CN113590820A

CN113590820A - Text processing method, device, medium and electronic equipment

Info

Publication number: CN113590820A
Application number: CN202110806970.0A
Authority: CN
Inventors: 冯旻伟; 尹竞成; 杨萌; 阮良
Original assignee: Hangzhou Netease Zhiqi Technology Co Ltd
Current assignee: Hangzhou Netease Zhiqi Technology Co Ltd
Priority date: 2021-07-16
Filing date: 2021-07-16
Publication date: 2021-11-02

Abstract

The embodiment of the disclosure provides a text processing method. The method can comprise the following steps: acquiring a plurality of pieces of text data; obtaining semantic information recognition results respectively corresponding to the text data based on a semantic information recognition network; determining a target text in the plurality of pieces of text data according to the semantic information recognition result; and clustering the determined target texts according to the semantics of the target text representation, and outputting a clustering result. The target text expressing the target semantics is captured by utilizing the neural network through an automatic means, and the target text is classified and summarized according to the semantics represented by the target text to form a knowledge base without manual participation, so that the collection efficiency, timeliness, normalization and safety of the target text are obviously improved, and better experience is brought to a user. In addition, the embodiment of the disclosure provides a text processing device, a medium and an electronic device.

Description

Text processing method, device, medium and electronic equipment

Technical Field

Embodiments of the present disclosure relate to the field of natural language processing, and more particularly, to a text processing method, apparatus, medium, and electronic device.

Background

This section is intended to provide a background or context to the embodiments of the disclosure recited in the claims. The description herein is not admitted to be relevant prior art by inclusion in this section.

In the process of providing service for users, the users often express target semantic information. The information may contain the emotions of dissatisfaction, doubts and the like of the user for the product service. Service personnel need to sensitively capture the target semantic information and reasonably make a solution, so that the user satisfaction can be effectively improved. In order to achieve the purpose, target texts containing target semantic information need to be collected and sorted, and service personnel are trained in a targeted manner based on the target texts, so that the service personnel can capture the target semantic information better.

At present, target texts are mainly collected and sorted manually.

Disclosure of Invention

In this context, embodiments of the present disclosure are intended to provide a text processing method, apparatus, medium, and electronic device.

In a first aspect of embodiments of the present disclosure, there is provided a text processing method, including: acquiring a plurality of pieces of text data; obtaining semantic information recognition results respectively corresponding to the text data based on a semantic information recognition network; the semantic information identification network comprises a neural network obtained by training based on a plurality of first training samples containing target semantic information labeling data; determining a target text containing the target semantic information in the plurality of pieces of text data according to the semantic information identification result; and clustering the determined target texts according to the semantics of the target text representation, and outputting a clustering result.

In some embodiments shown in the present disclosure, the obtaining semantic information recognition results corresponding to the text data based on the semantic information recognition network includes: constructing input features respectively corresponding to the plurality of text data based on the semantics respectively represented by the plurality of text data and/or the behavior features of the user respectively corresponding to the plurality of text data; and obtaining semantic information recognition results respectively corresponding to the plurality of text data by utilizing the semantic information recognition network based on the input characteristics.

In some embodiments shown in the present disclosure, the constructing input features corresponding to the plurality of pieces of text data based on semantics respectively characterized by the plurality of pieces of text data and/or behavior features of a user corresponding to the plurality of pieces of text data includes: for each text data of the plurality of text data, adding a first identifier in the input feature in response to the text data including a phrase that characterizes a target semantic; the first identification indicates that the text data represents target semantics; and/or adding a second identifier in the input characteristic in response to the behavior characteristic of the user representing the user to execute a preset behavior; the second identification indicates that the user performed a preset action.

In some embodiments shown in the present disclosure, the clustering the determined target texts according to semantics of target text representations includes: clustering the determined target texts by using a preset clustering algorithm; in the preset clustering algorithm, the method for determining the similarity between two target texts comprises the following steps: acquiring target phrases which represent preset type semantics and are respectively contained in the two target texts; and determining the similarity between the two target texts according to the similarity between the target phrases respectively contained in the two target texts.

