which can reflect the problems encountered by users and contain rich and valuable information. Therefore, user complaint texts are an important indicator to measure the quality of products or services.

[0003] Analyzing user complaint texts can not only quickly reveal urgent issues regarding product attributes or service quality, but also understand and respond to changes in user preferences, discover users' potential needs, thereby guiding the improvement of goods or services and preventing user churn.

[0004] Therefore, there is an urgent need for an accurate, timely and efficient method or device for processing user complaint texts, so as to establish an intuitive connection between the complaint texts and product improvements and service improvements. Summary of the invention

[0005] The present application provides a user complaint text processing method, a user complaint text processing device, an electronic device, a computer-readable storage medium, and a computer program product that can at least partially solve the above-mentioned problems or other problems.

[0006] On the one hand, the present application provides a method for processing user complaint texts, the method comprising: determining the distribution characteristics of the complaint texts, wherein the distribution characteristics of the complaint texts include complaint problem classification labels; performing topic modeling on the complaint texts to obtain the topics and topic keywords of the complaint texts; determining association rules between the topic keywords and the complaint problem classification labels; and improving products or services corresponding to the complaint problem classification labels based on the topic keywords for which the association rules have been determined with the complaint problem classification labels.

[0007] In one embodiment of the present application, subject modeling is performed on the complaint text to obtain the subject and subject keywords of the complaint text, including: using an unsupervised probability model to perform subject modeling on the complaint text to obtain the subject and subject keywords, wherein the unsupervised probability model includes a latent Dirichlet distribution topic model.

[0008] In one embodiment of the present application, the method also includes: performing sentiment analysis on the multiple complaint texts based on a sentiment dictionary or machine learning to determine complaint texts with high negative sentiment intensity among the multiple complaint texts; and performing topic modeling on the complaint texts with high negative sentiment intensity.

[0009] In one embodiment of the present application, the sentiment analysis includes sentiment polarity and sentiment intensity, and the sentiment polarity includes positive and negative, wherein based on a sentiment dictionary or machine learning, sentiment analysis is performed on the multiple complaint texts to determine the complaint texts with high negative sentiment intensity among the multiple complaint texts, which includes: determining the sentiment polarity scores of the multiple complaint texts based on the sentiment dictionary or the machine learning; and determining the texts among the multiple complaint texts whose sentiment polarity scores are less than a first threshold as the complaint texts with high negative sentiment intensity.

[0010] In one embodiment of the present application, subject modeling is performed on the complaint text to obtain the subject and subject keywords of the complaint text, including: using at least one of a statistical method and an expert method to determine the number of topics for the subject modeling.

[0011] In one embodiment of the present application, using at least one of a statistical method and an expert method to determine the number of topics for the topic modeling includes: using at least one of a statistical method and an expert method to determine a plurality of different initialization topic numbers; performing the topic modeling on the complaint text based on the initialization topic, and obtaining a word group corresponding to the initialization topic; determining a consistency score of the initialization topic based on the semantic similarity values ​​between a plurality of words with relatively high frequency of occurrence in the word group; and determining the number of initialization topics corresponding to the highest consistency score as the number of topics for the topic modeling.

[0012] In one embodiment of the present application, topic modeling of the complaint text also includes: based on the bag-of-words model, converting the text data of the complaint text into word vectors, and inputting the word vectors into an unsupervised probability model to perform the topic modeling; and assigning weight parameters of different sizes to word vectors with different occurrence frequencies, wherein the higher the frequency of occurrence, the smaller the weight parameter of the word vector.

[0013] In one embodiment of the present application, determining the association rules between the subject keywords and the complaint problem classification labels includes: matching the subject keywords and the complaint problem classification labels to form an item set; processing the item set based on an association rule algorithm to obtain an initial rule set; and determining the association rules from the initial rule set based on a screening rule.

[0014] In one embodiment of the present application, processing the item set based on an association rule algorithm to obtain an initial rule set includes: determining the initial rule set based on a predetermined minimum support and a predetermined minimum confidence.

[0015] In one embodiment of the present application, based on the screening rules, determining the association rules from the initial rule set includes: determining the confidence, support and lift of each of the initial rule set; and determining the association rules based on a minimum confidence threshold, a minimum support threshold and a minimum lift threshold.

[0016] In one embodiment of the present application, the method also includes preprocessing the complaint text, and the preprocessing includes: deleting punctuation marks, numbers, modal particles and sparse words in the complaint text to obtain valid data in the complaint text, wherein the sparse words are words whose probability of appearing in the complaint text is less than a second threshold; and segmenting the valid data to obtain word segments of the complaint text.

[0017] In one embodiment of the present application, the distribution characteristics of the complaint text also include: complaint category distribution and word frequency distribution of the segmented words.

[0018] On the other hand, the present application provides a user complaint text processing device, which includes: a distribution feature acquisition module, configured to determine the distribution features of the complaint text, wherein the distribution features of the complaint text include complaint problem classification labels; a modeling module, configured to perform topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text; and an association rule determination module, configured to determine the association rules between the topic keywords and the complaint problem classification labels; and an improvement module, configured to improve the product or service corresponding to the complaint problem classification label based on the topic keywords for which the association rules have been determined with the complaint problem classification label.

[0019] In one embodiment of the present application, the modeling module is further configured to use an unsupervised probability model to perform topic modeling on the complaint text to obtain the topic and the topic keywords, wherein the unsupervised probability model includes a latent Dirichlet distribution topic model.

[0020] In one embodiment of the present application, the device also includes a sentiment analysis module, which is configured to perform sentiment analysis on the multiple complaint texts based on a sentiment dictionary or machine learning to determine the complaint texts with high negative sentiment intensity among the multiple complaint texts; and the modeling module is further configured to perform topic modeling on the complaint texts with high negative sentiment intensity.

[0021] In one embodiment of the present application, the sentiment analysis includes sentiment polarity and sentiment intensity, the sentiment polarity includes positive and negative, and the sentiment analysis module is further configured to determine the sentiment polarity scores of the multiple complaint texts based on the sentiment dictionary or the machine learning; and determine the text in the multiple complaint texts whose sentiment polarity score is less than a first threshold as the complaint text with high negative sentiment intensity.

[0022] In one embodiment of the present application, the modeling module is further configured to determine the number of topics for the topic modeling using at least one of a statistical method and an expert method.

[0023] In one embodiment of the present application, the modeling module is further configured to determine a plurality of different numbers of initialization topics using at least one of a statistical method and an expert method; perform the topic modeling on the complaint text based on the initialization topics, and obtain a word group corresponding to the initialization topic; determine the consistency score of the initialization topic based on the semantic similarity values ​​between a plurality of words with relatively high occurrence frequencies in the word group; and determine the number of initialization topics corresponding to the highest consistency score as the number of topics for the topic modeling.

