CN116108858A

CN116108858A - Text processing method and device

Info

Publication number: CN116108858A
Application number: CN202310141120.2A
Authority: CN
Inventors: 范晓宁; 杨洪鑫; 郑岩; 张翌阳
Original assignee: Alipay Labs Singapore Pte Ltd
Current assignee: Advanced Nova Technology Singapore Holdings Ltd
Priority date: 2023-02-20
Filing date: 2023-02-20
Publication date: 2023-05-12

Abstract

The embodiment of the specification provides a text processing method and a text processing device, wherein the text processing method comprises the following steps: receiving a text to be processed; inputting the text to be processed into a text expression layer of a text processing model, and obtaining a semantic unit expression vector of a semantic unit in the text to be processed, which is output by the text expression layer, wherein the semantic unit expression vector comprises characteristic elements of at least two dimensions; inputting the semantic unit expression vector to a fusion layer of the text processing model to obtain semantic unit fusion characteristics output by the fusion layer; inputting the semantic unit fusion features to a classification layer of the text processing model to obtain the category information of the text to be processed output by the classification layer; according to the category information, determining a history quantity threshold corresponding to the category information; and generating early warning information under the condition that the quantity of the texts to be processed corresponding to the category information reaches the historical quantity threshold value.

Description

Text processing method and device

Technical Field

Embodiments of the present disclosure relate to the field of computer technologies, and in particular, to a text processing method.

Background

In daily business, a large number of work orders with technical problems or fault consultation as main contents are usually required to be processed, and the work orders contain feedback of business from users or technical partners, wherein part of information can reflect risks and faults existing in the current business.

However, in the processing of the work order, manual processing is generally required by distributing to different business personnel, so that the feedback problem and the fault in the work order are dispersed, and the missing and the delay processing of the fault are caused. Therefore, an effective solution is needed to solve the above problems.

Disclosure of Invention

In view of this, the present embodiments provide a text processing method. One or more embodiments of the present specification also relate to a text processing apparatus, a computing device, a computer-readable storage medium, and a computer program that solve the technical drawbacks existing in the prior art.

According to a first aspect of embodiments of the present specification, there is provided a text processing method, including:

receiving a text to be processed;

inputting the text to be processed into a text expression layer of a text processing model, and obtaining a semantic unit expression vector of a semantic unit in the text to be processed, which is output by the text expression layer, wherein the semantic unit expression vector comprises characteristic elements of at least two dimensions;

Inputting the semantic unit expression vector to a fusion layer of the text processing model to obtain semantic unit fusion characteristics output by the fusion layer;

inputting the semantic unit fusion features to a classification layer of the text processing model to obtain the category information of the text to be processed output by the classification layer;

according to the category information, determining a history quantity threshold corresponding to the category information;

and generating early warning information under the condition that the quantity of the texts to be processed corresponding to the category information reaches the historical quantity threshold value.

According to a second aspect of embodiments of the present specification, there is provided a text processing apparatus comprising:

a receiving module configured to receive text to be processed;

the first input module is configured to input the text to be processed into a text expression layer of a text processing model, and obtain a semantic unit expression vector of a semantic unit in the text to be processed, which is output by the text expression layer, wherein the semantic unit expression vector comprises characteristic elements of at least two dimensions;

the second input module is configured to input the semantic unit expression vector to a fusion layer of the text processing model, and obtain semantic unit fusion characteristics output by the fusion layer;

The third input module is configured to input the semantic unit fusion features to a classification layer of the text processing model, and obtain the category information of the text to be processed output by the classification layer;

the determining module is configured to determine a history quantity threshold corresponding to the category information according to the category information;

and the early warning module is configured to generate early warning information under the condition that the quantity of the texts to be processed corresponding to the category information reaches the historical quantity threshold value.

According to a third aspect of embodiments of the present specification, there is provided a computing device comprising:

a memory and a processor;

the memory is configured to store computer-executable instructions that, when executed by the processor, perform the steps of the text processing method described above.

According to a fourth aspect of embodiments of the present specification, there is provided a computer-readable storage medium storing computer-executable instructions which, when executed by a processor, implement the steps of the text processing method described above.

According to a fifth aspect of embodiments of the present specification, there is provided a computer program, wherein the computer program, when executed in a computer, causes the computer to perform the steps of the text processing method described above.

One embodiment of the present specification provides a text processing method, which receives text to be processed; inputting the text to be processed into a text expression layer of a text processing model, and obtaining a semantic unit expression vector of a semantic unit in the text to be processed, which is output by the text expression layer, wherein the semantic unit expression vector comprises characteristic elements of at least two dimensions; inputting the semantic unit expression vector to a fusion layer of the text processing model to obtain semantic unit fusion characteristics output by the fusion layer; inputting the semantic unit fusion features to a classification layer of the text processing model to obtain the category information of the text to be processed output by the classification layer; according to the category information, determining a history quantity threshold corresponding to the category information; and generating early warning information under the condition that the quantity of the texts to be processed corresponding to the category information reaches the historical quantity threshold value.

According to the method, the text processing model is utilized to automatically classify the worksheets, category information of the worksheets is obtained, semantic unit expression vectors comprising at least two dimension characteristic elements are processed, so that text classification results in the worksheets are more accurate by the text processing model, when the fact that the worksheets of certain category information reach a historical quantity threshold value is determined, the quantity of the worksheets reflecting the category is more, early warning information can be generated for the worksheets of the category, timely processing of faults of the same category reflected in a large number of worksheets is achieved, and labor cost can be reduced while omission of the faults is avoided.

