A customer service interface docking management system

Abstract

The present invention relates to the technical field of interface management, and discloses a customer service interface docking management system, including: obtaining a first conversation text within a first preset time period; preprocessing the first conversation text and mapping it to a word vector space to obtain a text vector; determining the logical entities of the system to be integrated, and using the database tables corresponding to the logical entities as true labels; training an adversarial network model based on the true samples and the text vector; extracting the generator and discriminator of the adversarial generation model within a second preset time period; performing translation on the target conversation text based on the generator to obtain the translation data of the integrated system; constructing a rule mapping processing unit based on the second conversation text and the translation database table; and reversely optimizing the mapping relationship of the rule mapping processing unit based on the discriminator.

Description

Technical Field

[0001] The present invention relates to the technical field of interface management, and more specifically, it relates to a customer service interface docking management system. Background Art

[0002] In an intelligent customer service system, the conversations between the customer service and users often contain rich information. With the expansion of the platform's business, the intelligent customer service needs to be integrated with various business systems to fully explore and store diverse data. Traditional system integration solutions usually rely on a dedicated system docking team to develop docking interfaces to achieve system integration, resulting in high system docking costs.

[0003] An integration interface of a form management information system with the publication number CN102982419B discloses an integration interface of a form management information system, including three types of interface modules: single sign-on and page jump, interface call, and data access and push. It can avoid re-login and jump pages, facilitating operation. It can initiate cross-system business processes in this system, and the data is updated and transferred in a timely manner. After transformation, it can be seamlessly connected with the production information system, the two-ticket expert system, and the mobile terminal to achieve system data sharing, collaborative management, and process integration.

[0004] The integration interface of this form information management system integrates multiple systems through the interface. However, if new business systems are generated for re-integration as the business system expands, this integration interface is difficult to adapt to the integration of business systems and there is a problem of difficult integration. Summary of the Invention

[0005] The present invention provides a customer service interface docking management system to solve the technical problems raised in the background art.

[0006] The present invention provides a customer service interface docking management system, including:

[0007] A data collection module, which is used to obtain a first conversation text within a first preset time period; preprocess the first conversation text and map it to a word vector space to obtain a text vector;

[0008] An adversarial generation module, which is used to determine the logical entities of the systems to be integrated and use the database tables corresponding to the logical entities as real labels; train an adversarial network model based on real samples and text vectors;

[0009] A translation management module, which is used to obtain a second conversation text within a second preset time period and perform a translation operation on the second conversation text, including:

[0010] Extract the generator and discriminator of the adversarial generation model;

[0011] Perform translation on the target dialogue text based on the generator to obtain the translation data of the integrated system;

[0012] Construct a rule mapping processing unit based on the second dialogue text and the translation database table;

[0013] Based on the discriminator, reverse-optimize the mapping relationship of the rule mapping processing unit;

[0014] The mapping management module is used to obtain the third dialogue text within the third preset time period and perform a translation operation on the third dialogue text, including:

[0015] Classify the third dialogue text into: rule text and non-rule text;

[0016] If it is rule text, perform translation processing on the third dialogue text based on the mapping processing unit;

[0017] If it is non-rule text, perform translation processing on the third dialogue text based on the generator.

[0018] Furthermore, preprocess the first dialogue text and map it to the word vector space to obtain a text vector;

[0019] The preprocessing includes: text cleaning to remove special characters, punctuation marks, and stop words in the first dialogue text to obtain the preprocessed text of the first dialogue text;

[0020] Map the preprocessed text to the word vector space to obtain an initial text vector;

[0021] Perform convolution processing on the initial text vector based on a one-dimensional convolutional neural network to obtain a text vector.

[0022] Furthermore, train an adversarial network model based on real samples and text vectors;

[0023] The adversarial network model includes: a generator and a discriminator;

[0024] Input the text vector into the generator to obtain the translation database table, and input the real label into the discriminator to determine the real probability for the translation database table;

[0025] If the absolute difference between the mean of the real probabilities for consecutive P times and the preset convergence threshold is less than or equal to the preset difference threshold, the training converges to obtain the adversarial network model.