[0024] In one embodiment of the present application, the modeling module is further configured to convert the text data of the complaint text into word vectors based on a bag-of-words model, and input the word vectors into an unsupervised probability model to perform the topic modeling; and assign weight parameters of different sizes to word vectors with different occurrence frequencies, wherein the higher the frequency of occurrence, the smaller the weight parameter of the word vector.

[0025] In one embodiment of the present application, the association rule determination module is further configured to match the topic keywords and the complaint problem classification labels to form an item set; based on the association rule algorithm, process the item set to obtain an initial rule set; and based on the screening rules, determine the association rules from the initial rule set.

[0026] In one embodiment of the present application, the association rule determination module is further configured to determine the initial rule set based on a predetermined minimum support and a predetermined minimum confidence.

[0027] In one embodiment of the present application, the association rule determination module is further configured to determine the confidence, support and lift of each of the initial rule sets; and determine the association rules based on the minimum confidence threshold, the minimum support threshold and the minimum lift threshold.

[0028] In one embodiment of the present application, the device also includes a preprocessing module, which is configured to delete punctuation marks, numbers, modal particles and sparse words in the complaint text to obtain valid data in the complaint text, wherein the sparse words are words whose probability of appearing in the complaint text is less than a second threshold; and to segment the valid data to obtain word segments of the complaint text.

[0029] In one embodiment of the present application, the distribution characteristics of the complaint text also include: complaint category distribution and word frequency distribution of the segmented words.

[0030] On the other hand, the present application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the user complaint text processing method provided in any embodiment of the first aspect of the present application.

[0031] On the other hand, the present application provides a computer-readable storage medium storing a computer program, characterized in that when the computer program is executed by a processor, it implements the user complaint text processing method provided in any implementation method in one aspect of the present application.

[0032] On the other hand, the present application provides a computer program product storing a computer program, characterized in that when the computer program is executed by a processor, it implements the user complaint text processing method provided in any implementation method in one aspect of the present application.

[0033] According to the user complaint text processing method, user complaint text processing method device, electronic device, computer-readable storage medium and computer program product provided by at least one embodiment of the present application, the subject and subject keywords of the complaint text can be obtained by performing topic modeling on the user's complaint text, and by establishing association rules between subject keywords and complaint problem classification labels, an intuitive connection can be established between the complaint text and product improvements and service improvements, so that products or services corresponding to the complaint problem classification labels can be improved accurately, timely and efficiently based on the subject keywords. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings, in which: Figure 1 is a flowchart of a method for processing user complaint text according to one embodiment of the present application; Figure 2is a distribution feature table of complaint text according to an implementation method of the present application; Figure 3 It is a complaint category distribution statistical chart of the complaint text according to an implementation method of the present application; Figure 4 This is a statistical diagram of the complaint problem classification label distribution of the complaint text according to an implementation method of this application. Figure 5 It is a word frequency distribution cloud diagram of the word segmentation of the complaint text according to one implementation mode of the present application; Figure 6 is a word segmentation table of the complaint text after preprocessing according to an implementation mode of the present application; Figure 7 is a distribution diagram of sentiment polarity scores of multiple complaint texts according to an embodiment of the present application; Figure 8 is a distribution diagram of sentiment polarity scores of multiple complaint texts with high negative sentiment intensity according to an embodiment of the present application; Fig. 9 is a working process diagram of an unsupervised probability model according to one embodiment of the present application; Fig.10 is a distribution diagram of the consistency scores of the models under different subject settings according to one embodiment of the present application; Fig.11 It is a visual distribution diagram of 78 topics according to one embodiment of the present application and an appearance frequency arrangement diagram of each word in the word group corresponding to topic No. 1; Fig.12 is a topic keyword table of some topics with a number of 78 topics according to one embodiment of the present application; Fig.13 is a topic distribution diagram with 78 topics according to one embodiment of the present application; Fig.14 is a topic distribution diagram of a complaint text with a sentiment polarity score of 0 according to an embodiment of the present application; Fig.15 It is a flowchart for determining association rules between topic keywords and complaint problem classification labels according to one implementation of the present application; Fig.16 is a list of some initial rules with higher confidence in the initial rule set according to one embodiment of the present application; Fig.17 is a list of some initial rules with higher support in the initial rule set according to an implementation mode of the present application Fig.18 is a list of some initial rules with higher improvement in the initial rule set according to one embodiment of the present application; Fig.19 is a schematic diagram of a framework of a user complaint text processing device according to an embodiment of the present application; and Fig. 20 It is a structural diagram of a computer system according to one embodiment of the present application, which is suitable for implementing the electronic device of the embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to better understand the present application, a more detailed description will be made of various aspects of the present application with reference to the accompanying drawings. It should be understood that these detailed descriptions are only descriptions of exemplary embodiments of the present application, and are not intended to limit the scope of the present application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0036] It should be noted that in this specification, the expressions first, second, third, etc. are only used to distinguish one feature from another feature area, and do not represent any limitation on the features, especially do not represent any order of precedence. Therefore, without departing from the teaching of this application, the first threshold discussed in this application may also be referred to as the second threshold, and vice versa.

[0037] As used herein, the terms "substantially," "approximately," and similar terms are used as terms of approximation, not degree, and are intended to account for the inherent variations in measurements or calculations that would be recognized by those of ordinary skill in the art.

[0038] It should also be understood that expressions such as "include", "including", "have", "contain" and / or "comprising" are open rather than closed expressions in this specification, which indicate the presence of the stated features, elements and / or components, but do not exclude the presence of one or more other features, elements, components and / or combinations thereof. In addition, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features rather than just the individual elements in the list. In addition, when describing embodiments of the present application, "may" is used to mean "one or more embodiments of the present application". And, the term "exemplary" is intended to refer to an example or illustration.

[0039] Unless otherwise defined, all words (including engineering terms and scientific and technological terms) used in this article have the same meaning as those commonly understood by ordinary technicians in the field to which this application belongs. It should also be understood that unless there is a clear explanation in this application, words defined in commonly used dictionaries should be interpreted as having the same meaning as their meaning in the context of the relevant technology, and should not be interpreted in an idealized or overly formal sense.

[0040] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. In addition, unless explicitly limited or contradictory to the context, the specific steps included in the method described in the present application are not necessarily limited to the order described, but can be performed in any order or in parallel. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0041] In addition, in the present application, when “connected” or “coupled” is used, it may indicate direct contact or indirect contact between corresponding components, unless otherwise clearly defined or inferred from the context.