Drawings

Fig. 1 is a schematic view of an application scenario of a text processing method according to an embodiment of the present disclosure;

FIG. 2 is a flow chart of a text processing method provided by one embodiment of the present description;

FIG. 3 is a schematic diagram of training a text processing model in a text processing method according to an embodiment of the present disclosure;

FIG. 4 is a schematic diagram of pre-training a text presentation layer in a text processing method according to an embodiment of the present disclosure;

FIG. 5 is a schematic diagram of determining a threshold of a number of histories in a text processing method according to an embodiment of the present disclosure;

FIG. 6 is a process flow diagram of a text processing method provided in one embodiment of the present disclosure;

FIG. 7 is a schematic diagram of a text processing device according to an embodiment of the present disclosure;

FIG. 8 is a block diagram of a computing device provided in one embodiment of the present description.

Detailed Description

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present description. This description may be embodied in many other forms than described herein and similarly generalized by those skilled in the art to whom this disclosure pertains without departing from the spirit of the disclosure and, therefore, this disclosure is not limited by the specific implementations disclosed below.

The terminology used in the one or more embodiments of the specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of the specification. As used in this specification, one or more embodiments and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used in one or more embodiments of the present specification refers to and encompasses any or all possible combinations of one or more of the associated listed items.

It should be understood that, although the terms first, second, etc. may be used in one or more embodiments of this specification to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, a first may also be referred to as a second, and similarly, a second may also be referred to as a first, without departing from the scope of one or more embodiments of the present description. The word "if" as used herein may be interpreted as "at … …" or "at … …" or "responsive to a determination", depending on the context.

In practical applications, when analyzing and processing a work order proposed by a user or a technical partner, the following problems often exist.

Firstly, a work order is proposed by a user, and due to view angle reasons and familiarity, the content fed back by the user cannot reflect the root cause or the real reason of a problem, and meanwhile, the user description is difficult to attach to the existing autonomous label. Multiple rounds of conversations are often required to verify the problem's authenticity and cause by business personnel, wasting labor costs, and being inefficient.

And secondly, the daily work orders contain a large number of consultation work orders or customer reasons and complaints, and the work orders are huge in quantity, and if the work orders are not processed, a large number of work orders can be misreported. Meanwhile, the problems of part serious faults or customer complaint worksheet reaction are more special, the worksheet quantity is small, and when manual treatment is carried out, the serious faults or special problems existing in the worksheet quantity are often ignored due to the small worksheet quantity, so that public opinion risks are caused.

In addition, although a large amount of historical work order information is accumulated in the service, the number of marking work orders meeting the requirements is less due to the difference of a label system and the service, and meanwhile, the marking of key information of the work orders is also less.

Therefore, an effective solution is needed to solve the above problems.

It should be noted that, the user information (including, but not limited to, user equipment information, user personal information, etc.) and the data (including, but not limited to, data for analysis, stored data, presented data, etc.) according to the embodiments of the present disclosure are information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related country and region, and provide corresponding operation entries for the user to select authorization or rejection.

In the present specification, a text processing method is provided, and the present specification relates to a text processing apparatus, a computing device, and a computer-readable storage medium, which are described in detail in the following embodiments one by one.

Referring to fig. 1, fig. 1 shows an application scenario of a text processing method according to an embodiment of the present disclosure.

Fig. 1 includes a client 102 and a server 104, where the server 104 may be configured to perform the text processing method.

In implementation, a service person may send a work order to be processed to the server 104 through the client 102, and after the server 104 receives the work order to be processed, the work order to be processed may be input into a text processing model deployed at the server 104, so as to obtain category information of the work order to be processed and keywords corresponding to the category information, which are output by the text processing model. And determining a history quantity threshold corresponding to the category information, generating early warning information according to the keywords corresponding to the category information and the work order information of the work order to be processed under the condition that the quantity of the work orders to be processed corresponding to the category information reaches the history quantity threshold, and sending the early warning information to the client 102. The method realizes the timely early warning of the feedback faults in a large number of work orders to be processed of the same class, and is convenient for subsequent processing.

Referring to fig. 2, fig. 2 shows a flowchart of a text processing method according to an embodiment of the present disclosure, which specifically includes the following steps.

Step 202: receiving text to be processed.

Specifically, the text processing method provided in the embodiments of the present disclosure may be applied to a server, and in particular, may be applied to a text processing system deployed at the server.

The text to be processed can be understood as text to be classified, for example, the text can be a work order, the work order can be understood as a dialogue list taking technical problems or fault consultation as main contents, and the dialogue list can be a work order initiated by a user or a work order initiated by a technical partner. For example, for e-commerce services, a user can initiate a fault consultation to a customer service aiming at the fault of the e-commerce platform, and then a dialogue list between the user and the customer service is a text to be processed. Or, aiming at the technical problem of the e-commerce platform, the fault feedback initiated by the technical partner is the text to be processed.

Based on this, conversation lists with technical problems or fault consultation as main contents can be received, and the conversation lists can be conversation lists received in real time, specifically, the conversation lists can be collected in real time by the work order distribution system and sent to the service end.

