Method, device and electronic equipment for processing customer information

Abstract

The application discloses a kind of processing method, device and electronic equipment of customer information, which is applied to big data field, and the method comprises: the customer information in multiple incoming call work orders of multiple customers is encoded and processed, to obtain multiple documents corresponding to multiple incoming call work orders;Obtain the weighting matrix model of each document in multiple documents;Obtain the theme word set of each document in multiple documents;According to the theme word set of each document and the weighting matrix model of each document, the target similarity between each document in multiple documents and each other document is calculated;According to target similarity, multiple documents are clustered based on PAM clustering algorithm, to obtain the clustered document set;From the clustered document set, obtain the target information of multiple customers.By the present application, the problem that the value information of incoming call work order is extracted by artificial labeling or natural language processing technology in related technology, leading to the inaccuracy of information extracted from incoming call work order is solved.

Description

Technical Field

[0001] This application relates to the field of big data, and more specifically, to a method, apparatus, and electronic device for processing customer information. Background Technology

[0002] When extracting potential value information from multiple call orders from multiple customers, manual classification and labeling are often used. However, manual labeling of multiple call orders is inefficient and is greatly affected by the subjectivity of different labelers, resulting in poor quality and accuracy of the labeling results, which cannot objectively reflect the potential value information in multiple call orders.

[0003] Currently, natural language processing (NLP) techniques are also used to calculate text similarity, extract potential value information from multiple call logs, and transform unstructured text into structured information that is easy for computers to recognize and process, thereby enabling the mining of textual information. Traditional text similarity calculation models can be mainly divided into three categories: Vector Space Model (VSM), Generalized Vector Space Model (GVSM), and Latent Semantic Indexing (LSI), also known as Latent Semantic Analysis (LSA). These three models generally use feature selection functions to extract keywords as text features, such as word contribution (TC), TFIDF, information entropy / information gain, mutual information, and CHI statistics. However, traditional text similarity models require large-scale corpora and often ignore the grammatical and organizational structure and semantic information in the text. For example, the Vector Space Model uses a bag-of-words model to construct the feature space, but this model typically uses a "hard matching" method in feature matching, which cannot solve the problems of "one meaning, multiple words" and "one word, multiple meanings."

[0004] There is currently no effective solution to the problem that the information extracted from call logs is inaccurate due to the use of manual annotation or natural language processing techniques in related technologies. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, and electronic device for processing customer information, in order to solve the problem that the information extracted from call work orders is inaccurate when using manual annotation or natural language processing techniques to extract valuable information from call work orders in related technologies.

[0006] To achieve the above objectives, according to one aspect of this application, a method for processing customer information is provided. The method includes: encoding customer information from multiple call tickets for multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; obtaining a weighted matrix model for each document in the multiple documents; obtaining a set of keywords for each document in the multiple documents; calculating a target similarity between each document and every other document based on the set of keywords and the weighted matrix model of each document; clustering the multiple documents based on the target similarity using a PAM clustering algorithm to obtain a clustered document set; and obtaining target information for multiple customers from the clustered document set.

[0007] Further, obtaining the weighted matrix model for each document in the plurality of documents includes: deleting preset strings from the plurality of documents, and performing word segmentation and filtering on each document in the plurality of documents to obtain a first corpus; substituting the first corpus into a word2vec model for calculation to obtain word vectors for each document; substituting the plurality of documents into a TextRank model for calculation to obtain a keyword set for each document; sorting the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword sets of other documents to obtain a keyword vector for each document; replacing the keywords in the keyword vector with the word vector, and calculating the weights of the word vectors in the keyword vector using a TF-IDF feature selection function to obtain the replaced keyword vector and the word vector weights corresponding to the word vectors in the replaced keyword vector; and obtaining the weighted matrix model for each document based on the replaced keyword vector and the word vector weights.

[0008] Further, the process of substituting the multiple documents into the TextRank model for calculation to obtain the keyword set for each document includes: performing a first processing on each of the multiple documents to obtain words with preset parts of speech as candidate keywords for each document, wherein the first processing includes at least the following processing: sentence segmentation, word segmentation, filtering, and part-of-speech tagging; obtaining a second corpus based on each document and the candidate keywords of each document; using the TextRank model to convert the candidate keywords of each document into a directed keyword graph, and calculating the weight of the candidate keywords in each document to obtain a weighted directed keyword graph for each document; and obtaining candidate keywords with weights higher than preset weights from the weighted directed keyword graph to obtain the keyword set for each document.

[0009] Furthermore, obtaining the topic word set for each of the multiple documents includes: calculating the word influence score of the word vectors configured according to the TextRank model, as shown in the following formula: Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf a TF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. ba w represents the weights of keyword a and keyword b in the weighted keyword directed graph. bc The weights of keywords b and c in the weighted keyword directed graph are represented. A second processing step is performed on each of the multiple documents to obtain a third corpus and the keywords in the third corpus. The second processing step includes at least the following steps: word segmentation, filtering, initial keyword extraction, and frequency analysis. Based on the formula for calculating the keyword influence score, the keyword influence score of the keywords in the third corpus is iteratively calculated. The iterative calculation stops when the difference between the Nth iteration and the (N-1)th iteration is less than a preset threshold. From the calculation result of the Nth iteration, the keywords with keyword influence scores higher than the preset keyword influence score are obtained, thus obtaining the keyword set for each document.

[0010] Further, calculating the target similarity between each document and every other document based on the topic word set of each document and the weighted matrix model of each document includes: constructing a bipartite graph model based on the topic word set of each document to calculate a first similarity between the topic word set of each document and the topic word set of every other document; calculating a second similarity between the weighted matrix model of each document and the weighted matrix model of every other document based on the least squares distance formula of the matrix; and calculating the target similarity between each document and every other document based on the first similarity and the second similarity.

[0011] Further, constructing a bipartite graph model based on the topic word set of each document to calculate the first similarity between the topic word set of each document and the topic word set of each other document includes: constructing the bipartite graph model of each document and each other document based on the topic word set of each document; calculating the maximum weight of bipartite matching between each document and each other document using the Hungarian algorithm, and using the maximum weight of bipartite matching as the first similarity between the topic word set of each document and the topic word set of each other document.

[0012] Furthermore, obtaining target information for multiple customers from the clustered document set includes: processing each clustered document set according to the method for obtaining a topic word set to obtain a topic word set for each clustered document set; processing each clustered document set according to the method for obtaining a keyword set to obtain a keyword set for each clustered document set; and determining the target information for the multiple customers based on the topic word set and the keyword set of each clustered document set.

[0013] Furthermore, after calculating the target similarity between each document and every other document based on the topic word set of each document and the weighted matrix model of each document, the method further includes: adding the target similarity between each document and every other document to obtain a third similarity for each document; obtaining target documents from the multiple documents whose third similarity is higher than a preset similarity to obtain a target document set; and pushing the target document set to the target object.

[0014] To achieve the above objectives, according to another aspect of this application, a customer information processing apparatus is provided, comprising: a first acquisition unit, configured to encode customer information from multiple call tickets of multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; a second acquisition unit, configured to acquire a weighted matrix model of each document in the multiple documents; a third acquisition unit, configured to acquire a set of keywords for each document in the multiple documents; a calculation unit, configured to calculate the target similarity between each document and each other document based on the set of keywords and the weighted matrix model of each document; a fourth acquisition unit, configured to cluster the multiple documents based on the target similarity using a PAM clustering algorithm to obtain a clustered document set; and a fifth acquisition unit, configured to acquire target information of multiple customers from the clustered document set.