[0042] Figure 1 FIG. 1 is a flowchart of a method 1000 for processing user complaint text according to one embodiment of the present invention. Figure 1 As shown, the present application provides a user complaint text processing method 1000 including: S1, determine the distribution characteristics of the complaint text, where the distribution characteristics of the complaint text include the complaint problem classification label.

[0043] S2, conduct topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text.

[0044] S3, determine the association rules between topic keywords and complaint problem classification labels.

[0045] S4, based on the topic keywords that have determined the association rules with the complaint problem classification tags, improve the products or services corresponding to the complaint problem classification tags.

[0046] The following will be combined Figures 2 to 7 The specific steps of the above-mentioned user complaint text processing method 1000 are described in detail.

[0047] Step S1 Figure 2 It is a distribution feature table of complaint text according to an implementation method of the present application.

[0048] like Figure 1 and Figure 2 As shown, in some embodiments of the present application, the distribution characteristics of the complaint text may include the complaint problem classification label, the complaint category distribution, and the word frequency distribution of the complaint text segmentation. Optionally, the distribution characteristics of the complaint text may also include: the type of the complained unit and the content of the complaint text.

[0049] Taking the complaint text of passengers on public air transport as an example, Figure 2A distribution characteristic table of some complaint texts among the passenger complaints with a certain city as the departure point and destination during May-June 2023 is shown, of which there are a total of 5,627 complaint texts and 5 partial complaint texts are shown.

[0050] Figure 3 It is a statistical chart of complaint category distribution of complaint text according to an implementation method of the present application. Figure 4 It is a statistical chart of the complaint problem classification label distribution of the complaint text according to an implementation method of the present application. Figure 5 It is a word frequency distribution cloud diagram of the word segmentation of the complaint text according to an implementation mode of the present application.

[0051] refer to Figure 2-Figure 5 In this embodiment, the distribution characteristics of the complaint text of public air transport passengers include complaint category distribution, complaint problem classification label, type of complained unit, complaint text content and word frequency distribution of word segmentation of complaint text.

[0052] Specifically, refer to Figure 2-Figure 3 , Figure 3 The horizontal coordinates represent different complaint categories. As an option, the complaint category distribution may include ticketing services, abnormal flight services, check-in and boarding, baggage services, air services, membership services, customer service, information notification, special passenger services, other services, etc. Figure 3 The vertical coordinate of is the number of complaint texts corresponding to different complaint categories.

[0053] Among the 5,627 complaint texts, complaints about "ticketing services" accounted for the largest proportion, up to about 53%. Preliminary analysis shows that complaints corresponding to this complaint category may include passengers' belief that the refund fee rules are unreasonable and that ticket prices fluctuate too frequently. In addition, complaints about "abnormal flight services" accounted for about 25%. Preliminary analysis shows that complaints corresponding to this complaint category may include passengers' request for airlines to compensate passengers for the price difference of purchasing another ticket after the flight schedule is changed. Complaints about "check-in and boarding" accounted for about 8%. Preliminary analysis shows that complaints corresponding to this complaint category may include security inspection issues, passengers' disapproval of ticket changes, and refund processing plans. In other words, there may be multiple complaint problem classification labels corresponding to any complaint category in the complaint category distribution. In other words, multiple possible complaint issues may lead to the occurrence of the same complaint category.

[0054] refer to Figure 2 and Figure 4 , Figure 4 The horizontal coordinates are different complaint problem classification labels. As an option, the complaint problem classification labels may include compensation standards, ticket refund rules, wrong purchase, ticket refund, refund and change rules, change rules, service attitude, etc. Figure 4The vertical coordinate of is the number of complaint texts corresponding to different complaint problem classification labels. Optionally, each complaint text may be manually annotated to determine the complaint problem classification label of the complaint text. The complaint problem classification label may reflect the problem of the product or service complained by the user.

[0055] Among the 5,627 complaint texts, there were 151 complaints with categorized labels. Among them, complaints about "compensation standards" accounted for the largest proportion, at 13.1%, reflecting passengers' dissatisfaction with no compensation or insufficient compensation. Secondly, complaints about "refund rules" accounted for 9.29%, reflecting passengers' belief that refund fees were too high, and refund rules were unreasonable or non-compliant. In addition, complaints about "wrong purchases", "refunds", "ticket prices", and "refunds due to illness" accounted for 5%-7%, and the remaining complaints accounted for less than 5%.

[0056] refer to Figure 2 and Figure 5 As an option, the distribution characteristics of the complaint text may also include the word frequency distribution of the word segments of the complaint text.

[0057] Specifically, after some implementations, the user complaint text processing method 1000 may include preprocessing the complaint text, which may include: deleting punctuation marks, numbers, modal particles and sparse words in the complaint text to obtain valid data in the complaint text, wherein sparse words are words whose probability of appearing in the complaint text is less than a second threshold; and segmenting the valid data to obtain word segments of the complaint text.

[0058] Figure 6 It is a word segmentation table of the preprocessed complaint text according to an implementation mode of the present application.

[0059] Combination Figure 2 , Figure 5 and Figure 6 , you can use the word segmentation tool to segment the valid data formed after the deletion process and generate the word segmentation corresponding to the complaint text. Based on the frequency of occurrence of multiple word segments, the word frequency distribution of the word segmentation of the complaint text can be determined.

[0060] For example, the jieba module in the Python programming language can be used to segment complaints and analyze the word frequency distribution of complaint texts. In addition, words with relatively high frequency of occurrence can be selected to draw a word frequency distribution cloud map.

[0061] In addition, for multiple complaint texts, such as the 5,627 complaint texts mentioned above, preprocessing may include: deleting punctuation marks in multiple complaint texts, deleting numbers in multiple complaint texts; deleting sparse words; deleting modal particles; and segmenting the valid data formed after the deletion process.

[0062] As an option, the punctuation marks that can be deleted may include, but are not limited to, [! "# $% &' () * +, -. / : ≤>?@ [\] ^ _ '{|}∼]. In addition, since the numbers mentioned in the complaint text, such as telephone numbers, ticket numbers, mailbox numbers, etc., are not needed, such numbers can be deleted from the complaint text. In addition, sparse words including names of people, countries, cities, airport names, etc. can also be deleted from the complaint text. Optionally, a second threshold can be set to represent the minimum frequency of sparse words in the complaint text. When the probability of a word appearing is less than the second threshold, it can be identified as a sparse word and deleted. In addition, modal particles commonly used in the language can also be deleted from the complaint text because they do not convey important semantic content.