In some embodiments illustrated by the present disclosure, the preset type semantics comprise emotional semantics and a theme; the obtaining of the target phrases respectively contained in the two target texts and representing the preset type semantics includes: acquiring a first phrase representing emotional semantics and a second phrase representing a theme, which are respectively contained in the two target texts; determining the similarity between the two target texts according to the similarity between the target phrases respectively contained in the two target texts, including: acquiring first similarity between first phrases respectively contained in the two target texts and second similarity between second phrases respectively contained in the two target texts; and determining the similarity between the two target texts according to the first similarity and the second similarity.

In some embodiments illustrated by the present disclosure, the method further comprises: acquiring feature vectors respectively corresponding to the two target texts by utilizing the semantic information recognition network; determining a third similarity between the two target texts according to the feature vector; determining the similarity between the two target texts according to the first similarity and the second similarity, including: and determining the similarity between the two target texts according to the first similarity, the second similarity and the third similarity.

In some embodiments shown in the present disclosure, the determining method of the weights corresponding to the first similarity, the second similarity and the third similarity respectively includes: obtaining a plurality of second training samples; the second training sample comprises similarity annotation information; determining a network by utilizing the similarity to obtain the similarity corresponding to the plurality of second training samples respectively; wherein the similarity determination network comprises a neural network generated based on the weights; based on similarity marking information respectively corresponding to the plurality of second training samples and the obtained similarity, adjusting network parameters of the similarity determination network to enable the similarity determination network to converge; determining network parameters of the network according to the converged similarity, and determining the weight.

In some embodiments shown in the present disclosure, the determining, according to the semantic information recognition result, a target text containing the target semantic information in the text data includes: determining a target text in the plurality of text data and a confidence corresponding to the target text according to the semantic information identification result; wherein the confidence level indicates a degree of trustworthiness of text data identified as containing the target semantic information; and outputting the confidence level.

In a second aspect of embodiments of the present disclosure, there is provided an apparatus comprising: the acquisition module is used for acquiring a plurality of pieces of text data; the acquisition module is used for acquiring a plurality of pieces of text data; semantic information recognition results corresponding to the data respectively; the semantic information identification network comprises a neural network obtained by training based on a plurality of first training samples containing target semantic information labeling data; the determining module is used for determining a target text containing the target semantic information in the text data according to the semantic information recognition result; and the clustering and outputting module is used for clustering the determined target texts according to the semantics of the target text representations and outputting clustering results.

In a third aspect of the disclosed embodiments, there is provided a medium storing a computer program for causing a processor to execute a text processing method as illustrated in any one of the preceding embodiments.

In a fourth aspect of embodiments of the present disclosure, there is provided an electronic device comprising: a processor; a memory for storing processor-executable instructions; wherein the processor executes the executable instructions to implement the text processing method as shown in any one of the foregoing embodiments.

In the recorded technical scheme, the target text expressing the target semantic information can be captured by utilizing the neural network through an automatic means, and the target text is classified and summarized according to the semantics represented by the target text to form a knowledge base without manual participation, so that the collection efficiency, timeliness, normalization and safety of the target text are obviously improved, and better experience is brought to a user.

Drawings

The foregoing and other objects, features and advantages of exemplary embodiments of the present disclosure will be readily understood by reading the following detailed description with reference to the accompanying drawings. Several embodiments of the present disclosure are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which:

fig. 1 is a schematic diagram illustrating an application scenario of text processing according to an embodiment of the present disclosure;

FIG. 2 is a method flow diagram illustrating a method of text processing, in accordance with an embodiment of the present disclosure;

FIG. 3 is a schematic flow chart illustrating a target text recognition method according to an embodiment of the present disclosure;

fig. 4 is a schematic diagram illustrating a target text similarity determination method according to an embodiment of the present disclosure;

fig. 5 is a schematic flow chart illustrating a weight determination method according to an embodiment of the present disclosure;

FIG. 6 is a schematic structural diagram of a text processing apparatus according to an embodiment of the present disclosure;

FIG. 7 is a program product for a text processing method, shown in an embodiment of the present disclosure;

fig. 8 is a schematic structural diagram of an electronic device shown in the embodiment of the present disclosure.

