A government affair service hotline platform data governance and analysis method based on big data

Abstract

The present application relates to the technical field of semantic processing, in particular to a government service hotline platform data governance and analysis method based on big data. The present application first acquires communication phrases of historical call data, and screens out emphasized phrases and possible appeal phrases from the historical call data; further, in any two historical call data, the emphasized phrases of different calls are matched based on the similarity between the emphasized phrases; further, according to the consistency of the use difference between all matched emphasized phrases, a theme consistency coefficient is obtained and all historical call data are classified; finally, in any class of historical call data, according to the similarity between the possible appeal phrases, the distribution of the possible appeal phrases is combined, the appeal phrases are screened out and recorded as demand labels, effectively solving the technical problems of inaccurate user demand label classification and large data mining deviation in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semantic processing, and particularly relates to a government service hotline platform data governance and analysis method based on big data. BACKGROUND

[0002] The government service hotline platform collects various feedback data of citizens, but because it integrates many other non-emergency hotline phones, many citizen problems will be reflected through the hotline, so important data needs to be screened from the numerous data, and these data need to be deeply mined and analyzed.

[0003] The prior art is basically to directly use the LDA topic model or graph clustering algorithm to classify and process repeated complaints by using the work order data recorded by the operator when analyzing the work order of the hotline platform, but because the records on the work order are basically simplified, it does not consider the specific complaint information of different users when communicating, and some actual demand situations of users are easily ignored, the demand label classification of users is not accurate, and some errors occur when the existing method analyzes the hotline platform data. SUMMARY

[0004] In order to solve the technical problem that the demand label classification of users is not accurate when analyzing the data of the government service hotline platform, and affecting the deep mining of data, the purpose of the present application is to provide a government service hotline platform data governance and analysis method based on big data, and the technical solution adopted is as follows:

[0005] Obtain the communication word group of the historical call data of the government service hotline platform; filter out the emphasized word group according to the use frequency of the communication word group; filter out the possible complaint word group according to the similarity of the communication word group and the emphasized word group, and combine the use frequency of the emphasized word group;

[0006] In any two historical call data, the emphasized word groups of different calls are matched based on the similarity between the emphasized word groups; a theme consistency coefficient is obtained according to the consistency of the use difference between all matched emphasized word groups; and all historical call data are classified based on the theme consistency coefficient;

[0007] In any one of the historical call data, the complaint word group is filtered out according to the similarity between the possible complaint word groups, and is recorded as a demand label in combination with the distribution of the possible complaint word group.

[0008] Further, after obtaining the demand label, it further includes:

[0009] Obtain the latest call data and extract the possible complaint word group, and obtain the classification label of the latest call data according to the cosine similarity between the possible complaint word group and the demand label.

[0010] Further, the method for obtaining the emphasized word group comprises:

[0011] For any call process, a TF-IDF value of each of the communication word groups is obtained, and a preset first proportion of the communication word groups with the largest TF-IDF value is selected as an emphasized word group.

[0012] Further, the method for obtaining the possible appeal word group comprises:

[0013] According to a cosine similarity between any of the communication word groups and each of the emphasized word groups, and in combination with a TF-IDF value of the emphasized word group, an appeal relationship degree corresponding to the communication word group is obtained.

[0014] The communication word group with an appeal relationship degree greater than a preset appeal threshold value is taken as a possible appeal word group.

[0015] Further, the method for obtaining the theme consistency coefficient comprises:

[0016] In any two call data, a theme consistency coefficient is obtained according to a consistency of a difference in TF-IDF values between all matched emphasized word groups.

[0017] Further, the method for obtaining the appeal word group comprises:

[0018] In any type of the historical call data, according to a cosine similarity between a type of the possible appeal word group and another type of the possible appeal word group, in combination with a theme consistency coefficient between the historical call data in which the two types of the possible appeal word group are located, an appeal similarity between the two types of the possible appeal word group is obtained.

[0019] Based on the appeal similarity, a same-type appeal word group of each type of the possible appeal word group is screened out; according to a number of the historical call data in which each type of the possible appeal word group exists, in combination with a number of the historical call data in which a corresponding same-type appeal word group exists, and a type of the corresponding same-type appeal word group, an appeal representative degree is obtained.