[0026] Furthermore, construct a rule mapping processing unit based on the second dialogue text and the translation database table, including:

[0027] After preprocessing the second dialogue text, perform word segmentation on the preprocessed text based on a Chinese word segmentation tool to obtain M segmented fields;

[0028] Extract the values of N fields from the translation database table;

[0029] Establish the mapping relationship between M word segmentation fields and the values of N fields, specifically including:

[0030] Initialize and generate a mapping scheme that meets the constraint conditions; among them, the mapping scheme includes the mapping relationship between each field value and several word segmentation fields;

[0031] The constraint conditions include: each field value is mapped to at least one word segmentation field, each field value is mapped to at most M word segmentation fields, and each field value and each word segmentation field can establish a mapping relationship only once;

[0032] Construct a rule mapping processing unit based on the mapping scheme.

[0033] Furthermore, reverse-optimize the mapping relationship of the rule mapping processing unit based on the discriminator, including determining the score of the rule mapping processing unit, as follows:

[0034] Generate a corresponding rule translation database table for the second dialogue text based on the rule mapping processing unit;

[0035] Determine the true probability of the rule translation database table based on the discriminator of the adversarial network model;

[0036] Take the absolute difference between the true probability and the preset convergence threshold as the score of the rule mapping processing unit;

[0037] If the absolute difference between the score and the preset convergence threshold is greater than the preset difference threshold, update the mapping relationship of the mapping scheme based on the gradient descent method until the absolute difference between the score and the preset convergence threshold is less than or equal to the preset difference threshold.

[0038] Furthermore, classify the third dialogue text into: rule text and non-rule text, including:

[0039] If the absolute difference between the score of the rule mapping processing unit and the preset convergence threshold is less than or equal to the preset difference threshold, then regard the corresponding second dialogue text as the rule text;

[0040] Preprocess, perform word segmentation, and part-of-speech tagging on both the second dialogue text and the third dialogue text;

[0041] Among them, the part-of-speech tagging obtains the part of speech of each word segmentation field corresponding to the second dialogue text and the third dialogue text based on the part-of-speech tagging tool;

[0042] Based on the position order of the segmented fields in the second dialogue text and the third dialogue text, corresponding part-of-speech sequences are constructed respectively; where the i-th unit of the part-of-speech sequence represents the part of speech of the i-th segmented field in the second dialogue text or the third dialogue text.

[0043] Calculate the similarity value of the corresponding part-of-speech sequence, specifically including:

[0044] Load a sliding extraction window for converting the part-of-speech sequence into a convolutional sequence with dimension H; where the length of the sliding extraction window is dynamically set based on the dimension of the part-of-speech sequence.

[0045] Calculate the similarity value of the part-of-speech combination of the h-th unit in the corresponding convolutional sequence respectively; where the part-of-speech combination is obtained by extracting the part-of-speech sequence using the sliding extraction window.

[0046] If the similarity value is greater than or equal to the preset similarity threshold, the third dialogue text is classified as a regular text; otherwise, the third dialogue text is classified as an irregular text.

[0047] Further, calculating the similarity value of the part-of-speech combination of the h-th unit in the corresponding convolutional sequence includes: ;

[0048] where 1 ≤ h ≤ H, h is a positive integer, represents the similarity value of the part-of-speech sequence, represents the number of parts of speech in the part-of-speech combination of the h-th unit in the convolutional sequence corresponding to the second dialogue text, represents the number of parts of speech in the part-of-speech combination of the h-th unit in the convolutional sequence corresponding to the third dialogue text, represents the number of the same parts of speech in the part-of-speech combination of the h-th unit in the convolutional sequence corresponding to the second dialogue text and the part-of-speech combination of the h-th unit in the convolutional sequence corresponding to the third dialogue text.

[0049] Further, the word vector space is the Word2Ve word vector space.

[0050] The beneficial effects of the present invention are as follows:

[0051] 1. Through the dual interface management solutions of rule mapping translation and generator dynamic translation, an intelligent docking mechanism covering all scenarios is constructed. For the standardized text (rule text) that appears frequently in customer service conversations, the system relies on the pre-trained rule mapping processing unit to achieve rapid translation based on the mapping relationship between historical conversation word segmentation and business database fields, significantly improving the efficiency of standardized interface docking; and for non-regular texts with complex sentences and changeable semantics, the system dynamically generates adaptation data through the generator of the adversarial generation module, and uses the model generalization ability to solve docking problems. With the dual strategy of rule priority and model guarantee, the customer service interface docking management system can not only efficiently handle repetitive scenarios through solidified rules, but also flexibly respond to personalized needs with the help of generation models, providing a more universal docking management solution for multi-business system integration.