In the drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

Detailed Description

The principles and spirit of the present disclosure will be described with reference to a number of exemplary embodiments. It is understood that these embodiments are given solely for the purpose of enabling those skilled in the art to better understand and to practice the present disclosure, and are not intended to limit the scope of the present disclosure in any way. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

As will be appreciated by one skilled in the art, embodiments of the present disclosure may be embodied as a system, apparatus, device, method, or computer program product. Accordingly, the present disclosure may be embodied in the form of: entirely hardware, entirely software (including firmware, resident software, micro-code, etc.), or a combination of hardware and software.

According to the embodiment of the disclosure, a text processing method, a text processing medium, a text processing device and an electronic device are provided.

In this context, it is to be understood that the terms referred to are as follows.

The target semantic information may be semantic information specified in advance according to business requirements. The semantic information may include a plurality of types, and one type of semantic information may be designated in advance for collection in the present disclosure. In some embodiments, the target semantic information may refer to emotional semantic information that implies the emotion of the user. Such as a question expressed by the user for a product or service, or dissatisfaction.

The target text may refer to a text containing the target semantic information. The text data may generally contain semantics, and in some embodiments of the present disclosure, it is desirable to sort out a text containing target semantic information from a stack of text data.

The neural network is an arithmetic mathematical model which simulates the behavior characteristics of the animal neural network and performs distributed parallel information processing. The network can achieve the purpose of processing information by adjusting the mutual connection relationship among a large number of internal nodes.

Natural language processing is a process of performing mathematical modeling on human language, analyzing and processing by using a computer, exploring rules and patterns in the language according to actual requirements and mining values.

Clustering, refers to the process of dividing a collection of physical or abstract objects into multiple class clusters consisting of similar objects.

Moreover, any number of elements in the drawings are by way of example and not by way of limitation, and any nomenclature is used solely for differentiation and not by way of limitation.

The principles and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure.

Due to the manual participation, the problems that the target text collection efficiency is low, the collection is not timely, the sorting standards are not uniform, the data can be leaked and the like exist in the related technology.

Therefore, in the related art, the collation target text cannot be collected well. This is very annoying.

Therefore, an improved target text collection method is needed to capture and classify and summarize target texts containing target semantic information in a timely, efficient, normative and safe manner.

Summary of The Invention

On the one hand, the inventor finds that the neural network has certain learning and recognition capabilities. The network can learn the characteristics of the text containing the target semantic information based on the prior knowledge, and perform analog analysis on the text data based on the learned characteristics so as to identify whether the text data contains the target semantic information. Thus, the inventors can utilize neural networks to identify text that contains targeted semantic information.

On the other hand, the inventor finds that the target texts all have certain semantics. Therefore, the inventor can cluster the target texts based on the semantics of the target text representation, so that a reasonable classification and summarization effect can be obtained.

Having described the general principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

Application scene overview

Referring to fig. 1, fig. 1 is a schematic view illustrating an application scenario of text processing according to an embodiment of the present disclosure.

As shown in fig. 1, the foregoing application scenario may include a terminal such as a mobile phone 1012, a tablet 1013, and a computer 1011, and a server 102 loaded with text processing service logic.

Illustratively, the terminal may collect text data in the form of text, image, voice, etc., and transmit the collected text data to the server 102 for processing.

The server 102 can analyze the text data through the carried service logic to obtain a target text, and classify and summarize the target text to obtain a summarized text 103, so that the collection efficiency, timeliness, normalization and safety of the target text can be remarkably improved.