Step 204: inputting the text to be processed into a text expression layer of a text processing model, and obtaining a semantic unit expression vector of a semantic unit in the text to be processed, which is output by the text expression layer, wherein the semantic unit expression vector comprises characteristic elements with at least two dimensions.

Specifically, after receiving the text to be processed, the text to be processed may be input into a text processing model, and the text expression layer of the text processing model is used to process the text to be processed, so as to obtain a semantic unit expression vector of a semantic unit in the text to be processed.

Where a semantic unit may be understood as a single word or word in the text to be processed. For example, for the text to be processed, "you good, i have questions to ask", the semantic unit may be "you" or "ask", etc., or "you good" or "questions", etc. Then, for other languages of text to be processed, such as for english text "I have a question", the semantic unit may be "I" or "have" or "a", etc. The text processing method provided in the embodiment of the present specification may be applied to various languages, and the embodiment of the present specification is not limited herein. The semantic unit expression vector may be understood as a coded vector obtained by coding the semantic unit. Feature elements may be understood as vector features that constitute semantic unit expression vectors.

Specifically, the semantic unit expression vector at least comprises a feature element of a semantic dimension, a feature element of a Chinese order dimension, a feature element of a satisfaction dimension and a feature element of a feedback duration dimension.

Wherein, feature elements of the semantic dimension may be understood as feature elements for text semantics in the text to be processed. The feature elements of the order dimension may be understood as feature elements for the text order in the text to be processed. A feature element of the satisfaction dimension may be understood as a feature element for text emotion in the text to be processed. The feature elements of the feedback duration dimension may be understood as feature elements of the interval duration between dialog and answer in the text to be processed.

For example, for a semantic unit expression vector that includes 100 dimensions, where feature elements of the semantic dimensions may be in the 1 st to 50 th dimensions, feature elements of the order dimensions may be in the 51 st to 150 th dimensions, and so on, feature elements of each dimension are included in the semantic unit expression vector.

It will be appreciated that the dimensions included in the semantic unit expression vector may be determined according to actual requirements, and embodiments of the present disclosure are not limited herein.

In conclusion, the semantic unit expression vector comprises at least two dimension characteristic elements, the text structure and text information in the text to be processed can be fully utilized for learning, a large amount of marking data is not needed, and the accuracy of an output result of the text processing model in application is improved.

In practical application, a small amount of marking data can be used for carrying out supervised training on a text processing model, and the specific implementation mode is as follows:

the training steps of the text processing model are as follows:

determining a first text training sample and a first text training label corresponding to the first text training sample;

inputting the first text training sample into the text expression layer obtained through pre-training to obtain a predicted semantic unit expression vector which is output by the text expression layer and corresponds to a semantic unit in the first text training sample;

inputting the predicted semantic unit expression vector to a fusion layer of the text processing model to obtain predicted semantic unit fusion characteristics output by the fusion layer;

inputting the fusion characteristics of the prediction semantic units to a classification layer of the text processing model to obtain the prediction category information of the semantic unit training sample output by the classification layer;

And training the text processing model by utilizing the prediction category information and the first text training label until the text processing model meeting the training stopping condition is obtained.

The first text training sample may be understood as a historical work order, and the first text training label may be understood as a category label corresponding to the historical work order, for example, may be a category label of a problem fed back in the historical work order. The first text training label may be specified according to actual business requirements. The training stop condition may be understood as reaching a preset number of training or the model loss value reaching a preset loss value threshold.

Specifically, when training the text processing model, the model loss value can be calculated according to the prediction type information and the first text training label, and parameters of the text processing model are adjusted by using the model loss value until the text processing model meeting the training stop condition is obtained.

In practical applications, the fusion layer may be implemented using an attention handling mechanism.

Specifically, referring to fig. 3, fig. 3 is a schematic diagram illustrating training a text processing model in a text processing method according to an embodiment of the present disclosure.

As shown in fig. 3, when training the text processing model 300, a first text training sample W may be input to the text expression layer 302 in the text processing model, specifically, when inputting the text training sample, the first text training sample may be directly input, or the first text training sample may be split into semantic units W1, W2, W3 … … Wn and respectively input to the text expression layer 302, where the text expression layer 302 may be a pre-trained text expression layer, to obtain predicted semantic unit expression vectors V1, V2, V3 … … Vn corresponding to each semantic unit in the first text training sample output by the text expression layer 302. The predicted semantic unit expression vector corresponding to each semantic unit is input to a fusion layer 304 of the text processing model 300, weighting processing is carried out on the weight of each predicted semantic unit expression vector by using the fusion layer 304, a predicted semantic unit fusion feature Vf output by the fusion layer 304 is obtained, the predicted semantic unit fusion feature is input to a classification layer 306 of the text processing model 300, and prediction category information of each semantic unit output by the classification layer 306 is obtained. And training the text processing model 300 according to the prediction category information and the first text training label until a text processing model meeting the training stop condition is obtained.

In addition, in the training process of the text processing model, the fusion layer 304 may also be used to calculate the weight of the expression vector of the predicted semantic unit corresponding to each semantic unit, and when it is determined that the weight of the expression vector of the predicted semantic unit reaches the preset weight threshold, the predicted semantic unit corresponding to the expression vector of the predicted semantic unit is determined, and the predicted semantic unit is output as a keyword.