[0015] Further, the second acquisition unit includes: a first processing subunit, used to delete preset strings from the plurality of documents and perform word segmentation and filtering on each of the plurality of documents to obtain a first corpus; a first calculation subunit, used to substitute the first corpus into a word2vec model for calculation to obtain word vectors for each document; a second calculation subunit, used to substitute the plurality of documents into a TextRank model for calculation to obtain a keyword set for each document; a first acquisition subunit, used to sort the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword sets of other documents, to obtain a keyword vector for each document; a third calculation subunit, used to replace the keywords in the keyword vector with the word vectors and use a TF-IDF feature selection function to calculate the weights of the word vectors in the keyword vector, to obtain the replaced keyword vector and the word vector weights corresponding to the word vectors in the replaced keyword vector; and a second acquisition subunit, used to obtain a weighted matrix model for each document based on the replaced keyword vector and the word vector weights.

[0016] Further, the second calculation subunit includes: a processing module, used to perform a first processing on each of the plurality of documents to obtain words with preset parts of speech as candidate keywords for each document, wherein the first processing includes at least the following processing: sentence segmentation processing, word segmentation processing, filtering processing, and part-of-speech tagging processing; a first acquisition module, used to acquire a second corpus based on each document and the candidate keywords of each document; a first calculation module, used to convert the candidate keywords of each document into a keyword directed graph using the TextRank model, and calculate the weight of the candidate keywords in each document to obtain a weighted keyword directed graph for each document; and a second acquisition module, used to acquire candidate keywords with weights higher than preset weights from the weighted keyword directed graph to obtain a keyword set for each document.

[0017] Furthermore, the third acquisition unit includes a configuration subunit, used to configure the calculation formula for the word influence score of word vectors according to the TextRank model, the calculation formula being as follows: Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf aTF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. ba w represents the weights of keyword a and keyword b in the weighted keyword directed graph. bc The weights of keywords b and c in the weighted keyword directed graph are represented by: a second processing subunit, used to perform a second processing on each of the plurality of documents to obtain a third corpus and the topic words in the third corpus, wherein the second processing includes at least the following processing: word segmentation processing, filtering processing, initial topic word extraction processing, and statistical word frequency processing; a fourth calculation subunit, used to iteratively calculate the word influence score of the topic words in the third corpus according to the calculation formula of the word influence score, and stop the iterative calculation when the difference between the Nth iteration calculation and the N-1th iteration calculation is less than a preset threshold, and obtain the topic words whose word influence scores are higher than the preset word influence scores from the calculation results of the Nth iteration calculation, thereby obtaining the topic word set of each document.

[0018] Further, the calculation unit includes: a construction subunit, used to construct a bipartite graph model based on the topic word set of each document, to calculate a first similarity between the topic word set of each document and the topic word set of each other document; a fifth calculation subunit, used to calculate a second similarity between the weighted matrix model of each document and the weighted matrix model of each other document based on the least squares distance formula of the matrix; and a sixth calculation subunit, used to calculate a target similarity between each document and each other document based on the first similarity and the second similarity.

[0019] Further, the first construction subunit includes: a construction module, configured to construct the bipartite graph model of each document and each other document based on the topic word set of each document; and a second calculation module, configured to calculate the maximum weight of bipartite matching between each document and each other document using the Hungarian algorithm, and use the maximum weight of bipartite matching as the first similarity between the topic word set of each document and the topic word set of each other document.

[0020] Furthermore, the fifth acquisition unit includes: a third processing subunit, used to process each clustered document set according to the method for acquiring the topic word set, and acquire the topic word set of each clustered document set; a fourth processing subunit, used to process each clustered document set according to the method for acquiring the keyword set, and acquire the keyword set of each clustered document set; and a determination subunit, used to determine the target information of the multiple customers based on the topic word set and the keyword set of each clustered document set.

[0021] Furthermore, the apparatus further includes: a sixth acquisition unit, configured to, after calculating the target similarity between each document and every other document in the plurality of documents based on the topic word set and the weighted matrix model, add the target similarity between each document and every other document in the plurality of documents to obtain a third similarity for each document; a seventh acquisition unit, configured to acquire target documents from the plurality of documents whose third similarity is higher than a preset similarity, to obtain a target document set; and a push unit, configured to push the target document set to a target object.

[0022] To achieve the above objectives, according to one aspect of this application, an electronic device is provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the customer information processing method described in any of the above.

[0023] This application employs the following steps: encoding customer information from multiple call tickets for multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; obtaining a weighted matrix model for each document in the multiple documents; obtaining a set of keywords for each document in the multiple documents; calculating the target similarity between each document and every other document based on the set of keywords and the weighted matrix model of each document; clustering the multiple documents based on the target similarity using the PAM clustering algorithm to obtain a clustered document set; and extracting target information of multiple customers from the clustered document set. This solves the problem of inaccurate information extracted from call tickets due to manual annotation or natural language processing techniques. By calculating the weighted matrix model of each document in multiple call order forms and the keyword set of each document, the target similarity between multiple documents is calculated, thereby extracting target information from multiple documents. This algorithm automatically extracts target information from multiple documents, avoiding the impact of manual annotation of multiple documents on the information quality of target information, and achieving the effect of extracting more accurate target information from multiple call order forms. Attached Figure Description

[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 This is a flowchart of a customer information processing method provided in the embodiments of this application;

[0026] Figure 2 This is an illustration of an optional customer information processing method provided according to an embodiment of this application. Figure 1 ;

[0027] Figure 3 This is an illustration of an optional customer information processing method provided according to an embodiment of this application. Figure 2 ;

[0028] Figure 4 This is an illustration of an optional customer information processing method provided according to an embodiment of this application. Figure 3 ;

[0029] Figure 5 This is a schematic diagram of a customer information processing apparatus according to an embodiment of this application;

[0030] Figure 6 This is a schematic diagram of an electronic device for processing customer information according to an embodiment of this application. Detailed Implementation

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, customer call tickets, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0035] The present invention will now be described in conjunction with preferred implementation steps. Figure 1 This is a flowchart of a customer information processing method provided according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0036] Step S101: Encode the customer information in multiple call work orders from multiple customers to obtain multiple documents corresponding to the multiple call work orders, wherein the multiple documents have the same encoding format.

[0037] Currently, multiple call tickets from multiple customers are generally multiple text files recording the communication process between customers and staff. In this embodiment, in order to further process customer information, it is necessary to convert multiple text files into multiple documents with the same encoding format.

[0038] Step S102: Obtain the weighted matrix model for each document in the multiple documents.

[0039] Step S103: Obtain the set of keywords for each document in the multiple documents.

[0040] Step S104: Based on the topic term set of each document and the weighted matrix model of each document, calculate the target similarity between each document and each other in multiple documents.

[0041] In this embodiment, in order to calculate the target similarity between each document and every other document in multiple documents, after obtaining multiple documents with the same encoding format, it is necessary to calculate the weighted matrix model of each document in multiple documents and the topic word set of each document in multiple documents.

[0042] Step S105: Based on the target similarity, cluster multiple documents using the PAM clustering algorithm to obtain a clustered document set.

[0043] Step S106: Obtain target information for multiple customers from the clustered document set.

[0044] In existing technologies, when analyzing multiple call tickets from multiple customers to extract information of interest to staff, manual annotation is generally used. However, manual annotation of multiple call tickets is inefficient and cannot meet the needs of current big data analysis. Moreover, due to the subjective differences among annotators, there is no uniform standard for annotation results, leading to poor quality of information extracted from multiple call tickets.

[0045] In this embodiment, to improve the efficiency of annotating multiple call tickets and the quality of information extracted from them, the PAM clustering algorithm is used to cluster multiple documents, resulting in a clustered document set. Then, target information for multiple customers is obtained from this set. This target information includes customer feedback on the company's products and services, as well as potential business needs. Extracting information from multiple call tickets from multiple customers can provide valuable support for improving customer service and targeted marketing.