[0052] 2. In the customer service interface docking management system, the generator and discriminator of the adversarial network are creatively decoupled and used in collaboration to build a rule system: the generator generates a data table structure that fits the business system based on the conversation text vector, providing an initial mapping target for the interface docking; the discriminator uses the real business database table as a label, reversely evaluates the accuracy of the rule mapping, and dynamically optimizes the mapping relationship between the word segmentation field and the database field through gradient descent. Through the closed-loop mechanism of "generation-discrimination-optimization", the system can automatically learn the mapping rules between customer service conversation text and business data without relying on manual processing of preset rules, and build a highly adaptive rule mapping processing unit. Compared with the single generation function of the traditional adversarial network, this design gives the customer service interface docking management system the dual capabilities of "rule self-construction" and "dynamic optimization", which not only greatly reduces the cost of manual configuration, but also realizes the precise structured conversion from conversation text to business data through the deep integration of models and rules. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a module diagram of a customer service interface docking management system of the present invention. DETAILED DESCRIPTION

[0054] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0055] like Figure 1 As shown, a customer service interface docking management system includes:

[0056] A data acquisition module, configured to obtain a first dialogue text within a first preset time period; preprocess the first dialogue text and map it to a word vector space to obtain a text vector;

[0057] An adversarial generation module, configured to determine the logical entities of the system to be integrated and use the database tables corresponding to the logical entities as real labels; train an adversarial network model based on real samples and text vectors;

[0058] A translation management module, configured to obtain a second dialogue text within a second preset time period and perform a translation task on the second dialogue text, including:

[0059] Extract the generator and discriminator of the adversarial generation model;

[0060] Perform translation on the target dialogue text based on the generator to obtain the translation data of the integrated system;

[0061] Construct a rule mapping processing unit based on the second dialogue text and the translation database table;

[0062] Reverse-optimize the mapping relationship of the rule mapping processing unit based on the discriminator;

[0063] A mapping management module, configured to obtain a third dialogue text within a third preset time period and perform a translation task on the third dialogue text, including:

[0064] Classify the third dialogue text into: rule text and non-rule text;

[0065] If it is rule text, perform translation processing on the third dialogue text based on the mapping processing unit;

[0066] If it is non-rule text, perform translation processing on the third dialogue text based on the generator.

[0067] It should be noted that the system to be integrated refers to a business system to be integrated into the customer service system. Among them, the logical entity refers to an abstract data model in the system to be integrated for describing the core business object, usually corresponding to a table in the database. For example, if the system to be integrated is an e-commerce order system, the logical entities include: order entity or logistics entity; further, the order entity corresponds to the database table order_info, and the logistics entity corresponds to the database table logistics_info.

[0068] In an embodiment of the present invention, the first dialogue text is preprocessed and mapped to a word vector space to obtain a text vector;

[0069] The preprocessing includes: text cleaning to remove special characters, punctuation marks, and stop words in the first dialogue text to obtain the preprocessed text of the first dialogue text;

[0070] Map the preprocessed text to the word vector space to obtain the initial text vector;

[0071] Perform convolution processing on the initial text vector based on a one-dimensional convolutional neural network to obtain the text vector.

[0072] It should be noted that text cleaning is used to remove invalid information in the original dialogue text, retain the core semantic content, and reduce the interference of noise on subsequent semantic analysis. Specifically, it includes: matching and removing non-semantic characters such as emoji, format control characters, and illegal characters (such as "@#￥%") in the dialogue text through regular expressions; deleting punctuation marks such as full stops, commas, question marks, and exclamation marks (for example, converting "When will my order arrive?" to "When will my order arrive") to eliminate the impact of punctuation on the accuracy of word segmentation; filtering out stop words in the text through dictionary matching based on a predefined stop word list (such as high-frequency meaningless words like "of", "already", "in", "excuse me", etc.). For example, cleaning "Hello, excuse me, how is my order status?" to "order status" and retaining the core keywords. Technical value: Through standardized cleaning, the original dialogue text is converted into a "pure text" that only contains valid semantic vocabulary, providing high-quality input for subsequent word segmentation and semantic representation.