[0063] Therefore, by determining the distribution characteristics of complaint problem classification labels, complaint category distribution or word frequency distribution of word segments of multiple complaint texts, it is possible to identify the problems with the products or services complained by users, or the potential needs of users.

[0064] For example, during the period from May to June 2023 and before, the civil aviation industry was severely impacted by the coronavirus, and passengers' travel mentality was difficult to recover in a short period of time. Airlines urgently need to improve service management to enhance passengers' travel confidence. The complaint texts of public air transport passengers can reflect the service problems encountered by passengers during the corresponding time period. Therefore, it is necessary to conduct a more in-depth study on the complaint texts of public air transport passengers.

[0065] To this end, the complaint texts of public air transport passengers can be processed through the user complaint text processing method 1000, and based on the results of the complaint text processing method 1000, the causes of specific service problems can be analyzed and management suggestions and improvement strategies can be provided to airlines and airports, which can serve as a reference guide for airlines and airports to prevent, organize, summarize and resolve popular complaint issues.

[0066] Step S2 Figure 7 This is a distribution diagram of the sentiment polarity scores of multiple complaint texts according to an implementation mode of the present application. The vertical axis is bidirectional. The left axis represents the score interval [0,0.5] of the negative sentiment score, and the right axis represents the score interval [0,1] of the positive sentiment score. The horizontal axis represents the index number of the complaint text, totaling 5627 texts.

[0067] Figure 8This is a distribution diagram of the sentiment polarity scores of multiple complaint texts with high negative sentiment intensity according to an implementation method of the present application. The horizontal axis represents the index number of the negative complaint text, and the sequence numbers of the screened negative complaint texts are reordered starting from 1. The vertical axis represents the score interval [0, 0.000001] of the negative sentiment score.

[0068] like Figure 7 As shown, in some embodiments of the present application, in order to make the results of the topic model more representative, before performing topic modeling on the complaint text, sentiment analysis can be performed on multiple complaint texts based on a sentiment dictionary or machine learning to determine the complaint texts with high negative sentiment intensity among the multiple complaint texts; and topic modeling can be performed on the complaint texts with high negative sentiment intensity.

[0069] Specifically, sentiment analysis can be understood as the process of measuring people's feelings by extracting text, video or emoticons. Sentiment analysis may include sentiment polarity and sentiment intensity, where sentiment polarity may include positive and negative, positive may be understood as positive emotion, and negative may be understood as negative emotion; sentiment intensity may represent the degree of positivity of positive emotion and the degree of negativity of negative emotion. Through the above two parameters of sentiment analysis, the user's feelings can be measured and the user's subjective inner evaluation tendency of a product or service can be judged. Optionally, the sentiment polarity score can be used to represent the two parameters of sentiment polarity and sentiment intensity.

[0070] As an option, the sentiment polarity scores of multiple complaint texts can be determined based on the sentiment lexicon approach, e.g. Figure 7 The negative sentiment scores and positive sentiment scores are shown, where Figure 7 The horizontal coordinate of is the index number of the complaint text. For example, the complaint text includes the 5627 complaint texts mentioned above. Figure 7 The vertical coordinates include a first coordinate and a second coordinate that are relative to each other in the horizontal direction, the first coordinate is a score interval of a negative sentiment score, for example, its numerical distribution may be 0-0.5, and the second coordinate is a score interval of a positive sentiment score, for example, its numerical distribution may be 0-1.

[0071] refer to Figure 7 and Figure 8 , the sentiment polarity score represents the probability of semantic positivity, which is distributed between [0,1], with 0.5 as the boundary to distinguish positive and negative emotions. The closer the sentiment polarity score is to 1, the more positive the sentiment expression is; the closer the sentiment polarity score is to 0, the more negative the sentiment expression is. Therefore, if the sentiment polarity score is higher than 0.5, the corresponding emotion can be regarded as positive; otherwise, it can be regarded as negative.

[0072] After determining the sentiment polarity scores of multiple complaint texts, the texts with sentiment polarity scores less than the first threshold in the multiple complaint texts can be determined as complaint texts with high negative sentiment intensity. Among the 5627 complaint texts, the sentiment polarity scores of most complaint texts are less than 0.5, indicating negative sentiment, among which Figure 7 The negative sentiment score shown is represented by a vertical line and a circle connected to the vertical line and filled with a diagonal pattern. The negative sentiment score is represented by a circle filled with white. In addition, through statistical calculations, complaint texts with negative emotions account for 99% of the total complaint texts. Since the sentiment polarity scores of most complaint texts are usually low, the first threshold value can be set to, for example, 0.000001 for the 5627 complaint texts. Based on the first threshold value, 4097 complaint texts with high negative sentiment intensity can be obtained from the 5627 complaint texts. Topic modeling of these complaint texts with high negative sentiment intensity can make the extraction results of the topic model more representative. It should be noted that the value of the first threshold value can be determined according to actual needs, and this application does not limit this.

[0073] Figure 8 The distribution of sentiment polarity scores of multiple complaint texts with high negative sentiment intensity whose negative sentiment intensity is less than the first threshold is shown, wherein Figure 8 The horizontal coordinate of is the index number of the complaint text. For example, the complaint text includes the 4097 complaint texts mentioned above. Figure 8 The vertical coordinate of is the score interval of the negative sentiment score, for example, its value distribution is from zero to the first threshold (0-0.000001).

[0074] Furthermore, the SnowNLP module can be used to determine the sentiment polarity scores of multiple complaint texts. SnowNLP is a Chinese sentiment analysis library programmed in Python, which includes Chinese positive and negative sentiment training sets, and can use the Naive Bayes principle to train and predict data. Therefore, in the process of determining the sentiment polarity scores of multiple complaint texts based on the sentiment dictionary method, the part-of-speech tagging, sentiment analysis, keyword extraction, and text summarization functions of the SnowNLP module can be used to determine the sentiment polarity scores of multiple complaint texts. The analysis method for determining the sentiment polarity scores of multiple complaint texts based on the sentiment dictionary is simple to operate, and in addition, the sentiment vocabulary can be expanded according to the text to improve the accuracy.

[0075] As another option, in order to improve the precision and accuracy of sentiment analysis, the sentiment polarity scores of multiple complaint texts can be determined based on machine learning methods. For example, manually annotated corpus can be used as a training set to pre-train a sentiment polarity strength classification model, and the trained model can be used to output the sentiment polarity scores of multiple complaint texts, where the sentiment polarity strength classification model can include a BERT model (Bidirectional Encoder Representations from Transformers, a bidirectional encoder representation model based on Transformer).

[0076] Fig. 9 It is a working process diagram of an unsupervised probability model according to one embodiment of the present application.