[0046] In summary, the customer information processing method provided in this application embodiment encodes customer information from multiple call tickets for multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; obtains a weighted matrix model for each document in the multiple documents; obtains a set of keywords for each document in the multiple documents; calculates the target similarity between each document and each other document based on the set of keywords and the weighted matrix model of each document; clusters the multiple documents based on the target similarity using the PAM clustering algorithm to obtain a set of clustered documents; and obtains the target information of multiple customers from the set of clustered documents. This solves the problem that extracting value information from call tickets using manual annotation or natural language processing techniques leads to inaccurate information extracted from call tickets. By calculating the weighted matrix model of each document in multiple call order forms and the keyword set of each document, the target similarity between multiple documents is calculated, thereby extracting target information from multiple documents. This algorithm automatically extracts target information from multiple documents, avoiding the impact of manual annotation of multiple documents on the information quality of target information, and achieving the effect of extracting more accurate target information from multiple call order forms.

[0047] Optionally, in the customer information processing method provided in this application embodiment, obtaining the weighted matrix model of each document in multiple documents includes: deleting preset strings from multiple documents, and performing word segmentation and filtering on each document in multiple documents to obtain a first corpus; substituting the first corpus into a word2vec model for calculation to obtain word vectors for each document; substituting multiple documents into a TextRank model for calculation to obtain a keyword set for each document; sorting the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword sets of other documents to obtain a keyword vector for each document; replacing the keywords in the keyword vector with word vectors, and calculating the weights of the word vectors in the keyword vector using a TF-IDF feature selection function to obtain the replaced keyword vector and the word vector weights corresponding to the word vectors in the replaced keyword vector; and obtaining the weighted matrix model of each document based on the replaced keyword vector and the word vector weights.

[0048] Currently, in the process of extracting core content from text using natural language processing techniques, the common methods are to extract small sets of discrete categories from the text or to use word2vec models to calculate word vectors. Using small sets of discrete categories to process text means representing the text using noun phrases and verb phrases. However, this method cannot fully capture the richness of multiple phrases in the text and requires a large feature space. Comparing the word2vec model with previous best-performing techniques based on different types of neural networks, it was found that word2vec models offer significant accuracy improvements at a lower computational cost. However, an inherent limitation of using word vectors to represent text is that word vectors do not consider word order in the text, and they cannot represent phrases that conform to linguistic conventions. For example, word vectors cannot accurately represent the meaning of idioms. Furthermore, existing models using word vectors to represent text, such as word vector mean models, word vector clustering models, and doc2vec models, do not consider the influence of words within the text.

[0049] In this embodiment, to obtain a more accurate representation of the text's meaning, multiple documents are first segmented and filtered to obtain a first corpus for subsequent calculations. Next, the first corpus is fed into the word2vec model to obtain word vectors for each document. Then, multiple documents are fed into the TextRank model to obtain a keyword set for each document. The similarity between each keyword in each document's keyword set and each keyword in every other document's keyword set is calculated. Based on the similarity scores, a greedy selection algorithm is used to sort at least one keyword in each document's keyword set, obtaining a keyword vector for each document. The word vectors obtained from the word2vec model are used to replace the keywords in the keyword vectors, and the TF-IDF feature selection function is used to calculate the weights of the word vectors in the keyword vectors, resulting in the replaced keyword vectors and their corresponding weights. Based on the replaced keyword vectors and their weights, a weighted matrix model for each document is obtained.

[0050] Common text feature selection functions in existing technologies mainly include word contribution (TC), information entropy / information gain, mutual information, and χ². 2 Statistically, information entropy / information gain only considers the contribution of features to the overall context and cannot be applied to individual categories. Therefore, information entropy / information gain is only suitable for selecting "global" features, not "local" features. This makes it impossible to combine the contextual information adjacent to the word for calculation, which is not conducive to extracting keywords from the text. Mutual information is only sensitive to critical feature words, and therefore often tends to favor the influence of rare words on text classification. However, the words in call slips are generally everyday language and not rare words, so mutual information is also not suitable for calculating keywords in this embodiment. χ 2 Statistics are only performed when the correlation between the text's feature words and categories meets the χ² standard. 2 Accurate results can only be obtained under the condition of distribution. When this assumption is not met, the results will differ greatly from the actual situation. Furthermore, the wording in the work order cannot guarantee that χ² will be satisfied. 2 Distribution, therefore χ 2Distribution is also unsuitable for keyword extraction in this embodiment. Based on the above analysis, in this embodiment, the TF-IDF value in the word contribution TC is used as the feature selection function to calculate the weights of word vectors in the keyword vector. TF (Term Frequency) in the TF-IDF value is a measure of the local importance of a word, representing its frequency in the document; a larger TF value reflects a greater contribution of the word to the document. IDF (Inverse Document Frequency) refers to the inverse document frequency index, with the formula log(D / D). w D represents the total number of texts. w Let w be the number of texts in which word w appears. IDF is the cross-entropy (Kullback-Leibler Divergence) of the probability distribution of keywords under a specific condition. Therefore, the TF-IDF value measures the importance of a word from two aspects: the frequency of the word in the text and its distribution in the corpus.

[0051] By replacing keywords in keyword vectors calculated using the TextRank model with word vectors calculated using the word2vec model, and using the TF-IDF feature selection function to calculate the weights of the word vectors, a text-weighted matrix model that can more accurately represent text is obtained. This not only solves the problem of polysemy but also combines semantic information from the context, calculating the weights of word vectors from both the frequency of words in the document and their distribution in the corpus. This improves the accuracy of word vectors in representing text, achieving the effect of using a text-weighted matrix model to represent documents, and further improving the accuracy of similarity between multiple documents.

[0052] Optionally, in the customer information processing method provided in this application embodiment, substituting multiple documents into the TextRank model for calculation to obtain the keyword set of each document includes: performing a first processing on each of the multiple documents to obtain words with preset parts of speech as candidate keywords for each document, wherein the first processing includes at least the following processing: sentence segmentation, word segmentation, filtering, and part-of-speech tagging; obtaining a second corpus based on each document and the candidate keywords of each document; using the TextRank model to convert the candidate keywords of each document into a keyword directed graph, and calculating the weight of the candidate keywords in each document to obtain a weighted keyword directed graph for each document; and obtaining candidate keywords with weights higher than preset weights from the weighted keyword directed graph to obtain the keyword set for each document.

[0053] Generally, keywords can reflect the key information or core ideas of the entire text. Since TextRank or LDA models based on word graphs do not require prior training on a corpus, existing natural language processing techniques often use these two algorithms to extract keywords from text. However, when new text is added to the calculation, the LDA model has poor scalability, and if the LDA model is used for calculation, the model needs to be retrained. The TextRank model is more concise in calculation. Therefore, in this embodiment, the TextRank model is used as a method to extract keywords from multiple documents.

[0054] In this embodiment, to obtain the keyword set corresponding to multiple documents, it is necessary to first perform sentence segmentation, word segmentation, filtering, and part-of-speech tagging on multiple documents in the original corpus to obtain candidate keywords for each document; the candidate keywords corresponding to each document form a second corpus; then, the TextRank model is used to construct a candidate keyword graph G = (V, E), with the candidate keywords as nodes E in the keyword graph, and an edge V between any two nodes is constructed using co-occurrence relations, and the weight of each node is iteratively calculated until the error rate between any two nodes in the keyword graph is less than a preset threshold, at which point the iterative calculation of the node weights is stopped, resulting in a weighted directed keyword graph corresponding to each document; based on the weights of each node in the weighted directed keyword graph, the multiple nodes in the weighted directed keyword graph are sorted, and c candidate keywords with node weights greater than the preset weights are obtained from the multiple nodes as the keyword set corresponding to each document. If the c keywords in the keyword set of a document are adjacent phrases in the text of the call order, then the adjacent keywords are combined to obtain multi-word keywords.

[0055] By constructing a weighted keyword directed graph for each document in multiple documents using the TextRank model, a set of keywords for multiple documents is obtained, which improves the ability to represent text information using keywords, achieves a more accurate representation of the meaning of documents, and improves the accuracy of similarity between multiple documents.