Exemplary method

Referring to fig. 2, fig. 2 is a flowchart illustrating a method for processing a text according to an embodiment of the disclosure.

The text processing method illustrated in fig. 2 may be applied to an electronic device. The electronic device can execute the method by loading software logic corresponding to the text processing method. The type of the electronic device may be a notebook computer, a mobile phone, a PDA (Personal Digital Assistant), or the like. The type of the electronic device is not particularly limited in this disclosure. The electronic device may also be a client device or a server device, and is not particularly limited herein.

As shown in fig. 2, the text processing method may include S202-S208.

In step S202, a plurality of pieces of text data are acquired.

The text data may comprise any combination of words. The text data can cover certain semantic information, and the text containing the target semantic information can be screened out by analyzing the text data.

In different scenarios, the text data may be acquired in different ways. The present disclosure does not particularly limit the manner of acquiring text data.

For example, in some scenarios, the user makes opinion feedback via voice, such as over the phone. The text data can be acquired by recognizing the voice at this time. For another example, in some scenarios, the user may perform opinion feedback in text form, such as email, short message, chat software, and so on. The text fed back by the user can be directly used as text data. For another example, in some scenarios, text data may be obtained from an image that encompasses user feedback content through text Recognition, such as OCR (Optical Character Recognition).

S204, obtaining semantic information recognition results respectively corresponding to the text data based on a semantic information recognition network; the semantic information identification network comprises a neural network obtained by training based on a plurality of first training samples containing target semantic information labeling data.

The semantic information identifying network may include building a classification or regression network based on a neural network. The network may be used to identify text of the target semantic information.

In some embodiments, the semantic information recognition network may be a NLP (Natural Language Processing) network. The text data can be analyzed by utilizing the NLP network, and a semantic information identification result of the text data is obtained.

To improve the recognition effect, in some embodiments, the semantic information recognition network (hereinafter referred to as a recognition network) may include BERT (Bidirectional Encoder with transform) and a classifier. The BERT can be used for extracting features of text data to obtain feature vectors. The classifier can perform classification processing based on the feature vector output by the BERT to obtain a semantic information identification result aiming at the text data.

The semantic information recognition result may indicate whether the corresponding text data contains the target semantic information and a confidence level that the corresponding text data is recognized as containing the target semantic information. For example, the recognition result for text a may indicate that text a contains target semantics with a confidence of 0.8. As another example, the recognition result for text B may indicate that text B does not contain target semantic information with a confidence of 0.8.

In some embodiments, in performing S204, the recognition network may be trained first using training samples.

In particular, a set of training samples may be obtained, wherein the training samples may correspond to target semantic information tagging data. The target semantic information tagging data may indicate whether a training sample contains target semantic information. In some embodiments, if a sample contains target semantic information, the target semantic information tag data corresponding to the sample may be set to 1. If the sample does not contain the target semantic information, the target semantic information marking data corresponding to the sample can be set to be 0.

The recognition network may then be supervised trained based on training samples until network convergence.

After training is completed, the network can be used to identify whether the input text contains the target semantic information.

In step S204, the obtained text data may be input to the recognition network, so as to obtain semantic information recognition results corresponding to the text data.

S206, determining a target text containing the target semantic information in the text data according to the semantic information recognition result.

In executing S206, semantic information recognition results corresponding to respective text data of the plurality of text data may be determined. And then screening out a target text containing target semantic information from the plurality of text data.

And S208, clustering the determined target texts according to the semantics of the target text representation, and outputting a clustering result.

The semantic meaning of the target text representation may refer to a meaning conforming to language logic represented by a character corresponding to the target text. In some embodiments, the semantics may include emotionalized semantics and themes, among other semantics, divided by type. Wherein, the emotional semantics can include question semantics, negative semantics, comparison semantics and the like. The subject may include a subject indicated by the target text. For example, the subject matter may include price, product quality, warranty policy, and the like.