In addition, in the training process of the text processing model, the fusion characteristic Vf of the prediction semantic unit may be further processed to obtain a secondary fusion characteristic Vf2, and the classification layer 306 is utilized to output a prediction secondary category for the first text training sample, where the prediction secondary category may be understood as a sub-category under the prediction category information. For example, for the prediction result of the first text training sample, the prediction category information may be a question feedback category, and the prediction secondary category may be a payment question feedback category.

In sum, the text processing model is trained by using a small amount of marking data, so that the training cost of the model is reduced, the method can be quickly adapted to the business classification and change rhythm, and maintenance is not required by extra labor cost, so that the precision and accuracy of the output result of the text processing model in subsequent application can be ensured.

In practical application, although a large amount of historical work order information exists, due to different services, the category label information of the historical work order has timeliness, the re-marking efficiency is lower, and if the part of historical work order information is not utilized, a large amount of semantic information can be lost, so that a self-supervision training method can be utilized to pretrain a text expression layer, so that the pretrained text expression layer has the semantic analysis capability, and the concrete implementation mode is as follows:

the step of pre-training the obtained text expression layer comprises the following steps:

determining a text training corpus, wherein the text training corpus comprises at least two training texts;

constructing a second text training sample and a corresponding second text training label according to the text training corpus, and pre-training the text expression layer according to the second text training sample and the second text training label to obtain a pre-trained text expression layer.

Where the text training corpus may be understood as a set of historical worksheets, then the training text may be understood as any one of the set of historical worksheets. Pre-training the text presentation layer can be understood as self-supervising training of the text presentation layer by using the constructed second text training sample and the second text training label.

Based on the method, a historical work order set can be determined, a second text training sample and a second text training label are constructed according to the historical work order set, and the constructed second text training sample and second text training label are utilized to conduct self-supervision training on the text expression layer.

It can be understood that the second text training sample is data without marking, and self-supervision training of the text expression layer can be realized by constructing a second text training label corresponding to the second text training sample.

Specifically, when the text expression layer is subjected to self-supervision training, the text expression layer can be trained through different dimensions, for example, the text expression layer can be trained through four dimensions, namely a semantic dimension, a word order dimension, a satisfaction degree dimension and a feedback time length dimension, so that the semantic information extraction capability and the emotion analysis capability of the text expression layer are improved, and the accuracy of a text classification result and an emotion classification result of a text processing model is further improved.

Referring to fig. 4, fig. 4 is a schematic diagram illustrating pre-training of a text expression layer in a text processing method according to an embodiment of the present disclosure.

As shown in fig. 4, the text expression layer may be implemented by using a BERT framework, specifically, a converter may be used to construct the text expression layer, and when the text expression layer is pre-trained, self-supervision training may be performed by using a semantic dimension prediction task 402, a word order dimension prediction task 404, a satisfaction dimension prediction task 406, and a feedback duration dimension prediction task 408, so as to obtain a trained text expression layer.

When the method is implemented, during training of semantic dimension prediction tasks on the text expression layer, single words and/or words in any work order can be hollowed randomly, and mask processing is carried out on the hollowed positions, so that the hollowed single words and/or words can be predicted by the training text expression layer, the capability of extracting semantic information of the training text expression layer is achieved, and the capability of understanding natural language of the text expression layer is enhanced, and the method is implemented as follows:

constructing a second text training sample and a corresponding second text training label according to the text training corpus, pre-training a text expression layer according to the second text training sample and the second text training label, and obtaining a pre-trained text expression layer, wherein the pre-trained text expression layer comprises the following components:

extracting semantic units in any training text, and performing mask processing on the semantic units to obtain a mask text;

and taking the mask text as a second text training sample, taking the semantic unit as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

Specifically, in the training process of the semantic dimension prediction task, characters and words in the work order can be extracted, and mask processing is performed on the positions of the characters and the words in the work order to obtain mask text. And taking the mask text as a second text training sample, taking the extracted characters and words as a second text training label, and training the text expression layer to obtain a pre-trained text expression layer.

For example, for a sentence "hello, i have questions to ask" in a work order, the word "questions" may be masked to obtain the mask text "hello, i have a question. And taking the mask text 'hello, I have a question' as a second text training sample, taking the word 'question' as a second text training label, and pre-training the text expression layer.

In conclusion, through the pre-training of semantic dimensions, the text processing model can capture basic semantic information in the text to be processed, and the understanding capability of the text processing model to natural language is enhanced.

determining a text sentence and any training text in the text training corpus, and determining the inclusion relation between the text sentence and any training text;

and taking the text sentence and any training text as a second text training sample, taking the inclusion relationship as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

Specifically, in at least two training texts included in the text training corpus, any one training text may be determined, and any one text sentence may be determined, where the text sentence may be extracted from the determined training text, or may be extracted from other training texts. And determining whether the text sentence belongs to the inclusion relation of the training text according to the text sentence and the training text. Taking the text sentence and the training text as a second text training sample, taking the inclusion relationship as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

For example, determining a text sentence "the commodity link is not opened" and any training text "you good, how the commodity link is not opened, asking how to return to the question", determining that the text sentence and the training text have a inclusion relationship of "the text sentence belongs to the training text", opening the commodity link of the text sentence "the commodity link is not opened" and any training text "you good, how to return to the question" as a second text training sample, taking the inclusion relationship "the text sentence belongs to the training text" as a second text training label, and pre-training the text expression layer.