[0056] Optionally, in the customer information processing method provided in this application embodiment, obtaining the topic word set for each document in multiple documents includes: calculating the word influence score of word vectors based on the TextRank model, as shown in the following formula: Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf a TF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. ba w represents the weights of keyword a and keyword b in a weighted directed graph. bc This represents the weights of keywords b and c in the weighted keyword directed graph. For each document in the multiple documents, a second processing step is performed to obtain a third corpus and the keywords within it. This second processing step includes at least the following steps: word segmentation, filtering, initial keyword extraction, and frequency analysis. Based on the formula for calculating keyword influence scores, the keyword influence scores of the keywords in the third corpus are iteratively calculated. The iteration stops when the difference between the Nth iteration and the (N-1th)th iteration is less than a preset threshold. From the results of the Nth iteration, keywords with keyword influence scores higher than the preset threshold are obtained, resulting in a keyword set for each document.

[0057] Currently, topic extraction models are commonly used to identify topic information in large-scale documents. However, when faced with dynamically growing text, topic models struggle to find suitable topic projection dimensions, resulting in inaccurate representations of document topics. Furthermore, current topic models utilize the WordNet and HowNet semantic dictionaries, which are semantic dictionaries specific to both Chinese and English. When dealing with highly specialized text processing domains, the lack of timely inclusion of specialized terms hinders the calculation of word similarity and the resolution of polysemy. In contrast, word vector training captures contextual information, which not only binds word relationships but also effectively addresses semantic gaps caused by missing specialized terms in dictionaries. Therefore, in this embodiment, word vectors are used to measure semantic relationships between words to obtain a set of topic terms for each document across multiple documents.

[0058] In this embodiment, because word vectors and keywords are introduced in the process of obtaining the topic word set corresponding to the document, the original word influence formula needs to be improved when calculating the word influence between words. The improved word influence scoring formula is as follows:

[0059]

[0060] Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf a TF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. ba w represents the weights of keyword a and keyword b in a weighted directed graph. bc This represents the weights of keywords b and c in the weighted keyword directed graph. After determining the formula for calculating word influence, multiple documents undergo word segmentation, filtering, initial topic word extraction, and frequency counting to obtain the third corpus and its topic words. Finally, based on the formula for calculating word influence scores, the word influence score of each topic word in the third corpus is iteratively calculated until the difference between the Nth iteration and the (N-1)th iteration is less than a preset threshold, or the maximum number of iterations is reached. At this point, the iteration stops, and from the results of the Nth iteration, topic words with word influence scores higher than the preset word influence scores are obtained in each document, resulting in the topic word set corresponding to each document.

[0061] By introducing the calculation formulas for word vectors and keywords when calculating the word influence of topic words, using word vectors and keywords to measure the relationship between topic words, and improving the calculation formula for word influence to calculate the weight of topic words, a set of topic words corresponding to each document is obtained. The calculated topic words combine contextual semantic information, achieving the effect of more accurately extracting topic words from multiple documents, and further improving the accuracy of similarity between multiple documents.

[0062] Optionally, in the customer information processing method provided in this application embodiment, calculating the target similarity between each document and each other document based on the topic word set of each document and the weighted matrix model of each document includes: constructing a bipartite graph model based on the topic word set of each document to calculate a first similarity between the topic word set of each document and the topic word set of each other document; calculating a second similarity between the weighted matrix model of each document and the weighted matrix model of each other document based on the least squares distance formula of the matrix; and calculating the target similarity between each document and each other document based on the first similarity and the second similarity.

[0063] To calculate the target similarity between each document and every other document, we first need to construct a bipartite graph model using the topic word set of each document and calculate the first similarity between the topic word set of each document and the topic word sets of every other document. Then, using the least squares distance formula of matrices, we calculate the second similarity between the weighted matrix model of each document and the weighted matrix model of every other document. For example, the specific steps to calculate the similarity between text matrices doc1 (N×m dimensional matrix, where N is the word vector dimension) and doc2 (N×n dimensional matrix) are as follows: perform orthogonal triangular decomposition on matrix doc1 to obtain orthogonal matrix X; calculate X... T Then, perform orthogonal triangular decomposition on matrix doc2 to obtain the difference matrix D between matrix doc1 and matrix doc2; calculate the distance (i.e., the second similarity) between matrix doc1 and matrix doc2 using the following formula:

[0064]

[0065] Among them, a ij It is the element in the i-th row and j-th column of the difference matrix D. Finally, based on the first and second similarities, the target similarity between each document and every other document is calculated. The formula for calculating the target similarity is as follows:

[0066] similarity(doc1,doc2)=αDSM(doc1,doc2)+βTSM(doc1,doc2)

[0067] Where doc1 and doc2 are any two documents, DSM(doc1,doc2) represents the first similarity between matrix doc1 and matrix doc2, TSM(doc1,doc2) represents the second similarity between matrix doc1 and matrix doc2, and α and β are the weights of the first and second similarities, respectively.

[0068] By combining the topic word set and weighted matrix model of each document, the target similarity between each document and every other document is calculated. This not only avoids the influence of semantic blanks, but also incorporates the contextual semantic information in the document, making the target similarity between each document and every other document more accurate, thus improving the accuracy of similarity between multiple documents.

[0069] Optionally, in the customer information processing method provided in this application embodiment, constructing a bipartite graph model based on the topic word set of each document to calculate the first similarity between the topic word set of each document and the topic word sets of all other documents includes: constructing a bipartite graph model between each document and all other documents based on the topic word set of each document; calculating the maximum weight of bipartite matching between each document and all other documents using the Hungarian algorithm, and using the maximum weight of bipartite matching as the first similarity between the topic word set of each document and the topic word sets of all other documents.

[0070] In this embodiment, in order to calculate the first similarity between the topic word set of each document and the topic word set of every other document, after obtaining the topic word set of each document, a bipartite graph model between each document and every other document is constructed using the topic word set of each document. That is, the bipartite graph model between two documents is used to represent the correlation between the two documents. The specific process of constructing a bipartite graph model between two documents is as follows: Given two topic sets T_d1 and T_d2 of two documents, construct a bipartite graph B(T_d1,T_d2), where |V(T_d1)| represents a node in topic set T_d1, and |V(T_d2)| represents a node in topic set T_d2. Use b(u) to represent a related node u in B(T_d1,T_d2). For each node u∈V(T_d1) in the bipartite graph B(T_d1,T_d2), select the node v with the highest similarity to u in node v∈V(T_d2) to form an edge. If there are multiple nodes v such that the similarity w(b(u),b(v)) between node u and node v is maximized, then multiple edges are formed. For each node v∈ For V(T_d2), similarly, in nodes u∈V(T_d1), select the node with the highest similarity to v to form a connection. If the connection already exists, there is no need to add another connection. Then, use the Hungarian algorithm to find the edge with the highest similarity among the connected edges. That is, starting from an unmatched point, keep searching for augmenting paths. If an augmenting path is found, match the matching edge on the augmenting path with the matching point. If no augmenting path is found, it is confirmed that the bipartite graph has reached the maximum matching. After traversing all nodes in the bipartite graph once, obtain the unmatched nodes, construct a new bipartite graph and match again until there are no unmatched nodes in the bipartite graph. Then, obtain the maximum similarity w(b(u), b(v)) (i.e., the maximum weight) from the matched bipartite graph. Finally, use the maximum weight after matching the bipartite graph as the first similarity between the topic word set of each document and the topic word set of every other document.

[0071] By introducing a bipartite graph model and the Hungarian algorithm, the first similarity between the topic word sets of each document in multiple documents is calculated, making the first similarity between the topic word sets of each document in multiple documents more accurate, thereby improving the accuracy of similarity between multiple documents.