In some embodiments, the clustering criteria may be determined according to the business requirements, i.e., which semantics are represented based on the target text are determined to perform clustering according to the business requirements. For example, clustering may be performed according to topics, or according to target semantics. As another example, the subject matter can be clustered with the target semantics. Clustering may then be performed according to the determined clustering criteria.

In some embodiments, the semantics of the phrase representations included in the target text may be taken as the semantics of the target text representations. In step S208, target phrases with relatively definite semantics in each target text may be obtained, and then target texts corresponding to target phrases with similar semantics are divided into a cluster class to obtain a clustering result.

In some embodiments, a corpus of phrases corresponding to various semantics, respectively, may be pre-deployed. Wherein the phrases in each phrase bank have similar relatively unambiguous semantics. For example, phrases such as "how long," "how much," "why," etc., that indicate the semantics of the question, can be divided into a corpus of phrases. As another example, phrases such as "too bad," "too slow," "something bad," etc. that represent negative semantics can be partitioned into a corpus of phrases. For another example, phrases such as "than others", "family from others", "business from others", etc. that identify the comparison semantics may be partitioned into a phrase library. For another example, phrases identifying price topics such as "price," "spread," etc. may be partitioned into a corpus of phrases. The method for dividing the phrase library and the types of the various semantics are not listed one by one.

In step S208, word segmentation may be performed on each target text to obtain phrases included in each target text. And then, phrases included in each target text can be matched with phrases in each phrase library deployed in advance, so that target phrases hitting phrases in any phrase library are obtained. And then, according to the semantics indicated by the phrase library hit by the obtained target phrases, dividing the target texts corresponding to the target phrases with similar semantics into a cluster class to obtain a clustering result, and outputting the clustering result.

Therefore, the target text expressing the target semantic information can be captured by utilizing the neural network through an automatic means, the target text is classified and summarized according to the semantics represented by the target text to form a knowledge base without manual participation, the collection efficiency, timeliness, normalization and safety of the target text are obviously improved, and better experience is brought to a user.

The disclosure also provides a text processing method. The steps of the method may be seen in S202-S208. When S204 is executed and semantic information recognition is performed by using the semantic information recognition network, the input features can be improved, so that information beneficial to target text recognition is added to the input features, the target text recognition accuracy is improved, and the target text classification and summarization effect is further improved. The steps of S202 and S206 to S208 will not be described repeatedly below.

Referring to fig. 3, fig. 3 is a schematic flow chart illustrating a target text recognition method according to an embodiment of the present disclosure.

As shown in FIG. 3, in performing S204, S302-S304 may be performed.

And S302, constructing input features respectively corresponding to the text data based on the semantics respectively represented by the text data and/or the behavior features of the user respectively corresponding to the text data.

Three modes can be included in S302. First, input features respectively corresponding to the plurality of pieces of text data may be constructed based on semantics respectively represented by the plurality of pieces of text data.

The semantics of the textual data representation may include target semantics. In some embodiments, the semantics of the text data representation may be determined by target semantics represented by the phrases comprised by the text data. Therefore, the input features are constructed according to the target semantics of the text data, so that the information representing the text semantics can be added to the input features, the recognition of the target text by the recognition network is assisted, the recognition accuracy is improved, and the classification and summarization effects of the target text are further improved.

In some embodiments, a first identification may be added to the input feature in response to the text data including a phrase that characterizes a target semantic for each of the plurality of pieces of text data. Therefore, the semantics of the text data representation can be introduced into the input features, and the identification accuracy is improved.

The first identification indicates that the text data characterizes a target semantic. In this example, the input feature may include a first preset feature bit, which may indicate whether the corresponding text data includes target semantics. In this example, word segmentation processing may be performed on text data to obtain each phrase included in the text data, and then it may be determined whether a semantic meaning represented by each phrase is a target semantic meaning, and if any phrase represents the target semantic meaning, the first identifier may be added to the first preset feature bit. If all the phrases are not used for representing the target semantics, a third identifier can be added to the first preset feature bit. The third identification indicates that the text data does not characterize target semantics.