[0020] Based on the appeal representative degree, an appeal word group is screened out.

[0021] Further, the method for screening out the appeal word group based on the appeal representative degree comprises:

[0022] A preset second proportion of the possible appeal word groups with the largest appeal representative degree is selected as an appeal word group.

[0023] Further, the method for obtaining the communication word group comprises:

[0024] The historical call data is converted into text and segmented into words. At the same time, the Chinese Dictionary and local dialect dictionaries are called to remove stop words, and the remaining word groups are used as communication word groups.

[0025] Furthermore, the method for matching the emphasized word groups in different calls based on the similarity between the emphasized word groups includes:

[0026] The emphasized word group in the call data is matched with the emphasized word group in another call data that has the highest cosine similarity.

[0027] Furthermore, the method for classifying all historical call data based on the subject consistency coefficient includes:

[0028] All historical call data were clustered and classified using the DBSCAN clustering algorithm.

[0029] The present invention has the following beneficial effects:

[0030] This invention first obtains communication phrases from historical call data to provide an analytical foundation; it then further filters out emphasized phrases to highlight the information the user is focusing on, and identifies potential appeal phrases to narrow down the analysis scope; it further matches emphasized phrases from different calls to analyze differences in word choice and usage habits among different users, identifying two words from two calls describing the same thing; it further obtains a thematic consistency coefficient based on the consistency of usage differences among all matched emphasized phrases, characterizing the degree of thematic consistency between two call data sets; based on the thematic consistency coefficient, similar historical call data are grouped together to compensate for expression biases caused by word differences and improve the accuracy of tag extraction; finally, appeal phrases are selected and recorded as demand tags to extract common demands, accurately representing user needs and providing a foundation for in-depth data mining. This invention effectively solves the problems of inaccurate user demand tag classification and large data mining bias in existing technologies by filtering emphasized phrases and extracting potential appeal phrases, classifying them using the thematic consistency coefficient between call data, and then obtaining demand tags based on similarity and distribution. Attached Figure Description

[0031] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart illustrating a data governance and analysis method for a government service hotline platform based on big data, provided as an embodiment of the present invention;

[0033] Figure 2 A flow chart of a method for obtaining a complaint phrase according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to further clarify the technical means and effects of the present application taken to achieve the predetermined object of the present application, the following describes in detail the specific implementation, structure, features and effects of a government service hotline platform data management and analysis method based on big data according to the present application, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0036] The following describes in detail the specific scheme of a government service hotline platform data management and analysis method based on big data according to the present application in combination with the accompanying drawings.

[0037] Please refer to Figure 1 which shows a flow chart of a government service hotline platform data management and analysis method based on big data according to an embodiment of the present application, specifically including:

[0038] Step S1: Obtain the communication phrase of the historical call data of the government service hotline platform; filter out the emphasized phrase according to the usage frequency of the communication phrase; and filter out the possible complaint phrase according to the similarity between the communication phrase and the emphasized phrase, combined with the usage frequency of the emphasized phrase.

[0039] In an embodiment of the present application, the hotline platform is set to automatically record the call recording of each call, and at the same time, collect the relevant text records of the call by the operator during the call. At the same time, according to the telephone number of the user's call and the user's relevant information, the user's region is obtained, etc. The obtained data includes user call records, corresponding work order data, and related information such as involved regions. Among them, the relevant data of the hotline platform in the past 30 days is set, and the historical call data includes call recording and text data recorded by the operator.

[0040] Considering that the call recording cannot be directly analyzed for complaints, the historical call data is first preprocessed, specifically including: converting the historical call data into text and performing word segmentation, and calling the Chinese Dictionary and local dialect dictionary to remove stop words, and the remaining phrase is used as a communication phrase.

[0041] Among them, a plurality of call audios are converted into texts by automatic speech recognition (ASR) technology; the converted texts and recorded texts are segmented by a JieBa segmentation tool, which are technical means known to those skilled in the art and will not be described in detail.