[0072] Optionally, in the customer information processing method provided in this application embodiment, obtaining target information of multiple customers from the clustered document set includes: processing each clustered document set according to the method for obtaining a topic word set, and obtaining a topic word set for each clustered document set; processing each clustered document set according to the method for obtaining a keyword set, and obtaining a keyword set for each clustered document set; and determining the target information of multiple customers based on the topic word set and the keyword set of each clustered document set.

[0073] To extract potential customer value information (i.e., target information) from multiple documents, we first need to extract keywords from each of the multiple clustered documents, using the method for obtaining keyword sets from documents. The keyword set corresponding to each clustered document is then used as the customer information for the customer caller information service subclass. Next, using the method for obtaining keyword sets from documents, we extract keywords from each of the multiple clustered documents. The keyword set corresponding to each clustered document is then used as the customer information for the customer caller information problem summary class. Finally, by combining the customer information from the customer caller information service subclass and the customer information from the customer caller information problem summary class, we obtain the target information extracted from multiple call tickets from multiple customers.

[0074] By obtaining a set of thematic terms and a set of keywords from the documents, the clustered documents are processed to obtain customer information for the customer call information business subclass and customer information for the customer call information problem summary class (i.e., target information). This achieves the goal of fully considering the contextual information in the call order when extracting target information, thereby achieving the effect of mining the target information that customers value more from the call order, and further achieving the effect of extracting more accurate potential customer value information from the call order.

[0075] Optionally, in the customer information processing method provided in this application embodiment, after calculating the target similarity between each document and each other document in multiple documents based on the topic word set of each document and the weighted matrix model of each document, the above method further includes: adding the target similarity between each document and each other document in multiple documents to obtain a third similarity of each document; obtaining target documents with a third similarity higher than a preset similarity from multiple documents to obtain a target document set; and pushing the target document set to the target object.

[0076] After calculating the target similarity between each document in multiple documents, the solution implemented in this embodiment also includes obtaining at least one of the most popular call tickets from multiple call tickets, that is, at least one call ticket whose core content appears most frequently in multiple call tickets: First, the sum of the target similarity between each document and each other in multiple documents is calculated to obtain the third similarity corresponding to each document; then, target documents with a third similarity higher than a preset similarity are obtained from multiple documents to obtain a set of target documents, which is at least one of the most popular call tickets among multiple call tickets; finally, the set of target documents is pushed to relevant staff (i.e., target objects) to improve the quality of customer service and provide customers with more precise marketing to meet their urgent needs.

[0077] By calculating the third similarity of multiple call orders, the call orders that customers value are identified and pushed to relevant staff. This improves the quality of service provided by staff and facilitates targeted marketing to meet customers' urgent needs. It achieves the effect of more accurately extracting call orders that customers value, and further enhances customer satisfaction.

[0078] Optionally, in this embodiment, the process of calculating the similarity between multiple weighted matrix models can be as follows: Figure 2 As shown, the first corpus is obtained by preprocessing and word segmentation of multiple documents in the corpus. Then, the TextRank model is used to extract the keyword set corresponding to each document in the first corpus. The Word2vec model is used to train the first corpus to obtain word vectors. Then, each keyword in the keyword set corresponding to each document is replaced with the corresponding word vector to obtain keyword vectors. The weights of the keyword vectors are calculated based on the TF-IDF value to obtain the text matrix model of each document. The similarity between multiple text matrix models is calculated using the least matrix squares distance. Finally, the similarity between multiple text matrix models is substituted into the formula for calculating text similarity to obtain the text similarity value between multiple documents.

[0079] Optionally, in this embodiment, the main workflow for extracting valuable information from multiple call tickets from multiple customers can be as follows: Figure 3As shown. The process involves five steps: First, semantic retrieval of preprocessed work orders is performed using a Word2vec word vector model. Second, keywords are extracted from multiple work orders using a TextRank model, and topic words are mined based on keyword influence. Third, the text content of multiple work orders is represented using a text weighted matrix model, and the text content is represented using topic words based on a text graph model. Fourth, the keyword model similarity between multiple text weighted matrices is calculated using a matrix least squares distance algorithm, and the topic model similarity between multiple topic word sets is calculated using a Hungarian bipartite graph matching algorithm. Fifth, potential customer intent is mined using a clustering analysis algorithm, and popular work orders are mined from multiple work orders based on the calculated text similarity.

[0080] Optionally, in this embodiment, the workflow for extracting valuable information from multiple call tickets from multiple customers can be as follows: Figure 4 As shown. The first step is to preprocess the original corpus by using a Word2vec-based word vector model to calculate the corresponding word vectors. The second step is to extract keywords from the processed corpus using the TextRank model, obtaining a set of text keywords. Weighted scores for the keywords are then calculated iteratively based on their term frequency / inverse document frequency, resulting in a set of text topics. The third step is to construct a weighted text matrix based on the keyword set, and simultaneously construct a bipartite graph model and calculate topic similarity based on the topic set. The fourth step is to calculate the semantic similarity between the weighted text matrices using the matrix least squares distance algorithm, calculate the topic similarity using text topic similarity matching, and finally calculate the overall similarity between the texts based on both semantic and topic similarities. The fifth step involves preprocessing multiple call orders, using the PAM clustering algorithm to analyze and calculate the preprocessed call orders, extracting topic words and keywords from the call orders, and finally mining customer call intent based on the topic words and keywords extracted from the call orders, and mining the top call orders (i.e. popular call orders) from the multiple call orders based on the text similarity in the fourth step.

[0081] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0082] This application also provides a customer information processing apparatus. It should be noted that the customer information processing apparatus of this application can be used to execute the customer information processing method provided in this application. The customer information processing apparatus provided in this application will be described below.

[0083] Figure 5 This is a schematic diagram of a customer information processing apparatus according to an embodiment of this application. Figure 5 As shown, the device includes: a first acquisition unit 501, a second acquisition unit 502, a third acquisition unit 503, a calculation unit 504, a fourth acquisition unit 505, and a fifth acquisition unit 506.

[0084] Specifically, the first acquisition unit 501 is used to encode customer information in multiple call work orders from multiple customers to obtain multiple documents corresponding to the multiple call work orders, wherein the multiple documents have the same encoding format.

[0085] The second acquisition unit 502 is used to acquire the weighted matrix model of each document in multiple documents.

[0086] The third acquisition unit 503 is used to acquire the set of keywords for each document in multiple documents.

[0087] The calculation unit 504 is used to calculate the target similarity between each document and every other document in multiple documents, based on the topic term set of each document and the weighted matrix model of each document.

[0088] The fourth acquisition unit 505 is used to cluster multiple documents based on the target similarity using the PAM clustering algorithm to obtain a clustered document set.

[0089] The fifth acquisition unit 506 is used to acquire target information of multiple customers from the clustered document set.

[0090] The customer information processing apparatus provided in this application embodiment includes a first acquisition unit 501, which encodes customer information from multiple call orders from multiple customers to obtain multiple documents corresponding to the multiple call orders, wherein the multiple documents have the same encoding format; a second acquisition unit 502, which acquires a weighted matrix model for each document in the multiple documents; a third acquisition unit 503, which acquires a set of keywords for each document in the multiple documents; a calculation unit 504, which calculates the target similarity between each document and each other document in the multiple documents based on the set of keywords and the weighted matrix model of each document; a fourth acquisition unit 505, which clusters the multiple documents based on the target similarity using the PAM clustering algorithm to obtain a clustered document set; and a fifth acquisition unit 506, which acquires target information of multiple customers from the clustered document set. This solves the problem that extracting value information from call orders using manual annotation or natural language processing techniques leads to inaccurate information extracted from call orders. By calculating the weighted matrix model of each document in multiple call order forms and the keyword set of each document, the target similarity between multiple documents is calculated, thereby extracting target information from multiple documents. This algorithm automatically extracts target information from multiple documents, avoiding the impact of manual annotation of multiple documents on the information quality of target information, and achieving the effect of extracting more accurate target information from multiple call order forms.