In conclusion, through the pre-training of the word order dimension, the basic semantic understanding capability of the text processing model can be improved.

determining at least two text sentences in the text training corpus, and determining the word order between the at least two text sentences;

adjusting the word order between the at least two text sentences to obtain an adjusted text;

and taking the adjusted text as a second text training sample, taking the word order as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

Specifically, any at least two text sentences can be determined in the training text in the text training corpus, and the word order between the at least two text sentences is determined, for example, the text training expectation comprises a training text 1 and a training text 2, the at least two text sentences can be determined in the training text 1, and the at least two text sentences can be determined from the training text 2. Adjusting the word sequence between at least two text sentences to obtain an adjusted text, taking the adjusted text as a second text training sample, taking the word sequence as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

For example, two text sentences are determined to be "you good", "i have questions to ask", the word order between the two text sentences is adjusted, the adjusted text is obtained to be "i have questions to ask", you good ", the adjusted text" i have questions to ask ", you good" is used as a second text training sample, the word order "you good, i have questions to ask" is used as a second text training label, and the text expression layer is pre-trained.

In conclusion, training of understanding capability of model natural language is achieved through text expression layer pre-training of the word order dimension.

determining satisfaction corresponding to any one training text, and determining a satisfaction interval corresponding to any one training text according to the satisfaction;

and taking any training text as a second text training sample, taking the satisfaction interval as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

Specifically, the satisfaction degree corresponding to the training text can be determined according to the feedback information of the training text. Because the training text is a historical work order, the satisfaction degree corresponding to the historical work order can be obtained based on feedback information of a user aiming at the historical work order or based on evaluating the historical work order manually. And determining a satisfaction interval corresponding to the training text according to the satisfaction.

The satisfaction interval may be understood as an interval range to which satisfaction belongs. For example, the satisfaction interval may be "very satisfactory" for a satisfaction of 90 minutes to 100 minutes, and may be "satisfactory" for a satisfaction of 80 minutes to 90 minutes. Alternatively, the satisfaction interval may be expressed by a character, for example, the satisfaction is 90 minutes to 100 minutes, the satisfaction interval may be "5", and the satisfaction interval may be "4" for the satisfaction is 80 minutes to 90 minutes. The satisfaction interval may be expressed in any form according to actual requirements, and the embodiments of the present specification are not limited herein.

Based on the above, the satisfaction degree corresponding to any one training text in at least two training texts can be determined, the satisfaction degree interval to which the satisfaction degree belongs is determined, the training text is used as a second text training sample, the satisfaction degree interval is used as a second text training label, and the text expression layer is subjected to training to obtain a pre-trained text expression layer.

In sum, through the training of satisfaction dimension, the text expression layer can be trained to recognize negative or more intense emotion in the dialogue of the text to be processed, so that the weight of the text to be processed is improved, a high public opinion or customer complaint risk work order is exposed as early as possible, and delayed discovery or missing report is prevented.

determining a dialogue sample in the text training corpus;

determining a feedback duration interval corresponding to the dialogue sample according to the timestamp of the dialogue sample;

and taking the dialogue sample as a second text training sample, taking the feedback time interval as a second text training label, and pre-training the text expression layer to obtain a pre-trained text expression layer.

Specifically, a dialogue sample can be determined in any training text in the text training corpus, the feedback duration of the dialogue sample is determined according to the timestamp of the dialogue sample, and the feedback duration interval corresponding to the dialogue sample is determined according to the feedback duration.

The feedback duration may be understood as the duration of the interval between the reply and the question in the dialogue sample. For example, for the dialogue sample in the training text, the user asks a, the customer service replies B, the time of the user asks a is 8 points, the time of the customer service replies B is 9 points, and then the feedback time is 1 hour.

It can be appreciated that the process of determining the feedback duration interval is similar to the process of determining the satisfaction interval, and will not be repeated here.

Based on the text training method, the dialogue sample can be used as a second text training sample, the feedback duration interval is used as a second text training label, and the text expression layer is pre-trained to obtain the pre-trained text expression layer.

In conclusion, through training of the feedback duration dimension, the understanding capability and the analysis capability of the model to the text to be processed are further improved.

Step 206: and inputting the semantic unit expression vector to a fusion layer of the text processing model to obtain semantic unit fusion characteristics output by the fusion layer.

Wherein, since the text to be processed comprises at least two semantic units, correspondingly, the semantic unit expression vectors are at least two;

correspondingly, the step of inputting the semantic unit expression vector to a fusion layer of the text processing model to obtain semantic unit fusion characteristics output by the fusion layer comprises the following steps:

Inputting at least two semantic unit expression vectors into a fusion layer of the text processing model, and calculating the weight of any semantic unit expression vector by using the fusion layer;

and weighting the at least two semantic unit expression vectors based on the weight of any semantic unit expression vector to obtain semantic unit fusion characteristics output by the fusion layer.

Specifically, the weight of each semantic unit expression vector can be calculated by using a fusion layer of the text expression model, and based on the weight of each semantic unit expression vector, weighting processing is performed on at least two semantic unit expression vectors to obtain semantic unit fusion characteristics output by the fusion layer.

In addition, after the weight of any semantic unit expression vector is calculated by using the fusion layer, the method further comprises the following steps:

under the condition that the weight of any semantic unit expression vector reaches a preset weight threshold value, the semantic unit expression vector is used as a target semantic unit expression vector;

and determining and outputting a target semantic unit corresponding to the target semantic unit expression vector according to the target semantic unit expression vector.