[0042] When a user dials a hotline, the user always has a demand or a complaint in his heart, which makes the user describe the complaint in the description to the call operator in the call, that is, the content similar to the content with a higher frequency of a plurality of word groups in a telephone communication is recorded as a possible complaint in the telephone communication.

[0043] In addition, the frequency of the communication word group represents the emphasized degree in the call, so the emphasized word group is screened according to the frequency of the communication word group, the information expressed by the user is highlighted, the possible complaint word group is screened according to the similarity between the communication word group and the emphasized word group, the word group possibly representing the user's complaint is determined, the analysis range is narrowed, the word group representing the user's complaint is further screened, and the most suitable demand label is constructed.

[0044] Preferably, in an embodiment of the present application, the importance of the word group is shown from the frequency of use by quantifying the local frequency and global scarcity of the word group, so for any call process, the TF-IDF value of each communication word group is obtained, and the preset first proportion of communication word groups with the largest TF-IDF value are selected as emphasized word groups.

[0045] As an example, the preset first proportion is 30%, the TF-IDF values of all communication word groups in any call process are arranged from large to small, and the top 30% of communication word groups are selected as emphasized word groups.

[0046] Preferably, in an embodiment of the present application, the process between the user and the call operator is basically around the complaint of the user, which leads to the fact that the chat content basically has some relevance to the complaint of the user in some aspects, the importance of the emphasized word group is higher, and the chat content in this call can be basically summarized, so when the cosine similarity between the communication word group and the emphasized word group is higher, it means that the similarity between the communication word group and the emphasized word group is higher, and it is more likely to be a complaint word group.

[0047] At the same time, the higher the TF-IDF value of the emphasized word group itself, the more critical the emphasized word group, and the higher the reference degree, so the complaint relationship degree of the corresponding communication word group is obtained according to the cosine similarity between any communication word group and each emphasized word group, and the TF-IDF value of the emphasized word group.

[0048] The communication word group with a persuasion relationship degree greater than a preset persuasion threshold is taken as a possible persuasion word group.

[0049] As an example, in any call process, any communication word group is selected as a target communication word group, the cosine similarity between the target communication word group and the emphasized word group is obtained, and a constant 1 is added for adjustment as a first similarity parameter, which represents the similarity between the communication word group and the emphasized word group. The higher the first similarity parameter is, the higher the similarity is.

[0050] The product of all the first similarity parameters of the target communication word group and the TF-IDF values of the corresponding emphasized word groups is taken as the numerator, the sum of the TF-IDF values of all the corresponding emphasized word groups corresponding to the target communication word group is taken as the denominator, the ratio of the numerator to the denominator is taken as the independent variable, the mapping value obtained by the softsign(x) function is taken as the persuasion relationship degree of the target communication word group; x is the independent variable.

[0051] The preset persuasion threshold is set to 0.6, and the possible persuasion word group is screened out.

[0052] It should be noted that one call process corresponds to one historical call data, and the processing process of all call processes is the same, and only one example is described here; before calculating the cosine similarity, the Skip-gram model in the bag-of-words model Word2Vec is used to construct the word vector corresponding to each word group, and the cosine similarity and the TF-IDF value have been prior art. In other embodiments of the present application, the implementer can adjust the preset first proportion and the preset persuasion threshold.

[0053] It should be noted that the emphasized word group may also be a persuasion word group, so when screening the possible emphasized word group, the communication word group determined as the emphasized word group also participates in the screening analysis.

[0054] Step S2: In any two historical call data, the emphasized word groups of different calls are matched based on the similarity between the emphasized word groups; the consistency coefficient of the main idea is obtained according to the consistency of the use difference between all matched emphasized word groups; and all historical call data are classified based on the consistency coefficient of the main idea.

[0055] It should be noted that the possible persuasion word groups screened out may not necessarily be the real persuasion, that is, the user may not clearly express the real need for persuasion, but instead describes a series of the persuasion. However, in the city, there are some users who can use simple language to describe the same persuasion. Therefore, all hotlines can be investigated, similar communication processes can be compared, and finally the persuasion labels corresponding to each communication process can be determined according to the similar communication processes.