[0077] like Fig. 9 As shown, in some embodiments of the present application, step S2 performs topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text, which may include: using an unsupervised probability model to perform topic modeling on the complaint text to obtain the topic and topic keywords, wherein the unsupervised probability model may include a latent Dirichlet allocation topic model (Latent Dirichlet Allocation, LDA).

[0078] Specifically, Theme Modeling (TM) is a probability-based statistical model used to find the semantic topic (hereinafter referred to as topic) that best describes the entire text content and the information of words closely associated with each semantic topic (hereinafter referred to as topic keywords). As an option, some embodiments of the present application may use an unsupervised probability model to perform topic modeling on the complaint text. Taking LDA as an example, the unsupervised probability model is simple to operate, intuitive, scalable and interpretable. In addition, the data set required for the unsupervised probability model is an unsupervised data set, which can be understood as an unlabeled data set. In other words, unsupervised data does not have clear annotations or labels, so compared to supervised data sets, unsupervised data sets are easier to collect and their collection costs are lower. Therefore, the operation of unsupervised probability models using unsupervised data sets is also simpler and less costly.

[0079] In addition, in some embodiments of the present application, in order to simplify the topic modeling process, the text data of the complaint text can be converted into word vectors based on the Bag of Words (BOW) model, and the word vectors are input into the above-mentioned unsupervised probability model to perform topic modeling. In addition, in order to improve the accuracy of topic modeling, for example, the term frequency-inverse document frequency (TF-IDF) can be used to assign weight parameters of different sizes to word vectors with different occurrence frequencies, wherein the higher the occurrence frequency, the smaller the weight parameter of the word vector.

[0080] In addition, in some embodiments of the present application, topic modeling of the complaint text to obtain the topic and topic keywords of the complaint text may include: using at least one of a statistical method and an expert method to determine the number of topics for topic modeling. The determination of the number of topics will affect the semantic coherence and interpretability of the final result. As an option, in the process of determining the number of topics for topic modeling using statistical methods, the number of topics can be determined by relying on the calculation of specific indicators, including indicators such as topic consistency (coherence) and perplexity (perplexity). As another option, in the process of determining the number of topics for topic modeling using expert methods, human intuition can be relied upon to examine the text and select the correct number of topics. As another option, the above two schemes can also be combined to use statistical methods and expert methods at the same time to determine the number of topics for topic modeling.

[0081] The following will describe the working process of topic modeling by combining the above two solutions and using statistical methods and expert methods to determine the number of topics for topic modeling. The number of topics obtained by combining the above two solutions can improve the semantic coherence and interpretability of the topic modeling results. In addition, in the process of determining the number of topics, the optimal topic can also be determined.

[0082] Fig.10 It is a distribution diagram of the consistency scores of the models under different subject settings according to one embodiment of the present application.

[0083] Specifically, Fig. 9 and Fig.10 As shown, the expected number of topics K (hereinafter referred to as the number of initial topics) and the TopN keywords N in the topic can be determined by statistical indicator method and expert experience. As an option, the number of initial topics can be a number in the range of 50-200.

[0084] In some embodiments of the application, a plurality of different numbers of initialization topics are determined using at least one of a statistical method and an expert method; topic modeling is performed on the complaint text based on the initialization topics, and a word group corresponding to the initialization topic is obtained; a consistency score of the initialization topic is determined based on the semantic similarity values ​​between multiple words with relatively high frequencies of occurrence in the word group; and the number of initialization topics corresponding to the highest initialization topic consistency score is determined as the number of topics for topic modeling.

[0085] For example, statistical methods and expert methods can be used to determine that the number of initialization topics is a number between 1 and 200. Optionally, the number of initialization topics can be, for example, 1, 7, 13, 19, 43, 55, 85, 121, 145, 175, 199, etc., where Fig.10 The horizontal coordinates of are the different numbers of initialization topics, Fig.10 The vertical coordinate of is the topic consistency score corresponding to different numbers of initialized topics.

[0086] After selecting and determining a number of different initialization topics, topic modeling can be performed on the complaint text based on the initialization topic, and a word group corresponding to the initialization topic can be obtained; based on the semantic similarity values ​​between multiple words with relatively high frequencies of occurrence in the word group, the consistency score of the initialization topic is determined, and the number of initialization topics corresponding to the highest initialization topic consistency score is determined as the number of topics for topic modeling. Taking the above 4097 complaint texts with high negative sentiment intensity as an example, the number of topics with the highest initialization topic consistency is 78, and the topic consistency score is 0.517. Therefore, the number of topics that can be selected for topic modeling is 78, and the topics under this number of topics are the optimal topics. In addition, based on business experience, the number of topic keywords representing the topic can be set to 3.

[0087] Reference again Fig. 9 Taking LDA as an example, the working process of the unsupervised probability model may include: sampling from the prior Dirichlet distribution to generate the topic distribution of the document; sampling from the topic distribution to generate the topic of the th word of the text; sampling from the Dirichlet distribution of the vocabulary to generate the word distribution of the topic; sampling from the word distribution to finally generate words.

[0088] In other words, based on the multiple optimal topics corresponding to the optimal number of topics mentioned above, the complaint text can be modeled and word groups corresponding to the topics can be obtained, where each topic can correspond to a different word group. LDA can calculate the prior distribution P(w|) of any word n ​​in the word group under its corresponding topic and the prior distribution P(|m) of the topic in the text m. TextRank (a graphical model for text summarization and keyword extraction) is performed once for each topic of the text, and the topic distribution of the word is used as the restart probability in the iterative formula of the PageRank algorithm, so that the PageRank related to each topic will prefer words in the word group that have a greater correlation with the topic. Finally, the score of each word under its topic multiplied by the weight of the topic is the final score of the word. Optionally, each topic can select 3 representative words from the top 30 words with the most prominent final scores as the topic keywords of the topic.

[0089] Fig.11 It is a visualization distribution diagram of 78 topics according to an implementation manner of the present application and an appearance frequency arrangement diagram of each word in the word group corresponding to topic No. 1. Fig.12 This is a topic keyword table of some topics, with a total number of 78 topics, according to one implementation of the present application. Fig.13 This is a topic distribution diagram with 78 topics according to an implementation method of the present application, and the topic distribution of multiple (all) complaint texts. Fig.14 This is a topic distribution diagram of a complaint text with a sentiment polarity score of 0 according to an implementation of the present application. Fig.13 , 14 In the figure, the horizontal axis represents the characteristic words of different topics, and the vertical axis represents the number of times the complaint text of the topic appears.