[0091] Optionally, in the customer information processing apparatus provided in this application embodiment, the second acquisition unit 502 includes: a first processing subunit, used to delete preset strings in multiple documents and perform word segmentation and filtering processing on each of the multiple documents to obtain a first corpus; a first calculation subunit, used to substitute the first corpus into the word2vec model for calculation to obtain word vectors for each document; a second calculation subunit, used to substitute multiple documents into the TextRank model for calculation to obtain a keyword set for each document; a first acquisition subunit, used to sort the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword set of each other document, to obtain a keyword vector for each document; a third calculation subunit, used to replace the keywords in the keyword vector with word vectors and use the TF-IDF feature selection function to calculate the weights of the word vectors in the keyword vector, to obtain the replaced keyword vector and the word vector weights corresponding to the word vectors in the replaced keyword vector; and a second acquisition subunit, used to obtain a weighted matrix model for each document based on the replaced keyword vector and the word vector weights.

[0092] Optionally, in the customer information processing apparatus provided in this application embodiment, the second calculation subunit includes: a processing module, used to perform a first processing on each of the multiple documents to obtain words with preset parts of speech as candidate keywords for each document, wherein the first processing includes at least the following processing: sentence segmentation processing, word segmentation processing, filtering processing, and part-of-speech tagging processing; a first acquisition module, used to acquire a second corpus based on each document and the candidate keywords of each document; a first calculation module, used to convert the candidate keywords of each document into a keyword directed graph using the TextRank model, and calculate the weight of the candidate keywords in each document to obtain a weighted keyword directed graph for each document; and a second acquisition module, used to acquire candidate keywords with weights higher than preset weights from the weighted keyword directed graph to obtain a keyword set for each document.

[0093] Optionally, in the customer information processing apparatus provided in this application embodiment, the third acquisition unit 503 includes: a configuration subunit, used to configure the calculation formula for the word influence score of word vectors according to the TextRank model, the calculation formula being as follows: Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf a TF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. ba w represents the weights of keyword a and keyword b in a weighted directed graph. bc The first subunit represents the weights of keywords b and c in the weighted keyword directed graph; the second processing subunit is used to perform a second processing on each of the multiple documents to obtain the third corpus and the topic words in the third corpus, wherein the second processing includes at least the following processes: word segmentation, filtering, initial topic word extraction, and statistical word frequency processing; the fourth calculation subunit is used to iteratively calculate the word influence score of the topic words in the third corpus according to the calculation formula of word influence score. When the difference between the Nth iteration and the (N-1)th iteration is less than a preset threshold, the iterative calculation stops. From the calculation result of the Nth iteration, the topic words with word influence scores higher than the preset word influence scores are obtained to obtain the topic word set of each document.

[0094] Optionally, in the customer information processing apparatus provided in this application embodiment, the above-mentioned calculation unit 504 includes: a construction subunit, used to construct a bipartite graph model based on the topic word set of each document, so as to calculate a first similarity between the topic word set of each document and the topic word set of each other document; a fifth calculation subunit, used to calculate a second similarity between the weighted matrix model of each document and the weighted matrix model of each other document based on the least squares distance formula of the matrix; and a sixth calculation subunit, used to calculate a target similarity between each document and each other document based on the first similarity and the second similarity.

[0095] Optionally, in the customer information processing apparatus provided in the embodiments of this application, the first construction subunit includes: a construction module, configured to construct a bipartite graph model of each document and each other document based on the subject word set of each document; and a second calculation module, configured to calculate the maximum weight of bipartite matching between each document and each other document using the Hungarian algorithm, and use the maximum weight of bipartite matching as the first similarity between the subject word set of each document and the subject word set of each other document.

[0096] Optionally, in the customer information processing apparatus provided in this application embodiment, the fifth acquisition unit 506 includes: a third processing subunit, used to process each clustered document set according to the method of acquiring a keyword set, and acquire a keyword set for each clustered document set; a fourth processing subunit, used to process each clustered document set according to the method of acquiring a keyword set, and acquire a keyword set for each clustered document set; and a determination subunit, used to determine target information of multiple customers based on the keyword set and the keyword set of each clustered document set.

[0097] Optionally, in the customer information processing apparatus provided in the embodiments of this application, the apparatus further includes: a sixth acquisition unit, configured to calculate the target similarity between each document in multiple documents and each other document based on the topic word set and the weighted matrix model, and then add the target similarity between each document in multiple documents and each other document to obtain a third similarity for each document; a seventh acquisition unit, configured to acquire target documents from multiple documents whose third similarity is higher than a preset similarity to obtain a target document set; and a push unit, configured to push the target document set to a target object.

[0098] The customer information processing device includes a processor and a memory. The first acquisition unit 501, the second acquisition unit 502, the third acquisition unit 503, the calculation unit 504, the fourth acquisition unit 505, and the fifth acquisition unit 506 are all stored in the memory as program units. The processor executes the program units stored in the memory to realize the corresponding functions.

[0099] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can improve the accuracy of similarity measurements between multiple documents.

[0100] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0101] This invention provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a method for processing customer information.

[0102] This invention provides a processor for running a program, wherein the program executes a method for processing customer information during runtime.

[0103] like Figure 6 As shown, this embodiment of the invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: encoding customer information from multiple call tickets from multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; obtaining a weighted matrix model for each document in the multiple documents; obtaining a set of keywords for each document in the multiple documents; calculating the target similarity between each document and each other document based on the set of keywords and the weighted matrix model of each document; clustering the multiple documents based on the target similarity using the PAM clustering algorithm to obtain a clustered document set; and obtaining target information of multiple customers from the clustered document set.

[0104] The processor also performs the following steps when executing the program: obtaining a weighted matrix model for each document in multiple documents includes: deleting preset strings from multiple documents and performing word segmentation and filtering on each document to obtain a first corpus; substituting the first corpus into the word2vec model for calculation to obtain word vectors for each document; substituting multiple documents into the TextRank model for calculation to obtain a keyword set for each document; sorting the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword sets of other documents to obtain a keyword vector for each document; replacing the keywords in the keyword vector with word vectors and using the TF-IDF feature selection function to calculate the weights of the word vectors in the keyword vector to obtain the replaced keyword vector and the corresponding word vector weights in the replaced keyword vector; and obtaining a weighted matrix model for each document based on the replaced keyword vector and word vector weights.

[0105] When the processor executes the program, it also performs the following steps: Substituting multiple documents into the TextRank model for calculation to obtain the keyword set for each document includes: performing a first processing on each of the multiple documents to obtain words with preset parts of speech as candidate keywords for each document. The first processing includes at least the following: sentence segmentation, word segmentation, filtering, and part-of-speech tagging; obtaining a second corpus based on each document and its candidate keywords; using the TextRank model to convert the candidate keywords of each document into a directed keyword graph, and calculating the weight of each candidate keyword in each document to obtain a weighted directed keyword graph for each document; and obtaining candidate keywords with weights higher than preset weights from the weighted directed keyword graph to obtain the keyword set for each document.

[0106] When the processor executes the program, it also performs the following steps: obtaining the topic word set for each document in multiple documents, including: calculating the word influence score of word vectors based on the TextRank model, as shown in the following formula: Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf a TF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. baw represents the weights of keyword a and keyword b in a weighted directed graph. bc This represents the weights of keywords b and c in the weighted keyword directed graph. For each document in the multiple documents, a second processing step is performed to obtain a third corpus and the keywords within it. This second processing step includes at least the following steps: word segmentation, filtering, initial keyword extraction, and frequency analysis. Based on the formula for calculating keyword influence scores, the keyword influence scores of the keywords in the third corpus are iteratively calculated. The iteration stops when the difference between the Nth iteration and the (N-1th)th iteration is less than a preset threshold. From the results of the Nth iteration, keywords with keyword influence scores higher than the preset threshold are obtained, resulting in a keyword set for each document.