[0056] Since different users have different differences in word choice, word usage, etc., the emphasized phrases of different calls are matched based on the similarity between the emphasized phrases in any two historical call data, and two words describing the same thing are screened out; then, the consistency coefficient of the main idea is obtained according to the consistency of the usage difference between all matched emphasized phrases, representing the consistency of the main idea of the two call data; then, all call data are classified based on the consistency coefficient of the main idea, to make up for the expression deviation caused by the word difference, improve the accuracy of label extraction, and facilitate the final determination of the appropriate demand label.

[0057] It should be noted that the matching of the emphasized phrases of any two historical call data and the analysis process of the consistency coefficient of the main idea are the same, and only one example is described here, and the description is not repeated.

[0058] In an embodiment of the present application, the similarity between the emphasized phrases is represented by cosine similarity, and the emphasized phrases of the call data are matched with the emphasized phrases with the largest cosine similarity in another call data. For example, the emphasized phrase a1 in the call data A, the cosine similarity between a1 and each emphasized phrase in the call data B is calculated, and the largest emphasized phrase b20 is selected for matching.

[0059] Preferably, in an embodiment of the present application, considering that users with the same appeal mostly have similar experiences, which leads to a high similarity between the cause and effect described by the user in the communication process with the call operator, that is, when describing the same event, similar phrases and sentences are used, and the difference is only in the number of words and the number of emphasized times when describing, which makes the corresponding many emphasized phrases and the proportion between these phrases similar in the communication process of users with the same appeal, so that the difference between the TF-IDF values of the matched emphasized phrases has strong consistency.

[0060] Based on this, in any two call data, the consistency coefficient of the main idea is obtained according to the consistency of the difference between the TF-IDF values of all matched emphasized phrases.

[0061] As an example, in any two call data, the difference between the TF-IDF value of each emphasized phrase of one call data and the TF-IDF value of the matched emphasized phrase is taken as the numerator, and the value is taken as the denominator, the ratio of the fraction is taken as the relative deviation factor, the standard deviation of the relative deviation factor is taken as the independent variable, and the negative correlation mapping is carried out through the exp(-x) function, and the mapping value is taken as the consistency coefficient of the main idea of the corresponding two call data, x represents the independent variable, and exp(-x) is the exponential function with natural constant e as the base.

[0062] The relative deviation factor comprises matching pairs of all emphasized phrases of the two call data; the TF-IDF value represents the use characteristics of the emphasized phrases, and the standard deviation represents the consistency of the use difference; the smaller the standard deviation, the stronger the consistency of the use difference, the closer the TF-IDF values between the emphasized phrases, the more similar the proportions of the two words at the same time, the more likely to have the same appeal, and the greater the theme consistency coefficient; since the standard deviation of a group of data after negation is unchanged, the emphasized phrases corresponding to the minuend in the numerator belong to the call data, which can be any one of the two call data, for example, call data A and call data B, and the emphasized phrases of any call data such as A are taken as the minuend.

[0063] In other embodiments of the application, the implementer can also take the average value of the absolute value of the relative deviation factor as the independent variable to represent the overall difference in the use of the emphasized phrases expressing the same event, and after negative correlation mapping by the exp(-x) function, the mapping value is taken as the theme consistency coefficient corresponding to the two call data.

[0064] Preferably, in an embodiment of the application, after obtaining the theme consistency coefficients between a plurality of historical call data, all historical call data are clustered and classified by the DBSCAN clustering algorithm.

[0065] Specifically, the greater the theme consistency coefficient, the closer the user appeal, and the smaller the clustering distance, so the reciprocal of the theme consistency coefficient of the two call data is taken as the clustering distance during clustering, the Eps neighborhood radius is set to the average value of the reciprocal of the theme consistency coefficient between all call data, the preset MinPts minimum point number is 5, all call data are clustered and classified by the DBSCAN clustering algorithm, a plurality of clustering clusters are obtained, and a plurality of call data groups with similar communication content are obtained; the DBSCAN clustering algorithm is a prior art and will not be described again.

[0066] Step S3: In any type of historical call data, according to the similarity between the possible appeal phrases, the possible appeal phrases are screened out in combination with the distribution of the possible appeal phrases and are marked as demand labels.