[0090] Fig.11 The left side of shows the distribution of 78 topics, which are evenly distributed with only a few overlapping topics. Therefore, the extracted user complaint topics cover a wide range and cover all types of services or products involved. The number of topics and topics determined based on this implementation method can effectively and comprehensively reflect the urgent issues related to product attributes or service quality involved in user complaints, as well as changes in user preferences.

[0091] Fig.11 The right side of shows the frequency of occurrence of each word in the word group corresponding to the No. 1 topic selected on the left. The light gray bar shows the overall frequency of occurrence of the word in the word group, and the dark gray bar shows the estimated frequency of occurrence of the word in the word group.

[0092] Combination Fig.11 and Fig.12, each subject can select the most frequently appearing word as the theme keyword of the topic. For example, taking the above 4097 complaint texts with high negative sentiment intensity as an example, the frequency of the topics represented by the theme keywords "submission, sick leave, materials" is much higher than other topics, and the themes represented by the theme keywords "guests, flight attendants, meals", "handling fees, collection, high amount", "delay, take-off, weather", "consumers, behavior, amount" and so on also appear frequently.

[0093] In addition, combined Fig.13 and Fig.14 , Fig.13 and Fig.14 The horizontal coordinates of are all different topic keywords. Fig.13 and Fig.14 The vertical coordinates of are the number of occurrences of different topic keywords. For example, the number of occurrences of the topic keyword "submit, sick leave, materials" is between 300-350. For the 4097 complaint texts with high negative sentiment intensity, the subjects represented by the topic keywords "submit, sick leave, materials", "delay, take-off, weather", "name, change, return", "price, price difference, fare" and "personnel, attitude, economy class" appear most frequently, so the topics represented by these topic keywords effectively and comprehensively reflect the urgent problems related to product attributes or service quality involved in user complaints, as well as changes in user preferences. In the topic distribution diagram of complaint texts with a sentiment polarity score of 0, the subjects represented by the topic keywords "consumer, behavior, amount", "submit, case, materials", "regulations, changes, conditions" and "luggage, compensation, consignment" appear most frequently, reflecting that users are more likely to have complaint texts with relatively high negative sentiment intensity when they are sick or suffer financial losses, and accordingly, users are more concerned about the improvement of products or services in these situations.

[0094] Step S3 Fig.15 It is a flowchart for determining the association rules between topic keywords and complaint problem classification labels according to one implementation of the present application. Fig.16 It is a list of some initial rules with higher confidence in the initial rule set according to one embodiment of the present application. Fig.17 It is a list of some initial rules with higher support in the initial rule set according to one implementation of the present application. Fig.18 It is a list of some initial rules with higher improvement in the initial rule set according to one implementation of the present application.

[0095] refer to Figure 15-17In some embodiments of the present application, step S3 determines the association rules between the subject keywords and the complaint problem classification labels, including: matching the subject keywords and the complaint problem classification labels to form an item set; processing the item set based on the association rule algorithm to obtain an initial rule set; and determining the association rules from the initial rule set based on the screening rules.

[0096] Specifically, as an option, association rules can be derived from ARM (Association Rule Mining), which is a part of Machine Learning (ML) and can be used to discover the correlation and interdependence between a large number of data items. ARM represents the above relationships between data items and intuitively determines how and why certain data items are connected through association rules.

[0097] Taking the complaint text of public air transport passengers as an example, passenger complaints can be regarded as a group of events. Using ARM, its causal factors can be analyzed, making it easier for airlines and airports to manage and prevent their services or products.

[0098] The association rules between topic keywords and complaint classification labels derived by using ARM may include: after forming the item set, determining the initial rule set based on the predetermined minimum support and the predetermined minimum confidence; and determining the association rules from the initial rule set based on the screening rules. Among them, the association rule algorithm may include the Apriori algorithm, and the Apriori algorithm may obtain frequent item sets layer by layer in a layer-by-layer iterative manner (which can be understood as the initial rule set).

[0099] refer to Fig.15 As shown in the flowchart of determining association rules, it is assumed that the transaction set for association rule mining is denoted as N. , the data item set is recorded as .exist , Under the condition that X⇒Y, an association rule is generated.

[0100] Optionally, determining the association rule from the initial rule set may include: determining the confidence, support and lift of each of the initial rule sets; and determining the association rule based on a minimum confidence threshold (min_Confidence), a minimum support threshold (min_Support) and a minimum lift threshold (min_Lift).

[0101] In other words, the association rules are subject to the minimum threshold of support, the minimum threshold of confidence, and the minimum threshold of lift. Among them, the support can represent the frequency of the rule, the ratio of the number of transactions X and Y contained in the transaction set to the number of all transactions, and the support can be recorded as: ; The confidence level can represent the strength of the rule, which is the ratio of the number of transactions containing X and Y to the number of transactions containing X in the transaction set. The confidence level can be expressed as: ; In addition to support and confidence, lift is also a measure of the intrinsic value of a rule using correlation analysis. Lift describes the influence of X on Y. Lift can be written as: ; When the Lift value is 1, it indicates that X and Y are independent of each other; if the value is less than 1, it indicates that the two transactions are mutually exclusive; and when the Lift value is greater than 3, the mined association rules are effective strong association rules.

[0102] Continue to refer Figure 15-18 Taking the above 4097 complaint texts with high negative sentiment intensity as an example, the initial rule set between the topic keywords and the complaint problem classification labels can be determined based on the predetermined minimum support and the predetermined minimum confidence. As an option, the predetermined minimum support can be 0.005, the predetermined minimum confidence can be 0.05, and the initial rule set includes 44 rules, among which Figure 16-18 The lhs (left hand side) represents the antecedent of the rule, and the rhs (righthand side) represents the consequent of the rule. Fig.16 The five rules with the highest confidence are shown; Fig.17 The five rules with the highest support are shown; Fig.18 The five rules with the highest lift are shown.

[0103] Based on the screening rules, association rules can be determined from the initial rule set. For example, after determining the confidence, support and lift of each rule in the initial rule set, the association rules are determined based on the minimum confidence threshold, the minimum support threshold and the minimum lift threshold.

[0104] Step S4 Step S4 improves the products or services corresponding to the complaint problem classification tags based on the topic keywords for which association rules have been determined with the complaint problem classification tags.