[0107] When the processor executes the program, it also performs the following steps: Calculating the target similarity between each document and every other document in multiple documents based on the topic word set of each document and the weighted matrix model of each document, including: constructing a bipartite graph model based on the topic word set of each document to calculate the first similarity between the topic word set of each document and the topic word set of every other document; calculating the second similarity between the weighted matrix model of each document and the weighted matrix model of every other document based on the least squares distance formula of the matrix; and calculating the target similarity between each document and every other document based on the first and second similarities.

[0108] The processor also performs the following steps when executing the program: constructing a bipartite graph model based on the topic word set of each document to calculate the first similarity between the topic word set of each document and the topic word set of every other document, including: constructing a bipartite graph model between each document and every other document based on the topic word set of each document; using the Hungarian algorithm to calculate the maximum weight of the bipartite matching between each document and every other document, and using the maximum weight of the bipartite matching as the first similarity between the topic word set of each document and the topic word set of every other document.

[0109] When the processor executes the program, it also performs the following steps: obtaining target information for multiple customers from the clustered document set, including: processing each clustered document set according to the method for obtaining the topic word set, and obtaining the topic word set for each clustered document set; processing each clustered document set according to the method for obtaining the keyword set, and obtaining the keyword set for each clustered document set; and determining the target information for multiple customers based on the topic word set and the keyword set of each clustered document set.

[0110] When the processor executes the program, it also performs the following steps: After calculating the target similarity between each document and every other document in multiple documents based on the topic word set of each document and the weighted matrix model of each document, the above method further includes: adding the target similarity between each document and every other document in multiple documents to obtain the third similarity of each document; obtaining target documents with a third similarity higher than a preset similarity from multiple documents to obtain a target document set; and pushing the target document set to the target object.

[0111] The devices mentioned in this article can be servers, PCs, tablets, mobile phones, etc.

[0112] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: encoding customer information from multiple call tickets for multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; obtaining a weighted matrix model for each document in the multiple documents; obtaining a set of keywords for each document in the multiple documents; calculating the target similarity between each document and each other document in the multiple documents based on the set of keywords and the weighted matrix model of each document; clustering the multiple documents based on the target similarity using the PAM clustering algorithm to obtain a clustered document set; and obtaining target information of multiple customers from the clustered document set.

[0113] When executed on a data processing device, it is also suitable to execute an initialization program with the following steps: obtaining a weighted matrix model for each document in multiple documents, including: deleting preset strings from multiple documents and performing word segmentation and filtering on each document to obtain a first corpus; substituting the first corpus into a word2vec model for calculation to obtain word vectors for each document; substituting multiple documents into a TextRank model for calculation to obtain a keyword set for each document; sorting the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword sets of other documents to obtain a keyword vector for each document; replacing the keywords in the keyword vector with word vectors and calculating the weights of the word vectors in the keyword vector using a TF-IDF feature selection function to obtain the replaced keyword vector and the corresponding word vector weights in the replaced keyword vector; and obtaining a weighted matrix model for each document based on the replaced keyword vector and word vector weights.

[0114] When executed on a data processing device, it is also suitable to execute an initialization program with the following steps: Substituting multiple documents into the TextRank model for calculation to obtain a keyword set for each document, including: performing a first processing on each of the multiple documents to obtain words with preset parts of speech as candidate keywords for each document, wherein the first processing includes at least the following processing: sentence segmentation, word segmentation, filtering, and part-of-speech tagging; obtaining a second corpus based on each document and its candidate keywords; using the TextRank model to convert the candidate keywords of each document into a directed keyword graph, and calculating the weight of the candidate keywords in each document to obtain a weighted directed keyword graph for each document; from the weighted directed keyword graph, obtaining candidate keywords with weights higher than preset weights to obtain a keyword set for each document.

[0115] When executed on a data processing device, it is also suitable to execute a program that initializes with the following method steps: obtaining the topic word set for each document in multiple documents, including: calculating the word influence score of word vectors based on the TextRank model, as shown in the following formula: Where d represents the preset damping coefficient, a represents keyword a in the weighted keyword directed graph, Out(b) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, In(a) represents the set of all keywords pointed to by keyword a in the weighted keyword directed graph, b represents keyword b pointing to keyword a, c represents keyword c pointed to by keyword a, S(b) represents the word influence of keyword b, tfidf a TF-IDF represents the product of the term frequency and inverse document frequency of keyword 'a'. c w represents the product of the term frequency and inverse document frequency of keyword c. ba w represents the weights of keyword a and keyword b in a weighted directed graph. bc This represents the weights of keywords b and c in the weighted keyword directed graph. For each document in the multiple documents, a second processing step is performed to obtain a third corpus and the keywords within it. This second processing step includes at least the following steps: word segmentation, filtering, initial keyword extraction, and frequency analysis. Based on the formula for calculating keyword influence scores, the keyword influence scores of the keywords in the third corpus are iteratively calculated. The iteration stops when the difference between the Nth iteration and the (N-1th)th iteration is less than a preset threshold. From the results of the Nth iteration, keywords with keyword influence scores higher than the preset threshold are obtained, resulting in a keyword set for each document.

[0116] When executed on a data processing device, it is also suitable to execute an initialization program with the following method steps: calculating the target similarity between each document and every other document in a plurality of documents based on the topic word set of each document and the weighted matrix model of each document, including: constructing a bipartite graph model based on the topic word set of each document to calculate the first similarity between the topic word set of each document and the topic word set of every other document; calculating the second similarity between the weighted matrix model of each document and the weighted matrix model of every other document based on the least squares distance formula of the matrix; and calculating the target similarity between each document and every other document based on the first similarity and the second similarity.

[0117] When executed on a data processing device, it is also suitable to execute an initialization procedure with the following steps: constructing a bipartite graph model based on the topic word set of each document, and calculating the first similarity between the topic word set of each document and the topic word sets of all other documents, including: constructing a bipartite graph model of each document and all other documents based on the topic word set of each document; calculating the maximum weight of bipartite matching between each document and all other documents using the Hungarian algorithm, and using the maximum weight of bipartite matching as the first similarity between the topic word set of each document and the topic word sets of all other documents.

[0118] When executed on a data processing device, it is also suitable to execute an initialization program with the following method steps: obtaining target information for multiple customers from a clustered document set, including: processing each clustered document set according to the method for obtaining the topic word set, and obtaining the topic word set for each clustered document set; processing each clustered document set according to the method for obtaining the keyword set, and obtaining the keyword set for each clustered document set; and determining the target information for multiple customers based on the topic word set and the keyword set of each clustered document set.

[0119] When executed on a data processing device, it is also suitable to execute an initialization program with the following steps: after calculating the target similarity between each document and every other document in multiple documents based on the topic word set of each document and the weighted matrix model of each document, the above method further includes: adding the target similarity between each document and every other document in multiple documents to obtain the third similarity of each document; obtaining target documents with a third similarity higher than a preset similarity from multiple documents to obtain a target document set; and pushing the target document set to the target object.