[0073] It should be noted that the one-dimensional convolutional neural network processing is used to extract local semantic features and context dependencies in the text through convolution operations, generating a text vector containing deep semantic information, including:

[0074] Take the initial text vector sequence as the input of the one-dimensional convolutional neural network, whose dimension is [number of samples × sequence length × word vector dimension]; use convolutional kernels with multiple different window sizes (such as window sizes 3, 5, 7) to perform sliding convolution on the input vector to extract n-gram level semantic features (such as a window size of 3 can capture the combined semantics of 3 consecutive words, such as "status query" in "order status query"); introduce non-linear transformation through the ReLU activation function, and then compress the feature dimension through a pooling layer (such as max pooling). Finally, concatenate the feature vectors output by each convolutional kernel to generate a text vector with a fixed dimension (such as a vector with a dimension of 300), which contains the local semantic features and global context information of the text. Different from the traditional method that only uses the average of word vectors, the one-dimensional convolutional neural network can automatically capture the combined semantics between words (such as "logistics exception" as an overall semantic unit), improve the semantic representation ability of the text vector, and provide more effective input features for the subsequent adversarial generation module.

[0075] In an embodiment of the present invention, an adversarial network model is trained based on real samples and text vectors;

[0076] The adversarial network model includes: a generator and a discriminator;

[0077] Input the text vector into the generator to obtain the translated database table, and input the true label into the discriminator to determine the true probability for the translated database table;

[0078] If the absolute difference between the mean of the true probabilities for consecutive P times and the preset convergence threshold is less than or equal to the preset difference threshold, the training converges, and an adversarial network model is obtained.

[0079] It should be noted that within the first preset time period, multiple first dialogue texts are obtained, so as to construct a sufficient number of text vectors as the input to the generator.

[0080] It should be noted that the adversarial network model architecture includes a generator and a discriminator.

[0081] The input of the generator is the text vector output by the data acquisition module, and the output is the "translated database table", that is, the structured data generated according to the text vector and corresponding to the logical entity of the system to be integrated.

[0082] The goal of the generator: By learning the mapping relationship between the text vector and the structure of the true database table, generate a translated database table that is as close as possible to the true label, so that the true probability output by the discriminator is close to 1.

[0083] The discriminator: The input is the "true label" (that is, the structure of the true database table corresponding to the logical entity of the system to be integrated, such as the original table fields and data formats of the order management system) and the "translated database table" output by the generator, and the output is a probability value, and the value range of the probability value is between 0 and 1, indicating the confidence of the discriminator in whether the input table structure "belongs to the data of the true business system".

[0084] The goal of the discriminator: Accurately distinguish the true label from the translated table output by the generator, output a high probability (close to 1) for the true label, and output a low probability (close to 0) for the generated table.

[0085] The generator is used to receive the text vector and output the translated database table. The discriminator simultaneously receives the true label and the translated database table and outputs the corresponding true probability.

[0086] The loss function includes the generator loss and the discriminator loss:

[0087] Generator loss: Maximize the misjudgment probability of the discriminator for the generated table through gradient ascent. Among them, represents the generator loss, represents the true probability of the translated database table, represents the translated database table, An operation that takes the expectation of the true probabilities for all translated database tables, used to perform statistical averaging on the true probabilities to measure the performance of the model over the overall data distribution.

[0088] Discriminator loss: Minimize the misjudgment probability of the true label and the missed judgment probability of the generated table through gradient descent. Among them, Represents the discriminator loss, Represents the true probability of the true label, Represents the true label.

[0089] Alternating optimization: The generator and the discriminator update their parameters alternately according to a preset number of rounds, forming an adversarial training loop of "generate - discriminate - feedback".

[0090] Training convergence conditions, including a preset convergence threshold, a preset difference threshold, and a consecutive number. The preset convergence threshold θ: Represents the expected true probability of the discriminator for the true label. For example, if set to 0.95, it means that the discriminator is expected to have a judgment probability of no less than 95% for the true label. The preset difference threshold : Represents the maximum deviation (such as 0.01) allowed between the mean true probability and the preset convergence threshold. The consecutive number P: Represents the consecutive valid training rounds used to judge convergence. For example, P = 10, that is, if the conditions are met for 10 consecutive trainings, it is considered convergent. Calculate the mean true probability Q of the discriminator for the generated table in P consecutive trainings. If is satisfied, it is considered that the distribution difference between the translated table output by the generator and the true label has been reduced to an acceptable range, the training process terminates, and the final adversarial network model is obtained.