[0105] Specifically, taking 5627 passenger complaint texts with a certain city as the departure point and destination as an example, through steps S1 to S3 of the user complaint text processing method 1000, it can be seen that the subject represented by the subject keywords "submit, sick refund, materials" belongs to the category of ticketing services, and the extracted association rule {sick refund}⇒{submit sick refund materials} also reveals that most passenger complaints under this subject are refund disputes caused by passengers being unable to take flights due to illness of themselves or their family members, most of which mentioned that they were unable to take the flight due to physical discomfort caused by the coronavirus. The main contradictions of this type of complaint include: some passengers choose to voluntarily refund their tickets because they do not know that there are channels for sick refunds, resulting in a large amount of handling fees being deducted; passengers submit sick proof materials that are not approved; they are refused sick refunds because they do not meet the refund material requirements; or the corresponding handling fees are still deducted after the approval, etc.

[0106] Airlines stipulate that if a full refund is required, it must be requested before departure and a diagnosis certificate, invoice, etc. must be issued. The refund material requirements are relatively strict. Therefore, before applying for a refund, airlines can strengthen information notification and add channels and material requirements for sick refunds to the refund and change instructions on the ticket interface. In addition, after the passenger successfully buys the ticket, a text message can be sent to remind the passenger of important information such as the departure time and refund and change instructions. In the process of applying for a refund, airlines can also provide passengers with templates of corresponding materials so that passengers can refer to them when providing materials. After the passenger applies for a refund, the airline can speed up the review to prevent delays. In addition, airlines can also increase investment in digital review, and after the passenger submits the refund materials, they can immediately know whether the submitted refund materials meet the specifications.

[0107] Therefore, according to the user complaint text processing method provided by at least one embodiment of the present application, the theme and theme keywords of the complaint text can be obtained by performing topic modeling on the user's complaint text, and by establishing association rules between theme keywords and complaint problem classification labels, an intuitive connection can be established between the complaint text and product improvements and service improvements, so that products or services corresponding to the complaint problem classification labels can be improved accurately, promptly and efficiently based on the theme keywords in a targeted manner.

[0108] Fig.19 It is a schematic diagram of the framework of a user complaint text processing device 2000 according to one embodiment of the present application.

[0109] like Fig.19As shown, another aspect of the present application further provides a user complaint text processing device 2000. The user complaint text processing device 2000 includes: a distribution feature acquisition module 100, a modeling module 200, an association rule determination module 300 and an improvement module 400. The distribution feature acquisition module 100 is configured to determine the distribution features of the complaint text, wherein the distribution features of the complaint text include complaint problem classification labels. The modeling module 200 is configured to perform topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text. The association rule determination module 300 is configured to determine the association rules between the topic keywords and the complaint problem classification labels. The improvement module 400 is configured to improve the products or services corresponding to the complaint problem classification labels based on the topic keywords for which the association rules have been determined with the complaint problem classification labels.

[0110] Therefore, according to the user complaint text processing device provided by at least one embodiment of the present application, the theme and theme keywords of the complaint text can be obtained by performing topic modeling on the user's complaint text, and by establishing association rules between theme keywords and complaint problem classification labels, an intuitive connection can be established between the complaint text and product improvements and service improvements, so that products or services corresponding to the complaint problem classification labels can be improved accurately, timely and efficiently based on the theme keywords.

[0111] Specifically, in one embodiment of the present application, the modeling module 200 is further configured to use an unsupervised probability model to perform topic modeling on the complaint text to obtain topics and topic keywords, wherein the unsupervised probability model includes a latent Dirichlet distribution topic model.

[0112] The unsupervised probability model is used to perform topic modeling on the complaint text. Taking LDA as an example, the unsupervised probability model is simple, intuitive, scalable and interpretable. In addition, the dataset required for the unsupervised probability model is an unsupervised dataset, which can be understood as an unlabeled dataset. In other words, unsupervised data has no clear annotations or labels, so it is easier to collect than supervised datasets, and its collection cost is also lower. Therefore, the operation of the unsupervised probability model using the unsupervised dataset is also simpler and the cost is lower.

[0113] In addition, the distribution characteristics of the complaint text also include: complaint category distribution and word frequency distribution of the segmented words.

[0114] In some embodiments of the present application, the device further includes a sentiment analysis module (not shown), which is configured to perform sentiment analysis on multiple complaint texts based on a sentiment dictionary or machine learning to determine complaint texts with high negative sentiment intensity among the multiple complaint texts; and the modeling module 200 is further configured to perform topic modeling on complaint texts with high negative sentiment intensity. Performing topic modeling on complaint texts with high negative sentiment intensity can make the extraction results of the topic model more representative. It should be noted that the value of the first threshold can be determined according to actual needs, and this application does not limit this.

[0115] In addition, sentiment analysis includes sentiment polarity and sentiment intensity, sentiment polarity includes positive and negative, and the sentiment analysis module is further configured to determine the sentiment polarity scores of multiple complaint texts based on a sentiment dictionary or machine learning; and determine the texts in the multiple complaint texts whose sentiment polarity scores are less than a first threshold as complaint texts with high negative sentiment intensity.

[0116] In some embodiments of the present application, the modeling module 200 is further configured to determine the number of topics for topic modeling using at least one of a statistical method and an expert method.

[0117] When the number of topics for topic modeling is determined by combining statistical methods and expert methods, the number of topics obtained by combining the two schemes can improve the semantic coherence and interpretability of the topic modeling results. In addition, in the process of determining the number of topics, the optimal topic can also be determined.

[0118] For example, the modeling module 200 is further configured to determine a plurality of different numbers of initialization topics using at least one of a statistical method and an expert method; perform topic modeling on the complaint text based on the initialization topics and obtain a word group corresponding to the initialization topic; determine a consistency score of the initialization topic based on the semantic similarity values ​​between multiple words with relatively high frequencies of occurrence in the word group; and determine the number of initialization topics corresponding to the highest consistency score as the number of topics for topic modeling.

[0119] Optionally, to simplify the topic modeling process, the modeling module 200 is further configured to convert the text data of the complaint text into word vectors based on the bag-of-words model, and input the word vectors into the above-mentioned unsupervised probability model to perform topic modeling. In addition, to improve the accuracy of topic modeling, the modeling module 200 is further configured to use, for example, word frequency-inverse document frequency to assign weight parameters of different sizes to word vectors with different frequencies of occurrence, wherein the weight parameter of the word vector with a higher frequency of occurrence is smaller.

[0120] In some embodiments of the present application, the association rule determination module 300 is further configured to match topic keywords and complaint problem classification labels to form an item set; based on the association rule algorithm, process the item set to obtain an initial rule set; and based on the screening rules, determine the association rules from the initial rule set.

[0121] Optionally, the association rule determination module 300 is further configured to determine the above-mentioned initial rule set based on a predetermined minimum support and a predetermined minimum confidence. Optionally, the association rule determination module 300 is further configured to determine the confidence, support and lift of each of the initial rule sets; and determine the association rule based on the minimum confidence threshold, the minimum support threshold and the minimum lift threshold.