[0120] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0125] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0126] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0127] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0128] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for processing customer information, characterized in that, include: Customer information from multiple call tickets for multiple customers is encoded to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format; Obtain the weighted matrix model for each of the multiple documents; Obtain the set of keywords for each of the multiple documents; Based on the topic word set of each document and the weighted matrix model of each document, the target similarity between each document and each of the other documents is calculated. Based on the target similarity, the multiple documents are clustered using the PAM clustering algorithm to obtain a clustered document set; Obtain target information for multiple customers from the clustered document set; Obtaining the weighted matrix model for each of the multiple documents includes: Delete the preset strings in the multiple documents, and perform word segmentation and filtering on each of the multiple documents to obtain the first corpus; The first corpus is substituted into the word2vec model for calculation to obtain the word vector of each document; Substitute the multiple documents into the TextRank model for calculation to obtain the keyword set for each document; Based on the similarity between the keyword set of each document and the keyword set of every other document, a greedy selection algorithm is used to sort the keywords in the keyword set of each document to obtain the keyword vector of each document; The keyword vector is replaced with the word vector, and the weights of the word vectors in the keyword vector are calculated using the TF-IDF feature selection function to obtain the replaced keyword vector and the word vector weights corresponding to the word vectors in the replaced keyword vector. Based on the replaced keyword vectors and the word vector weights, obtain the weighted matrix model for each document; Obtaining the topic keyword set for each of the multiple documents includes: The formula for calculating the word influence score of word vectors based on the TextRank model is shown below: Where d represents the preset damping coefficient, and a represents keyword a in the weighted keyword directed graph. This indicates the influence of keyword 'a'. Let represent the set of all keywords that keyword 'a' points to in the weighted keyword directed graph. Let represent the set of all keywords pointing to keyword 'a' in the weighted keyword directed graph, b represent keyword b pointing to keyword 'a', and c represent keyword c pointing to by keyword 'a'. This indicates the influence of keyword b. This represents the product of the term frequency and inverse document frequency of keyword 'a'. This represents the product of the term frequency and inverse document frequency of keyword c. This represents the weights of keyword a and keyword b in the weighted keyword directed graph. This represents the weights of keyword b and keyword c in the weighted keyword directed graph; wherein the weighted keyword directed graph is obtained through the following steps: using the TextRank model to convert the candidate keywords of each document into a keyword directed graph, and calculating the weights of the candidate keywords in each document to obtain the weighted keyword directed graph of each document; For each of the plurality of documents, a second processing is performed to obtain a third corpus and the topic words in the third corpus, wherein the second processing includes at least the following processing: word segmentation processing, filtering processing, initial topic word extraction processing, and word frequency statistics processing; Based on the formula for calculating the word influence score, the word influence score of the topic words in the third corpus is calculated iteratively. When the difference between the Nth iteration and the (N-1)th iteration is less than a preset threshold, the iterative calculation is stopped. From the calculation result of the Nth iteration, the topic words whose word influence scores are higher than the preset word influence scores are obtained, and the topic word set of each document is obtained.

2. The method according to claim 1, characterized in that, Substituting the multiple documents into the TextRank model for calculation, the keyword set for each document is obtained as follows: For each of the plurality of documents, a first processing is performed to obtain words with preset parts of speech as candidate keywords for each document. The first processing includes at least the following processes: sentence segmentation, word segmentation, filtering, and part-of-speech tagging. A second corpus is obtained based on each document and its candidate keywords. The candidate keywords for each document are converted into a directed keyword graph using the TextRank model, and the weights of the candidate keywords in each document are calculated to obtain a weighted directed keyword graph for each document. From the weighted keyword directed graph, candidate keywords with weights higher than preset weights are obtained to obtain the keyword set for each document.

3. The method according to claim 1, characterized in that, Based on the keyword set of each document and the weighted matrix model of each document, the target similarity between each document and every other document is calculated, including: A bipartite graph model is constructed based on the keyword set of each document to calculate the first similarity between the keyword set of each document and the keyword set of each other document. Based on the least squares distance formula of the matrix, calculate the second similarity between the weighted matrix model of each document and the weighted matrix model of each other document; Based on the first similarity and the second similarity, the target similarity between each document and every other document is calculated.

4. The method according to claim 3, characterized in that, Constructing a bipartite graph model based on the keyword set of each document, and calculating the first similarity between the keyword set of each document and the keyword sets of all other documents, includes: Based on the keyword set of each document, construct the bipartite graph model of each document and every other document; The Hungarian algorithm is used to calculate the maximum weight of the bipartite matching between each document and every other document, and the maximum weight of the bipartite matching is used as the first similarity between the topic word set of each document and the topic word set of every other document.

5. The method according to claim 1, characterized in that, Obtaining target information for multiple customers from the clustered document set includes: Based on the method for obtaining the keyword set, each clustered document set is processed to obtain the keyword set for each clustered document set; Based on the method for obtaining the keyword set, each clustered document set is processed to obtain the keyword set for each clustered document set; Based on the topic word set and keyword set of each clustered document set, the target information of the multiple customers is determined.

6. The method according to claim 1, characterized in that, After calculating the target similarity between each document and every other document based on the topic term set of each document and the weighted matrix model of each document, the method further includes: The target similarity between each document and each of the multiple documents is added together to obtain the third similarity of each document; From the plurality of documents, target documents with a third similarity higher than a preset similarity are obtained to obtain a target document set; Push the target document set to the target object.

7. A customer information processing device, characterized in that, include: The first acquisition unit is used to encode customer information from multiple call tickets of multiple customers to obtain multiple documents corresponding to the multiple call tickets, wherein the multiple documents have the same encoding format. The second acquisition unit is used to acquire the weighted matrix model of each of the plurality of documents; The third acquisition unit is used to acquire the keyword set of each of the plurality of documents; The calculation unit is used to calculate the target similarity between each document and each of the other documents based on the topic word set of each document and the weighted matrix model of each document; The fourth acquisition unit is used to cluster the multiple documents based on the target similarity using the PAM clustering algorithm to obtain a clustered document set. The fifth acquisition unit is used to acquire target information of multiple customers from the clustered document set; The second acquisition unit includes: a first processing subunit, used to delete preset strings from the plurality of documents and perform word segmentation and filtering on each of the plurality of documents to obtain a first corpus; a first calculation subunit, used to substitute the first corpus into a word2vec model for calculation to obtain word vectors for each document; a second calculation subunit, used to substitute the plurality of documents into a TextRank model for calculation to obtain a keyword set for each document; a first acquisition subunit, used to sort the keywords in the keyword set of each document using a greedy selection algorithm based on the similarity between the keyword set of each document and the keyword sets of other documents, to obtain a keyword vector for each document; a third calculation subunit, used to replace the keywords in the keyword vector with the word vectors and use a TF-IDF feature selection function to calculate the weights of the word vectors in the keyword vectors, to obtain the replaced keyword vectors and the corresponding word vector weights in the replaced keyword vectors; and a second acquisition subunit, used to obtain a weighted matrix model for each document based on the replaced keyword vectors and the word vector weights. The third acquisition unit includes a configuration subunit, used to configure the calculation formula for the word influence score of word vectors according to the TextRank model, as shown in the following formula: Where d represents the preset damping coefficient, and a represents keyword a in the weighted keyword directed graph. This indicates the influence of keyword 'a'. Let represent the set of all keywords that keyword 'a' points to in the weighted keyword directed graph. Let represent the set of all keywords pointing to keyword 'a' in the weighted keyword directed graph, b represent keyword b pointing to keyword 'a', and c represent keyword c pointing to by keyword 'a'. This indicates the influence of keyword b. This represents the product of the term frequency and inverse document frequency of keyword 'a'. This represents the product of the term frequency and inverse document frequency of keyword c. This represents the weights of keyword a and keyword b in the weighted keyword directed graph. This represents the weights of keywords b and c in the weighted keyword directed graph; wherein the weighted keyword directed graph is obtained through the following steps: using the TextRank model to convert the candidate keywords of each document into a keyword directed graph, and calculating the weights of the candidate keywords in each document to obtain the weighted keyword directed graph of each document; a second processing subunit is used to perform a second processing on each of the multiple documents to obtain a third corpus and the topic words in the third corpus, wherein the second processing includes at least the following processing: word segmentation processing, filtering processing, initial topic word extraction processing, and statistical word frequency processing; a fourth calculation subunit is used to iteratively calculate the word influence score of the topic words in the third corpus according to the calculation formula of the word influence score, and stop the iterative calculation when the difference between the Nth iteration calculation and the N-1th iteration is less than a preset threshold, and obtain the topic words whose word influence scores are higher than the preset word influence scores from the calculation results of the Nth iteration calculation to obtain the topic word set of each document.

8. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the customer information processing method of any one of claims 1 to 6.

Images