[0091] In an embodiment of the present invention, based on the second dialogue text and the translated database table, a rule mapping processing unit is constructed, including:

[0092] After preprocessing the second dialogue text, perform word segmentation on the preprocessed text based on a Chinese word segmentation tool to obtain M word segmentation fields;

[0093] Extract N field values from the translated database table;

[0094] Establish a mapping relationship between the M word segmentation fields and the N field values, specifically including:

[0095] Initialize a mapping scheme that meets the constraint conditions; among them, the mapping scheme includes the mapping relationship between each field value and several word segmentation fields;

[0096] The constraint conditions include: Each field value is mapped to at least one word segmentation field, each field value is mapped to at most M word segmentation fields, and each field value and each word segmentation field can be mapped only once;

[0097] Construct a rule mapping processing unit based on a mapping scheme.

[0098] It should be noted that a Chinese word segmentation tool, including but not limited to: jieba, THULAC and other Chinese word segmentation tools, is used to segment the preprocessed text, splitting it into M independent word segmentation fields. For example, for the sentence "Query the logistics status of order 123", after word segmentation, it may obtain M word segmentation fields such as "query", "order", "123", "logistics", "status".

[0099] Extract N field values from the translation database table. The translation database table is structured data generated by the generator in the adversarial network model according to the text vector and corresponding to the logical entities of the system to be integrated. For example, for the order_id in the order table, the corresponding field value is 123, indicating the 123rd order.

[0100] Generate an initial mapping scheme according to specific constraint conditions. The mapping scheme describes the correspondence between each field value and several word segmentation fields. Among them, each field value has at least one mapping relationship with a word segmentation field: ensure that each field value in the translation database table can find the corresponding semantic information source from the second dialogue text to ensure the integrity of the mapping. For example, the order_id field value should have at least one mapping with word segmentation fields such as "order" and "123", otherwise there will be no corresponding information for this field value in the dialogue. Each field value has at most M mapping relationships with word segmentation fields, which limits the number of word segmentation fields associated with a field value to avoid confusion in the mapping relationship caused by excessive association. Each field value and each word segmentation field can establish a mapping relationship only once to ensure the uniqueness of the mapping, avoid having multiple conflicting mappings for a field value or a word segmentation field, and ensure the accuracy and certainty of the mapping relationship.

[0101] Construct a rule mapping processing unit based on the generated mapping scheme. This unit will serve as a core component, and during the subsequent customer service interface docking process, it can accurately convert the natural language information in the customer service dialogue into structured data recognizable by the business system according to the established mapping relationship, achieving efficient interface docking.

[0102] In an embodiment of the present invention, the mapping relationship of the rule mapping processing unit is reversely optimized based on the discriminator, including determining the score of the rule mapping processing unit as follows:

[0103] Generate a corresponding rule translation database table based on the rule mapping processing unit for the second dialogue text;

[0104] Determine the true probability of the rule translation database table based on the discriminator of the adversarial network model;

[0105] Take the absolute difference between the true probability and the preset convergence threshold as the score of the rule mapping processing unit;

[0106] If the absolute difference between the score and the preset convergence threshold is greater than the preset difference threshold, update the mapping relationship of the mapping scheme based on the gradient descent method until the absolute difference between the score and the preset convergence threshold is less than or equal to the preset difference threshold.

[0107] It should be noted that the generation of the rule translation database table includes:

[0108] Input: The second dialogue text (M tokenized fields after preprocessing and tokenization) and the current mapping scheme of the rule mapping processing unit (the corresponding relationship between the tokenized fields and the database fields);

[0109] Output: The rule translation database table (structured data), that is, map the tokenized fields in the dialogue text to the N field values of the translation database table according to the mapping scheme to form a data table structure recognizable by the business system. For example, in the order query scenario, map "order 123" to order_id = 123.