[0067] When a user dials a hotline to make an appeal or inquiry, he / she will describe the related event and the specific demand around the appeal, so when the specific demand and the event described by the user are basically similar, it can be considered that the user appeal is similar, and then there will be some possible appeal phrases that are generally similar to other possible appeal phrases, which can represent the user appeal;

[0068] Meanwhile, the distribution of the possible appeal word group reflects the general degree of the user's expression of appeal and the representative degree of the word group. Therefore, in any type of historical call data, the appeal word group is screened out as a demand label according to the similarity between the possible appeal word groups and in combination with the distribution of the possible appeal word groups. The word group that best represents the appeal in the historical call data is selected to extract the common appeal and accurately represent the user demand, thereby providing a basis for deep data mining.

[0069] It should be noted that the analysis process is the same for each type of call data, and only any one of them is described here.

[0070] Preferably, in an embodiment of the present application, please refer to Figure 2 which shows a flowchart of an appeal word group acquisition method provided by an embodiment of the present application, and specifically includes:

[0071] Step S301: In any type of historical call data, the appeal similarity between two possible appeal word groups is obtained according to the cosine similarity between the two possible appeal word groups and in combination with the theme consistency coefficient between the historical call data where the two word groups are located.

[0072] It is considered that a possible appeal word group can be distributed in multiple historical call data. When comparing the similarity between two possible appeal word groups, different historical call data are compared. The higher the theme consistency coefficient between the historical call data, the more likely the appeal theme of the two calls is consistent, and the higher the reference value of the similarity of the possible appeal word groups. Therefore, the appeal similarity between the possible appeal word groups is obtained to represent the similarity between the possible appeal word groups.

[0073] As an example, the calculation formula of the appeal similarity includes:

[0074]

[0075] wherein v represents the classification serial number of the historical call data, s and m both represent the category serial number of the possible appeal word group, and s≠m; L v,s,m represents the appeal similarity between the s-th possible appeal word group and the m-th possible appeal word group in the v-th historical call data; J v,m represents the number of historical call data containing the m-th possible appeal word group in the v-th historical call data; j represents the serial number of the historical call data containing the m-th possible appeal word group in the v-th historical call data; I v,s represents the number of historical call data containing the s-th possible appeal word group in the v-th historical call data; i represents the serial number of the historical call data containing the s-th possible appeal word group in the v-th historical call data; represents the word vector corresponding to the s-th possible appeal word group; represents the word vector corresponding to the mth possible appeal phrase; represents the cosine similarity between the s th possible appeal phrase and the mth possible appeal phrase; A i,j represents the theme consistency coefficient corresponding to the i th historical call data containing the s th possible appeal phrase and the j th historical call data containing the mth possible appeal phrase.

[0076] In the calculation formula of the appeal similarity, the cosine similarity is weighted and averaged by multiplying the theme consistency coefficient as the weight, to obtain the appeal similarity between the appeal phrases, the higher the appeal similarity, the closer the two phrases in appeal, and the higher the similarity. When calculating the cosine similarity in the same historical call data, the corresponding A i,j is set to 1.

[0077] In other embodiments of the present application, the implementer can also measure the similarity between vectors by the Euclidean distance, the smaller the Euclidean distance, the higher the similarity, and after mapping the Euclidean distance by taking the inverse or by exp(-x) function, the cosine similarity is replaced.

[0078] Step S302: screening the same appeal phrase of each possible appeal phrase based on the appeal similarity; and acquiring the appeal representative degree according to the number of historical call data of each possible appeal phrase, the number of historical call data of the corresponding same appeal phrase, and the type of the corresponding same appeal phrase.

[0079] Considering that the possible appeal phrases with high appeal similarity are similar in appeal, the same appeal phrase of each possible appeal phrase is screened based on the appeal similarity;

[0080] As an example, the similarity threshold is set to 0.6, and in all appeal similarities corresponding to each possible appeal phrase, the other possible appeal phrases corresponding to the appeal similarities greater than the similarity threshold are taken as the same appeal phrases of each possible appeal phrase.