Second, the input features respectively corresponding to the plurality of pieces of text data may be constructed based on the behavior features of the user respectively corresponding to the plurality of pieces of text data.

The behavioral characteristics of the user may indicate whether the user has performed a behavior that may cause the user to generate the target semantic information. Therefore, the input features are constructed according to the behavior features of the user, so that the input features can increase the information of the user behavior, assist the recognition network to recognize the target semantic information, improve the recognition accuracy and further improve the target text classification and summarization effect.

In some embodiments, a second identifier may be added to the input feature in response to the behavior feature of the user characterizing the user performing a preset behavior; the second identification indicates that the user performed a preset action.

The preset behavior may be preset empirically. For example, purchasing a product or service, a communication action, transacting a VIP action, and the like may be taken as the preset actions. If the user executes the preset behavior, the behavior that the user may execute to generate the target semantic information is described. Therefore, behavior characteristics can be introduced into the input characteristics, and therefore identification accuracy is improved.

In this example, the input feature may include a second preset feature bit, and the feature bit may indicate whether a user corresponding to the text data performs a preset action. In this example, behavior characteristics of a user corresponding to the text data before and after the text data is generated may be obtained, and then whether the user performs a preset behavior may be determined by analyzing the behavior characteristics of the user. If the user is determined to perform the preset action, the second identifier may be added to the second preset feature bit. If the user is determined not to execute the preset action, a fourth identifier may be added to the second preset feature bit.

Thirdly, the input features respectively corresponding to the plurality of text data can be constructed based on the semantics respectively represented by the plurality of text data and the behavior features of the user respectively corresponding to the plurality of text data. Therefore, text semantics and user behavior characteristics can be simultaneously introduced into the input characteristics, the recognition network is assisted to recognize semantic information, the recognition accuracy is improved, and the target text classification and summarization effect is further improved. The detailed description is not provided herein.

S304, based on the input features, utilizing the semantic information identification network to obtain semantic information identification results corresponding to the text data respectively. Therefore, semantic information identification based on the input features can be realized, the identification accuracy is improved, and the target text classification and summarization effect is further improved.

The disclosure also provides a text processing method. The steps of the method may be seen in S202-S208. In S208, the determined target texts may be clustered by using a preset clustering algorithm. The algorithm can determine the similarity between two target texts by utilizing the similarity between target phrases which are included in the two target texts and represent preset type semantics. Therefore, reasonable clustering based on preset type semantics can be achieved, and the classifying and summarizing effects are improved. The steps of S202 to S206 will not be described repeatedly below.

The preset type semantics may refer to preset semantics according to a service requirement.

For example, if clustering according to the theme is required, the preset type of semantics may be set as the semantics of the theme type, and then the similarity between the two target texts may be determined according to the similarity between the target phrases in the two target texts for characterizing the theme.

For another example, if clustering needs to be performed according to the emotional semantics, the preset type of semantics may be set as the emotional semantics, and then the similarity between the two target texts may be determined according to the similarity between the target phrases in the two target texts, which characterize the emotional semantics.

The preset clustering algorithm may be any preset clustering algorithm. For example, the preset clustering algorithm may be a K-Means (K-Means) algorithm, a mean shift clustering algorithm, a density-based clustering algorithm, and the like. This is not to be taken as an example.

It should be noted that the core of the preset clustering algorithm is to determine the similarity between two target texts. The specific steps of the algorithm can refer to the related art, and the following describes a method for determining the similarity between two target texts.

Referring to fig. 4, fig. 4 is a schematic diagram illustrating a method for determining similarity of target texts according to an embodiment of the present disclosure.

As shown in FIG. 4, in determining the similarity of two target texts, S402-S404 may be performed.

And S402, acquiring target phrases which represent preset type semantics and are respectively contained in the two target texts.

In some embodiments, a target phrase library corresponding to the preset type of semantics may be deployed in advance.