Specifically, after the weight of each semantic unit expression vector is calculated, the semantic unit expression vector with the weight reaching the preset weight threshold can be used as a target semantic unit expression vector, and a target semantic unit corresponding to the target semantic unit expression vector is determined and output.

In sum, by extracting the target semantic units according to the weights, keywords in the text to be processed can be effectively captured, noise can be filtered, and accuracy of subsequent text classification is improved.

Step 208: and inputting the semantic unit fusion features to a classification layer of the text processing model to obtain the category information of the text to be processed, which is output by the classification layer.

Specifically, after determining the semantic unit fusion features, the semantic unit fusion features may be input to a classification layer of the text processing model, so as to obtain category information of the text to be processed output by the classification layer.

The category information of the text to be processed can be understood as category information of text content reflected by the text to be processed, for example, the category information of the text to be processed can be a consultation category, a fault feedback category and the like.

In addition, the classification layer may output keywords, specifically, the fusion layer may calculate the weight of the text expression vector corresponding to each text, and when determining that the weight of the text expression vector reaches the preset weight threshold, determine the text corresponding to the text expression vector, and output the text as the keywords.

And, a secondary category for the text to be processed, which can be understood as a sub-category under category information, can also be output. For example, the category information may be a question feedback category and the secondary category may be a payment question feedback category.

Step 210: and determining a historical quantity threshold corresponding to the category information according to the category information.

Specifically, after the category information is determined, a history number threshold corresponding to the category information may also be determined.

In a specific implementation, the determining, according to the category information, a history number threshold corresponding to the category information includes:

determining the number of texts corresponding to the category information in the history time according to the category information;

and determining a period number threshold and/or a time number threshold corresponding to the category information according to the text number.

The period number threshold may be understood as a text number threshold that takes a certain period of time as a period, and the time number threshold may be understood as a text number threshold at a certain time.

Specifically, a number of day level dimensions threshold and a number of hour level dimensions threshold may be calculated.

In practical applications, the period number threshold and the time number threshold may be represented in the form of a data baseline. For example, the same category number of tools in the historical time period of 10 days is extracted, the mean and variance are calculated, and the base baseline of the day level dimension is calculated according to the 3sigma rule. The 3sigma rule is a rule of thumb for fast calculation of normal distribution data of known mean and standard deviation. In statistics, the rule of thumb is that in normal distribution, the more accurate figures are 68.27%, 95.45% and 99.73% from the average to less than one, two, three standard deviations. Rules of thumb are most often used in statistics to predict the final result. The rule may act as a rough estimate of the upcoming data result before the standard deviation of the data is obtained and before the exact data can be collected.

For example, for a historical time period of 5 days, the number of work orders for the problem feedback category is 20 on the first day, 40 on the second day, 30 on the third day, 10 on the fourth day, and 15 on the fifth day, then the mean and variance of the number of work orders per day can be calculated, and the base baseline for the day level dimension can be calculated according to the 3sigma rule.

Because the base line granularity is thicker, high false alarm and false alarm rate can be caused, the days are divided into barrels according to 30 minutes, and the base line of each barrel is calculated and cut off according to the method and is used as the base line of the hour level.

For example, the barrels are divided into 30 minutes, 12 points to 12 points are half as one barrel, the number of the class worksheets from 12 points to 12 points in each day in the past 10 days is calculated, and the base line of the class worksheets in the interval from 12 points to 12 points is calculated by the same method as the above, namely the base line of the hour level dimension.

If the work orders are gathered in a large quantity at a certain moment, but the accumulated quantity does not exceed the base line, the alarm is missed or delayed. For this spike type case, the average work order amount of 10 days per hour is obtained by adding the work order amounts of 10 days at the same time within 10 days after the hour barrel division and dividing by 10. The value was smoothed using the kernel density estimate, taking the 10% bin with the highest probability, multiplied by the average amount of work orders in a day as the short time spike baseline (i.e., instant number threshold). Referring to fig. 5, fig. 5 is a schematic diagram illustrating a method for determining a threshold of a history number in a text processing method according to an embodiment of the present disclosure. Fig. 5 is a schematic diagram of a nuclear density estimation. The left graph in fig. 5 shows the number of work orders per hour counted by hour for a certain work order category (average number of work orders per hour per day for the past 10 days), and the right graph shows the estimated core density.

Step 212: and generating early warning information under the condition that the quantity of the texts to be processed corresponding to the category information reaches the historical quantity threshold value.

In the implementation, the generating early warning information under the condition that the number of the texts to be processed corresponding to the category information reaches the historical number threshold value includes:

and under the condition that the number of the texts to be processed corresponding to the category information reaches the period number threshold or the moment number threshold, generating early warning information according to the text information of the texts to be processed.

Specifically, under the condition that the number of the texts to be processed corresponding to the category information reaches a period number threshold or a time number threshold, keywords corresponding to the category information, information of the texts to be processed and the like can be obtained to generate early warning information for early warning.

The information of the text to be processed may include attribute information, content information, and the like of the text to be processed. For example, when the text to be processed is a work order, the information of the text to be processed may be content information of the work order, a link of the work order, and the like.

In summary, the method utilizes the text processing model to automatically classify the worksheets to obtain class information of the worksheets, and processes semantic unit expression vectors comprising at least two dimension feature elements, so that the text processing model can more accurately classify the texts in the worksheets, and when the worksheets with certain class information reach the historical quantity threshold value, the quantity of the worksheets reflecting the class is more, early warning information can be generated for the worksheets with the class, timely processing of faults of the same class reflected in a large number of worksheets is realized, and the labor cost can be reduced while omission of the faults is avoided.