[0081] Considering that the more historical call data a possible appeal phrase exists, the more users use the phrase to express appeal, the stronger the representativeness, and the higher the appeal representative degree; at the same time, the more historical call data the corresponding same appeal phrase exists, and the more types of the corresponding same appeal phrase, the more different expressions the appeal represented by the phrase embodies, and the stronger the universality and coverage, the higher the appeal representative degree, so the appeal representative degree is acquired.

[0082] As an example, the possible appeal phrase is taken as the phrase to be analyzed in turn, the number of historical call data in which the phrase to be analyzed exists is taken as the numerator, the sum of the number of historical call data in which the corresponding similar appeal phrase exists is taken as the denominator, and the fraction ratio is taken as the appeal representative degree of the phrase to be analyzed.

[0083] In the formula, the number of historical call data in which the phrase exists means the number of historical call data containing the corresponding phrase; the sum of the number of similar appeal phrases and the constant 1 is to avoid the denominator being zero; and the distribution of the possible appeal phrase is shown by the existence of the possible appeal phrase and the similar appeal phrase in the historical call data.

[0084] Step S303: screening the appeal phrase based on the appeal representative degree.

[0085] As an example, the possible appeal phrases of the historical call data are sorted in descending order of the appeal representative degree, the top 10% of the possible appeal phrases are selected as the appeal phrases, and the appeal phrases are taken as the demand labels of the users in the city.

[0086] In other embodiments of the present application, the implementer can adjust the proportion number.

[0087] In an embodiment of the present application, after the demand labels are obtained by analyzing the historical data, the method further comprises:

[0088] The latest call data is obtained, the possible appeal phrases are extracted, and the classification labels of the latest call data are obtained according to the cosine similarity between the possible appeal phrases and the demand labels.

[0089] Specifically, the classification labels of the latest call data are obtained by selecting the demand labels corresponding to the possible appeal phrases with the highest cosine similarity, for example, the top 10% of the possible appeal phrases.

[0090] In an embodiment of the present application, the method further comprises analyzing the work order and related data according to the sorted classification work order data (call data), and specifically comprises:

[0091] Each classification label (demand label) is taken as a separate item for analysis.

[0092] (1) The number of a type of work order (divided into appeal work order and consultation work order) in each day in the investigation period (preset to be the past one year) is counted, and the abnormal value is removed according to the formula, so as to calculate the reference value of Monday to Sunday and statutory holidays in the three periods, that is, and Meanwhile, the number of each matter in daily consultation work orders and appeal work orders is calculated, and outliers are removed according to a formula, so as to calculate the reference value of each matter on Monday to Sunday and statutory holidays in the observation period, i.e. and In addition, the number of each matter in daily consultation work orders and appeal work orders, i.e. Q ic and Q ia .

[0093] Reference value Before calculation, outliers are removed according to the box plot method:

[0094] f outlier (X) = lb≤X≤ub;

[0095] lb = Q1-X×IQR;

[0096] ub = Q3+X×IQR;

[0097] Reference value Calculation:

[0098]

[0099] In the formula: Q1 represents the single matter reference value of Q i in consultation work orders; the reference value is calculated according to Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday and statutory holidays in the specified base period, and there are eight corresponding reference values; the reference value has a different set of reference values in different periods according to the change in personnel scale;

[0100] Q2 represents the single matter reference value of Q i in appeal work orders; the reference value is calculated according to Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday and statutory holidays in the specified base period, and there are eight corresponding reference values; the reference value has a different set of reference values in different periods according to the change in personnel scale;

[0101] Q3 represents the daily reference work order quantity of consultation work orders; the reference value is calculated according to Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday and statutory holidays in the specified base period, and there are eight corresponding reference values; the reference value has a different set of reference values in different periods according to the change in personnel scale;

[0102] Q4 represents the daily reference work order quantity of appeal work orders; the reference value is calculated according to Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday and statutory holidays in the specified base period, and there are eight corresponding reference values; the reference value has a different set of reference values in different periods according to the change in personnel scale;

[0103] This indicates the Q in the corresponding consultation ticket. i The average number of work orders for a single item within a certain period. The benchmark value is calculated separately for Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday, and statutory holidays within the specified base period. There are eight corresponding benchmark values. The benchmark value has a different set of benchmark values ​​in different periods according to changes in the number of personnel.