In S402, the phrases contained in the two target texts may be obtained first, and then the phrase hitting the target semantic library in each phrase may be used as the target phrase.

S404 may be executed to determine a similarity between the two target texts according to the similarity between the target phrases respectively contained in the two target texts.

In some embodiments, the phrases in the phrase library may be labeled with specific semantic types that they characterize in advance. For example, the semantics of the phrase token can be labeled as query semantics, negative semantics, or which of the semantics is compared. In S404, a similarity between the two target phrases may be determined according to the labeling information corresponding to the two target phrases. If the two target phrases are both target semantics representing the same type (for example, both target semantics representing comparison semantics), the similarity between the two target phrases can be determined to be 1, that is, the similarity between the two target texts can be determined to be 1; if the two target phrase representation target semantic types are not the same (e.g., one representation comparison semantic, the other representation question semantic), then it can be determined that the similarity between the two target phrases is 0 and the similarity between the two target texts is 0.

In some embodiments, in order to accurately determine the similarity between two target texts and further improve the classification and summarization effect, when determining the similarity between two target phrases, phrase vectors corresponding to the two target phrases respectively may be obtained through a pre-trained text processing model, then, normalization and other operations are performed on the distance (e.g., cosine distance, mahalanobis distance, etc.) between the two phrase vectors to obtain the similarity between the two target phrases, and the similarity is determined as the similarity between the two target texts. Therefore, the classification and summarization effect is improved.

In some embodiments, the target text may be clustered according to a combination of emotional semantics and topics. Therefore, the classification and summarization effect is improved. In this example, the preset type semantics may include target semantics and subject.

In executing S402, a first phrase representing an emotional semantic meaning and a second phrase representing a subject, which are respectively contained in the two target texts, may be obtained. The specific method can refer to the method for matching the phrase library, and is not described in detail herein.

Then, in S404, a first similarity between first phrases respectively contained in the two target texts and a second similarity between second phrases respectively contained in the two target texts may be obtained. The method for determining the first similarity and the second similarity may refer to the aforementioned method for determining the similarity between two target phrases, which is not described in detail herein.

The similarity between the two target texts may then be determined according to the first similarity and the second similarity. In some embodiments, weights corresponding to the first similarity and the second similarity respectively may be pre-specified, and then the similarity between the two target texts may be obtained by performing weighted summation on the first similarity and the second similarity.

In some embodiments, the target text may be clustered in conjunction with the emotional semantics, the topic, and the distance between the feature vectors of the two target texts. Therefore, the classification and summarization effect is improved.

In this example, the semantic information recognition network may be further used to obtain feature vectors corresponding to the two target texts, respectively. Wherein, the one-dimensional vector output by the full connection layer in the semantic information recognition network can be used as the feature vector.

And then determining a third similarity between the two target texts according to the feature vectors. In this step, a similarity determination method such as a cosine distance, a mahalanobis distance, or the like may be adopted to determine a distance between two feature vectors, and then a third similarity between two target texts may be obtained by normalizing the distance.

After obtaining the third similarity, in S404, the similarity between the two target texts may be determined according to the first similarity, the second similarity, and the third similarity.

In some embodiments, weights corresponding to the first similarity, the second similarity, and the third similarity may be pre-specified, and then the similarity between the two target texts may be obtained by performing weighted summation on the first similarity, the second similarity, and the third similarity.

In some embodiments, weights corresponding to the three similarities can be obtained through a model training method, so that the accuracy of the similarity between the two target texts can be improved through the accurate weights, and the effect of classification and summarization is further improved.

Referring to fig. 5, fig. 5 is a schematic flow chart illustrating a weight determination method according to an embodiment of the present disclosure.

As shown in FIG. 5, in determining the weights, S502-S508 may be performed.

S502, obtaining a plurality of second training samples; the second training sample includes similarity annotation information. The second training sample may include two text data, and the similarity label information may indicate whether the two text data are similar to each other.