The text processing method provided in the present specification is further described below with reference to fig. 6 by taking an application of the text processing method in work order pre-warning as an example. Fig. 6 shows a flowchart of a processing procedure of a text processing method according to an embodiment of the present disclosure, which specifically includes the following steps.

Step 602: receiving text to be processed.

The text to be processed may be a work order to be classified.

In particular, multiple worksheets may be received.

Step 604: and inputting the text to be processed into a text processing model to obtain category information output by the text processing model.

Specifically, each work order can be input into the text processing model, and category information of each work order output by the text processing model is obtained.

In implementation, a work order can be input into a text expression layer of a text processing model, and a text expression vector corresponding to each text in the work order, which is output by the text expression layer, is obtained, wherein the text expression vector can comprise characteristic elements of semantic dimension, characteristic elements of word order dimension, characteristic elements of satisfaction dimension and characteristic elements of feedback duration dimension. And inputting the text expression vectors into a fusion layer of the text processing model, calculating the weight of each text expression vector by using the fusion layer, and carrying out weighted summation on the text expression vectors of each text based on the weight of each text expression vector to obtain the text fusion feature. And inputting the character fusion features into a classification layer of the text processing model to obtain the class information of the work order output by the classification layer.

In addition, when the weight of each text expression vector is calculated by using the fusion layer, the text expression vector with the weight reaching the preset weight threshold value can be used as a target text expression vector, and the target text corresponding to the target text expression vector can be used as a keyword to be output.

Step 606: and determining a historical quantity threshold corresponding to the category information according to the category information.

Step 608: and generating early warning information under the condition that the quantity of the texts to be processed corresponding to the category information reaches the historical quantity threshold value.

Corresponding to the above method embodiments, the present disclosure further provides an embodiment of a text processing device, and fig. 7 shows a schematic structural diagram of a text processing device provided in one embodiment of the present disclosure. As shown in fig. 7, the apparatus includes:

a receiving module 702 configured to receive text to be processed;

a first input module 704, configured to input the text to be processed to a text expression layer of a text processing model, and obtain a semantic unit expression vector of a semantic unit in the text to be processed, which is output by the text expression layer, wherein the semantic unit expression vector includes feature elements of at least two dimensions;

a second input module 706, configured to input the semantic unit expression vector to a fusion layer of the text processing model, to obtain semantic unit fusion features output by the fusion layer;

a third input module 708, configured to input the semantic unit fusion feature to a classification layer of the text processing model, and obtain category information of the text to be processed output by the classification layer;

a determining module 710 configured to determine, according to the category information, a history number threshold corresponding to the category information;

And the early warning module 712 is configured to generate early warning information when the number of the texts to be processed corresponding to the category information reaches the historical number threshold.

In an alternative embodiment, the semantic unit expression vectors are at least two; the second input module 606 is further configured to:

In an alternative embodiment, the second input module 706 is further configured to:

In an optional embodiment, the semantic unit expression vector includes at least a feature element of a semantic dimension, a feature element of a word order dimension, a feature element of a satisfaction dimension, and a feature element of a feedback duration dimension.

In an alternative embodiment, the determining module 710 is further configured to:

In an alternative embodiment, the pre-warning module 712 is further configured to:

In an alternative embodiment, the apparatus further comprises a training module configured to:

In an alternative embodiment, the training module is further configured to:

determining a dialogue sample in the text training corpus;

In summary, the device automatically classifies the worksheets by using the text processing model to obtain class information of the worksheets, and processes semantic unit expression vectors including at least two dimension feature elements, so that the text processing model is more accurate in text classification results in the worksheets, and when the worksheets with certain class information reach the historical quantity threshold value, the quantity of the worksheets reflecting the class is more, early warning information can be generated for the worksheets with the class, timely processing of faults of the same class reflected in a large number of worksheets is realized, and labor cost can be reduced while omission of faults is avoided.

The above is an exemplary scheme of a text processing apparatus of the present embodiment. It should be noted that, the technical solution of the text processing apparatus and the technical solution of the text processing method belong to the same concept, and details of the technical solution of the text processing apparatus, which are not described in detail, can be referred to the description of the technical solution of the text processing method.

Fig. 8 illustrates a block diagram of a computing device 800 provided in accordance with one embodiment of the present description. The components of computing device 800 include, but are not limited to, memory 810 and processor 820. Processor 820 is coupled to memory 810 through bus 830 and database 850 is used to hold data.

Computing device 800 also includes access device 840, access device 840 enabling computing device 800 to communicate via one or more networks 860. Examples of such networks include public switched telephone networks (PSTN, public Switched Telephone Network), local area networks (LAN, local Area Network), wide area networks (WAN, wide Area Network), personal area networks (PAN, personal Area Network), or combinations of communication networks such as the internet. Access device 840 may include one or more of any type of network interface, wired or wireless, such as a network interface card (NIC, network interface controller), such as an IEEE802.11 wireless local area network (WLAN, wireless Local Area Network) wireless interface, a worldwide interoperability for microwave access (Wi-MAX, worldwide Interoperability for Microwave Access) interface, an ethernet interface, a universal serial bus (USB, universal Serial Bus) interface, a cellular network interface, a bluetooth interface, a near field communication (NFC, near Field Communication) interface, and so forth.