[0110] Based on the corresponding relationship between the "tokenized fields and field values" in the mapping scheme, extract the valid information in the dialogue text and fill it into the corresponding fields of the translation database table to realize the conversion from natural language text to structured data.

[0111] Use the discriminator trained by the adversarial network model to evaluate the authenticity of the generated rule translation database table, and output the matching probability (value range [0, 1]) between the table and the real database table (real label) of the system to be integrated, that is, the true probability. This probability reflects the degree of coincidence between the rule translation table and the real table in terms of field composition, data format, business constraints, etc. (such as no missing fields, correct data types, and compliance with business logic rules).

[0112] The score is determined by the absolute difference between the true probability and the preset convergence threshold (denoted as θ, such as 0.95), that is: . The smaller the score, the higher the matching degree between the rule translation table and the real table, and the better the performance of the rule mapping processing unit.

[0113] If the score is greater than the preset difference threshold (denoted as , such as 0.01), that is , it means that the difference between the rule translation table generated by the current mapping scheme and the real table exceeds the acceptable range, and the mapping relationship needs to be optimized.

[0114] In one embodiment of the present invention, updating the mapping relationship of the mapping scheme based on the gradient descent method includes: initializing the association weights between each field value and the M word segmentation fields, where the values of the association weights range from 0 to 1, and establishing a mapping relationship between the association weights greater than or equal to a preset association threshold and the field values. Calculating the gradient of the score with respect to the association weights by the gradient descent method to adjust the mapping parameters in the opposite direction of the gradient and dynamically update the association weights between the word segmentation fields and the field values.

[0115] In one embodiment of the present invention, classifying the third dialogue text into: regular text and irregular text, including:

[0116] If the absolute difference between the score of the rule mapping processing unit and the preset convergence threshold is less than or equal to the preset difference threshold, then the corresponding second dialogue text is regarded as regular text;

[0117] Performing preprocessing, word segmentation processing, and part-of-speech tagging on both the second dialogue text and the third dialogue text;

[0118] Among them, part-of-speech tagging is used to obtain the part of speech of each word segmentation field corresponding to the second dialogue text and the third dialogue text based on a part-of-speech tagging tool;

[0119] Based on the order of positions of the word segmentation fields in the second dialogue text and the third dialogue text, corresponding part-of-speech sequences are constructed respectively; among them, the i-th unit of the part-of-speech sequence represents the part of speech of the i-th word segmentation field in the second dialogue text or the third dialogue text;

[0120] Calculating the similarity value of the corresponding part-of-speech sequences, specifically including:

[0121] Loading a sliding extraction window for converting the part-of-speech sequence into a convolutional sequence with a dimension of H; where the length of the sliding extraction window is dynamically set based on the dimension of the part-of-speech sequence;

[0122] Calculating the similarity value of the part-of-speech combinations of the h-th unit in the corresponding convolutional sequences respectively; among them, the part-of-speech combinations are obtained by the sliding extraction window for the part-of-speech sequence;

[0123] If the similarity value is greater than or equal to the preset similarity threshold, then the third dialogue text is classified as regular text; otherwise, the third dialogue text is classified as irregular text.

[0124] In one embodiment of the present invention, calculating the similarity value of the part-of-speech combinations of the h-th unit in the corresponding convolutional sequences includes:

[0125] ;

[0126] Among them, 1 ≤ h ≤ H, h is a positive integer, represents the similarity value of the part-of-speech sequence, represents the number of parts of speech in the part-of-speech combination of the hth unit in the convolution sequence corresponding to the second dialogue text, represents the number of parts of speech in the part-of-speech combination of the hth unit in the convolution sequence corresponding to the third dialogue text, Indicates the number of parts of speech that are the same in the h-th unit's part of speech combination in the convolution sequence corresponding to the second dialogue text and the h-th unit's part of speech combination in the convolution sequence corresponding to the third dialogue text.

[0127] In one embodiment of the present invention, a third dialogue text is obtained within a third preset time period, and the third dialogue text is compared with a regular text (such as the second dialogue text). If the similarity value is greater than or equal to a preset similarity threshold, the third dialogue text is also a regular text.

[0128] When the dialogue text is a regular text, the rule mapping processing unit is used as the interface of the integrated system to achieve the connection between the customer service system and the business system. When the dialogue text is an irregular text, the generator is used as the interface of the integrated system to achieve the connection between the customer service system and the business system.