[0104] The Q in the work order indicates a request. i The average number of work orders for a single item within a certain period. The benchmark value is calculated separately for Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday, and statutory holidays within the specified base period. There are eight corresponding benchmark values. The benchmark value has a different set of benchmark values ​​in different periods according to changes in the number of personnel.

[0105] This represents the average number of consultation work orders within a certain period. The benchmark value is calculated separately for Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday, and statutory holidays within the specified base period, resulting in eight corresponding benchmark values. The benchmark value varies in different periods depending on changes in the number of personnel.

[0106] This represents the average number of work orders within a certain period. The benchmark value is calculated separately for Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday, and statutory holidays within the specified base period. There are eight corresponding benchmark values. The benchmark value varies in different periods depending on the size of the workforce.

[0107] Day represents a day from Monday to Sunday, or a public holiday.

[0108] lb represents the lower limit of the threshold range;

[0109] ub represents the upper limit of the threshold range;

[0110] Q1 represents the data grouped from low to high, located at the 25th percentile.

[0111] Q3 indicates that the grouped data is arranged from low to high, and the data is located at 75%;

[0112] IQR corresponds to Q3-Q1, which is the interval between the 25th percentile and the 75th percentile.

[0113] X represents the adjustable magnification factor, with a default value of 1.5. Spacing greater than or equal to 1.5 is considered an outlier.

[0114] (2) Calculate the consultation index and the demand index according to the formula. For example, when calculating the consultation index, first sum up the number of all items in the consultation work order, and then sum the results. Divide by the benchmark value of the same item Then multiply the result by the reference value corresponding to the date of the day The appeal index is calculated using the same logic.

[0115] (3) After the consultation index and the appeal index are calculated, add them together to obtain the 12345 hotline public voice index. The city index is the sum of the consultation index and the appeal index.

[0116] The calculation process of the city 12345 hotline public voice index is as follows:

[0117] I = I c + I a ;

[0118] Where:

[0119]

[0120] Day = Mon, Tue, Wed, Thu, Fri, Sat, Sun, Holiday;

[0121] In the above formula, I represents the city 12345 hotline public voice index; I c represents the city 12345 hotline public voice consultation index; I a represents the city 12345 hotline public voice appeal index; Q ic represents the number of single item work orders in the reporting period in the consultation work order; Q ia represents the number of single item work orders in the reporting period in the appeal work order; n represents the number of last-level items in the consultation work order; m represents the number of last-level items in the appeal work order; represents the single item reference value of Q i in the consultation work order; represents the single item reference value of Q i in the appeal work order; represents the daily reference work order quantity of the appeal work order; represents the daily reference work order quantity of the consultation work order.

[0122] Note:

[0123] I for Index (index);

[0124] C for Consultation (consultation);

[0125] A for Appeal (appeal);

[0126] Q for Quantity (quantity);

[0127] O for Order (work order);

[0128] B for Base Period (benchmark value);

[0129] D for District;

[0130] T fot Town (town / street);

[0131] i is a common letter representing a term in a sequence; n and m are common letters representing quantities; f is a common letter expressing a function; Day is a common word expressing a date; and Outlier is a common word expressing an outlier.

[0132] In summary, addressing the technical problem of inaccurate user demand tag classification in existing data analysis of government service hotline platforms, which hinders in-depth data mining, this invention proposes a big data-based data governance and analysis method for government service hotline platforms. This invention first obtains communication phrases from historical call data, filtering out emphasized phrases and potential request phrases. Further, in any two historical call data sets, emphasized phrases from different calls are matched based on the similarity between them. Then, based on the consistency of usage differences among all matched emphasized phrases, a theme consistency coefficient is obtained, and all historical call data is classified. Finally, in any category of historical call data, based on the similarity between potential request phrases and their distribution, request phrases are selected and recorded as demand tags. This invention effectively solves the problems of inaccurate user demand tag classification and large data mining biases in existing technologies by filtering emphasized phrases and extracting potential request phrases, classifying them based on the theme consistency coefficient between call data, and then obtaining demand tags based on similarity and distribution.