[0122] Fig. 20 It is a structural diagram of a computer system according to one embodiment of the present application, which is suitable for implementing the electronic device of the embodiment of the present application.

[0123] refer to Fig. 20 , which shows a schematic diagram of the structure of an electronic device 600 suitable for implementing the embodiments of the present application. The terminal device in the embodiments of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Fig. 20 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0124] like Fig. 20 As shown, the electronic device 600 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the electronic device 600 are also stored. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0125] Typically, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device 600 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 8 The electronic device 600 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead. Fig. 20 Each block shown in the figure may represent one device, or may represent multiple devices as required.

[0126] In addition, according to the embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, the above-mentioned functions defined in the method of the embodiment of the present application are executed. It should be noted that the computer-readable medium described in the embodiment of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection with one or more conductors, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In an embodiment of the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device, or device. In an embodiment of the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, device, or device. The program code embodied on the computer readable medium may be transmitted using any appropriate medium, including but not limited to: wire, optical cable, RF (radio frequency), etc., or any suitable combination of the foregoing.

[0127] The computer-readable medium may be included in the electronic device; or it may exist independently without being installed in the electronic device. The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: determines the distribution characteristics of the complaint text, wherein the distribution characteristics of the complaint text include complaint problem classification labels; performs topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text; determines the association rules between topic keywords and complaint problem classification labels; and based on the topic keywords for which the association rules have been determined with the complaint problem classification labels, improves the products or services corresponding to the complaint problem classification labels.

[0128] The computer program code for the operation of the embodiments of the present application can be written in one or more programming languages ​​or a combination thereof, and the programming language includes an object-oriented programming language, such as Java, Smalltalk, C++, and also includes a conventional procedural programming language, such as "C" language or similar programming language. The program code can be executed completely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., using an Internet service provider to connect through the Internet).

[0129] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0130] The units involved in the embodiments described in the present application may be implemented by software or hardware. The units described may also be set in a processor, for example, it may be described as: a processor includes an observation unit, an estimation unit, and a positioning unit. The names of these units do not constitute a limitation on the units themselves in some cases, for example, the positioning unit may also be described as a "positioning object position information unit".

[0131] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present application is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present application.

Claims

1. A method for processing user complaint text, characterized in that: include: Determining distribution characteristics of complaint texts, wherein the distribution characteristics of complaint texts include complaint problem classification labels; Performing topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text; Determine the association rules between the subject keywords and the complaint classification labels; and Based on the subject keywords for which the association rules have been determined with the complaint problem classification labels, the products or services corresponding to the complaint problem classification labels are improved.

2. The method according to claim 1, wherein: Performing topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text includes: using an unsupervised probability model to perform topic modeling on the complaint text to obtain the topic and the topic keywords, wherein the unsupervised probability model includes a latent Dirichlet distribution topic model; and / or, The distribution characteristics of the complaint text also include: complaint category distribution and word frequency distribution of the word segments of the complaint text.

3. The method according to claim 1, wherein: The method further comprises: Based on a sentiment dictionary or machine learning, sentiment analysis is performed on the plurality of complaint texts to determine complaint texts with high negative sentiment intensity among the plurality of complaint texts; and The topic modeling is performed on the complaint text with high negative sentiment intensity.

4. The method according to claim 3, wherein: The sentiment analysis includes sentiment polarity and sentiment intensity, wherein the sentiment polarity includes positive and negative, wherein based on a sentiment dictionary or machine learning, sentiment analysis is performed on the plurality of complaint texts to determine that the complaint texts with high negative sentiment intensity among the plurality of complaint texts include: Determining sentiment polarity scores of the plurality of complaint texts based on the sentiment dictionary or the machine learning; and The texts with the sentiment polarity scores less than the first threshold among the multiple complaint texts are determined as the complaint texts with high negative sentiment intensity.

5. The method according to claim 1, wherein: Performing topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text includes: using at least one of a statistical method and an expert method to determine the number of topics for the topic modeling; but, Determining the number of topics for the topic modeling by using at least one of a statistical method and an expert method includes: Determine a plurality of different numbers of initialization topics by using at least one of a statistical method and an expert method; Performing the topic modeling on the complaint text based on the initialization topic, and obtaining a word group corresponding to the initialization topic; Determining a consistency score of the initialization topic based on semantic similarity values ​​between a plurality of words with relatively high occurrence frequencies in the word group; and The number of initialized topics corresponding to the highest consistency score is determined as the number of topics for the topic modeling.

6. The method according to claim 1, wherein: The method also includes preprocessing the complaint text, the preprocessing including: deleting punctuation marks, numbers, modal particles and sparse words in the complaint text to obtain valid data in the complaint text, wherein the sparse words are words with a probability of appearing in the complaint text less than a second threshold; and segmenting the valid data to obtain word segments of the complaint text; and / or, The topic modeling of the complaint text further includes: based on the bag-of-words model, converting the text data of the complaint text into word vectors, and inputting the word vectors into an unsupervised probability model to perform the topic modeling; and assigning weight parameters of different sizes to word vectors with different frequencies of occurrence, wherein the weight parameter of the word vector with a higher frequency of occurrence is smaller; and / or, Determining the association rules between the subject keywords and the complaint problem classification labels includes: matching the subject keywords and the complaint problem classification labels to form an item set; processing the item set based on an association rule algorithm, and determining the initial rule set based on a predetermined minimum support and a predetermined minimum confidence; and, based on screening rules, determining the association rules from the initial rule set, and determining the confidence, support and lift of each one in the initial rule set; and determining the association rules based on a minimum confidence threshold, a minimum support threshold and a minimum lift threshold.

7. A user complaint text processing device, characterized in that: The device comprises: A distribution feature acquisition module is configured to determine the distribution features of the complaint text, wherein the distribution features of the complaint text include complaint problem classification labels; A modeling module is configured to perform topic modeling on the complaint text to obtain the topic and topic keywords of the complaint text; An association rule determination module, configured to determine an association rule between the subject keyword and the complaint problem classification label; and The improvement module is configured to improve the product or service corresponding to the complaint problem classification label based on the subject keyword for which the association rule has been determined with the complaint problem classification label.

8. An electronic device, characterized in that: include: at least one processor; as well as, A memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor so that the at least one processor can execute the user complaint text processing method described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the user complaint text processing method as described in any one of claims 1-6 is implemented.

10. A computer program product storing a computer program, characterized in that: When the computer program is executed by a processor, the user complaint text processing method as described in any one of claims 1-6 is implemented.

Images