[0129] In one embodiment of the present invention, the part-of-speech tagging tool includes but is not limited to jieba or HanLP. For example, the second dialogue text is "I want to check the order status", and the third dialogue text is "He wants to check the express delivery progress". Among them, "I" and "he" are usually marked as personal pronouns (such as "r"), "want" and "want" may be marked as modal verbs (such as "v"), "query" and "check" are verbs ("v"), "order" and "express delivery" are nouns ("n"), and "status" and "progress" are also nouns ("n"). Then the part-of-speech sequence of the second dialogue text is [r, v, v, n, n], and the part-of-speech sequence of the third dialogue text is [r, v, v, n, n].

[0130] In detail, part-of-speech sequences of different lengths (such as 8 and 10) cannot be directly compared based on fixed positions. By dynamically adjusting the sliding window length, they are unified into the same dimension H=6, so that the similarity between the two can be calculated under a consistent framework, avoiding comparison obstacles caused by different lengths.

[0131] The sliding window dynamically extracts local part-of-speech combinations and can adapt to the structural characteristics of sequences of different lengths. For example, the long sequence 10 can capture detailed features through a smaller window, and the short sequence 8 can ensure feature coverage through a larger window. Finally, under the dimension of H=6, the local structural patterns of the two are fully extracted and compared to explore potential similarities.

[0132] For example, the second dialogue text is "I'm going to school for class". After part-of-speech tagging, the part-of-speech sequence P2 is [r, v, n, v, n], with a length of 5. Set the sliding window length to 3 and perform sliding extraction. The obtained convolutional sequence C2 includes: the first unit [r, v, n], the second unit: [v, n, v], and the third unit: [n, v, n].

[0133] The third dialogue text is "He comes to the company", and the part-of-speech sequence P3 is [r, v, n, u], with a length of 4. For the third dialogue text, set the sliding window length to 2, and at the same time, perform a certain degree of overlap and integration during extraction to obtain a convolutional sequence with a length of 3. The obtained convolutional sequence C3 includes: the first unit [r, v], the second unit: [v, n], and the third unit: [n, u]. Here, u represents the particle part of speech.

[0134] Based on the formula, the similarity value of the first unit is 0.8, the similarity value of the second unit is 0.8, and the similarity value of the third unit is 0.8. Then the overall similarity value between the second dialogue text and the third dialogue text is 0.8. If the preset similarity threshold is 0.95, then the third dialogue text is an irregular text and needs to be translated by the generator. If the preset similarity threshold is 0.7, then the third dialogue text is a regular text and needs to be translated by the rule mapping processing unit.

[0135] In an embodiment of the present invention, the word vector space is the Word2Ve word vector space.

[0136] The above has described the embodiments of this embodiment, but this embodiment is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of this embodiment.

Claims

1. A customer service interface docking management system, characterized in that: include: A data acquisition module, used to obtain a first conversation text within a first preset time period; Map the first conversation text to the word vector space after preprocessing to obtain a text vector; The adversarial generation module is used to determine the logical entities of the system to be integrated and use the database tables corresponding to the logical entities as the true labels; The adversarial network model is trained based on real samples and text vectors; The translation management module is used to obtain the second dialogue text within a second preset time period and perform a translation operation on the second dialogue text, including: Extract the generator and discriminator of the adversarial generative model; Performing translation on the target dialogue text based on the generator to obtain translation data of the integrated system; Based on the second dialogue text and the translation database table, construct a rule mapping processing unit; Mapping relationship of processing units based on discriminator reverse optimization rule mapping; The mapping management module is used to obtain a third dialogue text within a third preset time period and perform a translation operation on the third dialogue text, including: The third dialogue text is classified into regular text and irregular text; If it is a regular text, the third dialogue text is translated based on the mapping processing unit; If it is a non-regular text, the third dialogue text is translated based on the generator.

2. A customer service interface docking management system according to claim 1, characterized in that: Map the first conversation text to the word vector space after preprocessing to obtain a text vector; The preprocessing includes: text cleaning to remove special characters, punctuation marks and stop words in the first dialogue text to obtain a preprocessed text of the first dialogue text; Map the preprocessed text to the word vector space to obtain the initial text vector; The initial text vector is convolved based on a one-dimensional convolutional neural network to obtain a text vector.