In one embodiment of the present application, the above-described components of computing device 800, as well as other components not shown in FIG. 8, may also be connected to each other, such as by a bus. It should be understood that the block diagram of the computing device illustrated in FIG. 8 is for exemplary purposes only and is not intended to limit the scope of the present application. Those skilled in the art may add or replace other components as desired.

Computing device 800 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., tablet, personal digital assistant, laptop, notebook, netbook, etc.), mobile phone (e.g., smart phone), wearable computing device (e.g., smart watch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or personal computer (PC, personal Computer). Computing device 800 may also be a mobile or stationary server.

Wherein the processor 820 is configured to execute computer-executable instructions that, when executed by the processor, perform the steps of the text processing method described above.

The foregoing is a schematic illustration of a computing device of this embodiment. It should be noted that, the technical solution of the computing device and the technical solution of the text processing method belong to the same concept, and details of the technical solution of the computing device, which are not described in detail, can be referred to the description of the technical solution of the text processing method.

An embodiment of the present disclosure also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the text processing method described above.

The above is an exemplary version of a computer-readable storage medium of the present embodiment. It should be noted that, the technical solution of the storage medium and the technical solution of the text processing method belong to the same concept, and details of the technical solution of the storage medium which are not described in detail can be referred to the description of the technical solution of the text processing method.

An embodiment of the present specification also provides a computer program, wherein the computer program, when executed in a computer, causes the computer to perform the steps of the text processing method described above.

The above is an exemplary version of a computer program of the present embodiment. It should be noted that, the technical solution of the computer program and the technical solution of the text processing method belong to the same conception, and details of the technical solution of the computer program, which are not described in detail, can be referred to the description of the technical solution of the text processing method.

The foregoing describes specific embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

The computer instructions include computer program code that may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth. It should be noted that the computer readable medium contains content that can be appropriately scaled according to the requirements of jurisdictions in which such content is subject to legislation and patent practice, such as in certain jurisdictions in which such content is subject to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

It should be noted that, for simplicity of description, the foregoing method embodiments are all expressed as a series of combinations of actions, but it should be understood by those skilled in the art that the embodiments are not limited by the order of actions described, as some steps may be performed in other order or simultaneously according to the embodiments of the present disclosure. Further, those skilled in the art will appreciate that the embodiments described in the specification are all preferred embodiments, and that the acts and modules referred to are not necessarily all required for the embodiments described in the specification.

In the foregoing embodiments, the descriptions of the embodiments are emphasized, and for parts of one embodiment that are not described in detail, reference may be made to the related descriptions of other embodiments.

The preferred embodiments of the present specification disclosed above are merely used to help clarify the present specification. Alternative embodiments are not intended to be exhaustive or to limit the invention to the precise form disclosed. Obviously, many modifications and variations are possible in light of the teaching of the embodiments. The embodiments were chosen and described in order to best explain the principles of the embodiments and the practical application, to thereby enable others skilled in the art to best understand and utilize the invention. This specification is to be limited only by the claims and the full scope and equivalents thereof.

Claims

1. A text processing method, comprising:

receiving a text to be processed;

2. The method of claim 1, the semantic unit expression vector being at least two;

3. The method of claim 2, further comprising, after calculating the weight of any semantic unit expression vector using the fusion layer:

4. The method of claim 1, the semantic unit expression vector comprising at least a feature element of a semantic dimension, a feature element of a prosody dimension, a feature element of a satisfaction dimension, and a feature element of a feedback duration dimension.

5. The method of claim 1, wherein determining, according to the category information, a history number threshold corresponding to the category information, comprises:

6. The method according to claim 5, wherein generating early warning information when determining that the number of texts to be processed corresponding to the category information reaches the history number threshold includes:

7. The method of claim 1, the training step of the text processing model being as follows:

8. The method of claim 7, the step of pre-training the obtained text-expression layer comprising:

9. The method of claim 8, wherein the constructing a second text training sample and a corresponding second text training tag according to the text training corpus, and pre-training a text expression layer according to the second text training sample and the second text training tag, to obtain a pre-trained text expression layer, comprises:

10. The method of claim 8, wherein the constructing a second text training sample and a corresponding second text training tag according to the text training corpus, and pre-training a text expression layer according to the second text training sample and the second text training tag, to obtain a pre-trained text expression layer, comprises:

11. The method of claim 8, wherein the constructing a second text training sample and a corresponding second text training tag according to the text training corpus, and pre-training a text expression layer according to the second text training sample and the second text training tag, to obtain a pre-trained text expression layer, comprises:

12. The method of claim 8, wherein the constructing a second text training sample and a corresponding second text training tag according to the text training corpus, and pre-training a text expression layer according to the second text training sample and the second text training tag, to obtain a pre-trained text expression layer, comprises:

13. The method of claim 7, wherein constructing a second text training sample and a corresponding second text training tag according to the text training corpus, and pre-training a text expression layer according to the second text training sample and the second text training tag, to obtain a pre-trained text expression layer, comprises:

Determining a dialogue sample in the text training corpus;

14. A computing device, comprising:

a memory and a processor;

the memory is configured to store computer executable instructions, the processor being configured to execute the computer executable instructions, which when executed by the processor, implement the steps of the method of any one of claims 1 to 13.