[0133] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0134] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A data governance and analysis method for a government service hotline platform based on big data, characterized in that, The method includes: Obtain communication phrases from historical call data of the government service hotline platform; filter out emphasis phrases based on the frequency of use of the communication phrases; filter out possible appeal phrases based on the similarity between the communication phrases and the emphasis phrases, combined with the frequency of use of the emphasis phrases; In any two historical call data sets, the emphasis phrases in different calls are matched based on the similarity between the emphasis phrases; the theme consistency coefficient is obtained based on the consistency of the usage differences among all matched emphasis phrases; and all historical call data are classified based on the theme consistency coefficient. In any of the historical call data, based on the similarity between the possible request phrases and the distribution of the possible request phrases, request phrases are selected and recorded as request tags. The method for obtaining the theme consistency coefficient includes: in any two call data, the difference between the TF-IDF value of each emphasis phrase in one call data and the TF-IDF value of the matched emphasis phrase is used as the numerator, the sum is used as the denominator, the ratio of the fractions is used as the relative deviation factor, the standard deviation of the relative deviation factor is used as the independent variable, and after negative correlation mapping through the exp(-x) function, the mapped value is used as the theme consistency coefficient of the corresponding two call data, where x represents the independent variable, and exp(-x) is an exponential function with the natural constant e as the base.

2. The data governance and analysis method for a government service hotline platform based on big data as described in claim 1, characterized in that, After obtaining the demand tags, it also includes: Obtain the latest call data and extract the possible request phrases. Based on the cosine similarity between the possible request phrases and the request tags, obtain the classification tags for the latest call data.

3. The data governance and analysis method for a government service hotline platform based on big data as described in claim 1, characterized in that, The method for obtaining the emphasized phrases includes: For any call, obtain the TF-IDF value of each of the communication phrases, and select the communication phrases with the largest TF-IDF value as the emphasis phrases.

4. The data governance and analysis method for a government service hotline platform based on big data as described in claim 1, characterized in that, The methods for obtaining the possible appeal phrases include: Based on the cosine similarity between any of the communication phrases and each of the emphasis phrases, and combined with the TF-IDF value of the emphasis phrases, the appeal relevance degree of the corresponding communication phrases is obtained; The communication phrases whose appeal relevance is greater than a preset appeal threshold are considered as possible appeal phrases.

5. The data governance and analysis method for a government service hotline platform based on big data as described in claim 1, characterized in that, The method for obtaining the aforementioned appeal phrases includes: In any type of historical call data, based on the cosine similarity between one possible appeal phrase and another possible appeal phrase, and combined with the theme consistency coefficient between the historical call data containing the two phrases, the appeal similarity between the two possible appeal phrases is obtained. Based on the similarity of the requests, similar request phrases for each possible request phrase are selected; based on the number of historical call data for each possible request phrase, combined with the number of historical call data for the corresponding similar request phrases, and the type of the corresponding similar request phrases, the representativeness of the requests is obtained. The phrases related to the demands were selected based on their representativeness.

6. The data governance and analysis method for a government service hotline platform based on big data as described in claim 5, characterized in that, The method for selecting appeal phrases based on the representativeness of the appeal includes: Select the possible phrases of appeal with the highest representativeness of the appeal as the appeal phrases.

7. The data governance and analysis method for a government service hotline platform based on big data as described in claim 1, characterized in that, The method for obtaining the communication phrases includes: The historical call data is converted into text and segmented into words. At the same time, the Chinese Dictionary and local dialect dictionaries are called to remove stop words, and the remaining word groups are used as communication word groups.

8. The data governance and analysis method for a government service hotline platform based on big data as described in claim 1, characterized in that, The method for matching emphasis phrases in different calls based on the similarity between the emphasis phrases includes: The emphasized word group in the call data is matched with the emphasized word group in another call data that has the highest cosine similarity.

9. The data governance and analysis method for a government service hotline platform based on big data according to claim 1, characterized in that, The method for classifying all historical call data based on the subject consistency coefficient includes: All historical call data were clustered and classified using the DBSCAN clustering algorithm.

Images