3. A customer service interface docking management system according to claim 2, characterized in that: The adversarial network model is trained based on real samples and text vectors; The adversarial network model includes: generator and discriminator; Input the text vector into the generator to obtain the translation database table, and input the true label into the discriminator to make a true probability judgment for the translation database table; If there are P consecutive times where the absolute difference between the mean of the true probability and the preset convergence threshold is less than or equal to the preset difference threshold, the training converges and the adversarial network model is obtained.

4. A customer service interface docking management system according to claim 3, characterized in that: Based on the second dialogue text and the translation database table, a rule mapping processing unit is constructed, including: After preprocessing the second dialogue text, segment the preprocessed text based on a Chinese word segmentation tool to obtain M segmentation fields; Extract and translate N field values ​​from the database table; Establish a mapping relationship between M word segmentation fields and N field values, including: Initialize and generate a mapping scheme that meets the constraint conditions; wherein the mapping scheme includes a mapping relationship between each field value and a number of word segmentation fields; The constraints include: each field value has a mapping relationship with at least one word segmentation field, each field value has a mapping relationship with at most M word segmentation fields, and each field value has and can only have one mapping relationship with each word segmentation field; A rule mapping processing unit is constructed based on the mapping scheme.

5. A customer service interface docking management system according to claim 4, characterized in that: The mapping relationship of the rule mapping processing unit is reversely optimized based on the discriminator, including determining the score of the rule mapping processing unit, as follows: Generate a corresponding rule translation database table for the second dialogue text based on the rule mapping processing unit; The discriminator based on the adversarial network model determines the true probability of the rule translation database table; The absolute difference between the true probability and the preset convergence threshold is used as the score of the rule mapping processing unit; If the absolute difference between the score and the preset convergence threshold is greater than the preset difference threshold, the mapping relationship of the mapping scheme is updated based on the gradient descent method until the absolute difference between the score and the preset convergence threshold is less than or equal to the preset difference threshold.

6. A customer service interface docking management system according to claim 5, characterized in that: The third dialogue text is classified into regular text and irregular text, including: If the absolute difference between the score of the rule mapping processing unit and the preset convergence threshold is less than or equal to the preset difference threshold, the corresponding second dialogue text is used as the rule text; Preprocessing, word segmentation and part-of-speech tagging of the second dialogue text and the third dialogue text; The part-of-speech tagging obtains the part-of-speech of each word segmentation field corresponding to the second dialogue text and the third dialogue text based on a part-of-speech tagging tool; Based on the position order of the word segmentation fields in the second dialogue text and the third dialogue text, corresponding part-of-speech sequences are constructed respectively; wherein the i-th unit of the part-of-speech sequence represents the part-of-speech of the i-th word segmentation field in the second dialogue text or the third dialogue text; Calculate the similarity value of the corresponding part-of-speech sequence, specifically including: Loading a sliding extraction window to convert the part-of-speech sequence into a convolution sequence with a dimension of H; wherein the length of the sliding extraction window is dynamically set based on the dimension of the part-of-speech sequence; Calculate the similarity value of the part-of-speech combination of the h-th unit in the corresponding convolution sequence respectively; wherein the part-of-speech combination is extracted from the part-of-speech sequence based on the sliding extraction window; If the similarity value is greater than or equal to a preset similarity threshold, the third dialogue text is classified as a regular text; otherwise, the third dialogue text is classified as an irregular text.

7. A customer service interface docking management system according to claim 6, characterized in that: Calculate the similarity value of the part-of-speech combination of the hth unit in the corresponding convolution sequence, including: ； Where, 1≤h≤H, h is a positive integer, Represents the similarity value of part-of-speech sequence, represents the number of parts of speech in the part-of-speech combination of the hth unit in the convolution sequence corresponding to the second dialogue text, represents the number of parts of speech in the part-of-speech combination of the hth unit in the convolution sequence corresponding to the third dialogue text, Indicates the number of parts of speech that are the same in the h-th unit's part of speech combination in the convolution sequence corresponding to the second dialogue text and the h-th unit's part of speech combination in the convolution sequence corresponding to the third dialogue text.

8. A customer service interface docking management system according to claim 7, characterized in that: The word vector space is Word2Ve word vector space.

Images