And S504, determining a network by utilizing the similarity to obtain the similarity corresponding to the plurality of second training samples respectively.

Wherein the similarity determination network may include a neural network generated based on the weights. In some embodiments, the mapping function corresponding to the similarity determination network may be generated based on the first similarity, the second similarity, the third similarity, and the weights corresponding to the three similarities, respectively.

In step S504, each second training sample may be input to the similarity determination network, so as to obtain similarities corresponding to the plurality of second training samples.

S506, based on the similarity marking information respectively corresponding to the plurality of second training samples and the obtained similarity, adjusting the network parameters of the similarity determination network to make the similarity determination network converge.

In S506, a preset loss function (e.g., a cross entropy loss function) may be used to calculate loss information, and then a descent gradient may be obtained using the calculated loss information, and the similarity may be adjusted to determine a parameter of a mapping function of the network based on back propagation. In some embodiments, the number of training iterations may be pre-specified, and then the training steps are iteratively performed until the iterations are completed and the network converges.

S508, determining network parameters of the network according to the converged similarity, and determining the weight. After the network is converged, the network parameters are relatively mature, and the similarity can be received by combining the importance degrees of the first similarity, the second similarity and the third similarity, so that an accurate similarity calculation result is obtained. Therefore, based on the network parameters of the network as the determined weights, the weights of the three similarities can be reasonably distributed based on the importance degrees of the first similarity, the second similarity and the third similarity, the accuracy of the similarity between the two target texts is improved, and the classifying and summarizing effects are further improved.

The disclosure also provides a text processing method. The steps of the method may be seen in S202-S208. In S204, the semantic information recognition result obtained by using the recognition network indicates whether the text data includes the target semantic information, and also includes a confidence level. In S206, a target text in the text data and a confidence corresponding to the target text may be determined according to the semantic information recognition result. The confidence is then output.

The confidence level indicates a degree of confidence that text data is identified as containing the target semantic information. Therefore, the confidence corresponding to the target text is output, reference can be provided for training, and the target text with high confidence is processed with priority and emphasis

When the target semantic information is emotional semantic information, the confidence may indicate an emotion value of the emotional semantic information representation. The greater the confidence, the more easily the text data is identified as containing emotional semantic information, i.e., the greater the degree of emotionality. Accordingly, the emotion value derived based on the confidence level may indicate the degree of emotionalization embodied by the emotionalized semantic information.

Therefore, the confidence coefficient is output as an emotion value corresponding to the emotional semantic information, reference can be provided for training, and the target text corresponding to the emotional semantic information with high emotional degree is processed with priority and emphasis.

The following description will be made with reference to an application scenario of fig. 1. It should be noted that the foregoing application scenarios are merely illustrative for the convenience of understanding the spirit and principles of the present disclosure, and that embodiments of the present disclosure are not limited in this respect. Rather, embodiments of the present disclosure may be applied to any scenario where applicable.

In the scenario shown in fig. 1, a marketing training as a Service (SAAS) platform may be deployed in the server 102.

The marketing training platform carries target text screening logic and clustering logic. The target text contains emotional semantic information. The target text screening logic may use a pre-trained semantic information recognition network for relevant screening. The distance logic may perform the distance using a K-Means algorithm. The formula for calculating the similarity by the K-Means algorithm comprises the following steps:

P(A，B)＝w1*cosine(A，B)+w2*sentiment(A，B)+w3*compare(A，B)+w4*topic(A，B)。

wherein, A and B represent two target texts of which the similarity needs to be calculated.

P (A, B) represents the degree of similarity.

cosine (A, B) represents cosine similarity between the feature vectors corresponding to A and B respectively.

sentiment (A, B) represents the similarity between the emotional semantics of the A and B tokens.

compare (A, B) represents the similarity between the comparative semantics of the A and B tokens.

topic (a, B) represents the similarity between topics characterized by a and B.

w1, w2, w3 and w4 represent the weights corresponding to the 4 similarities. May be preset empirically.

The

terminals

1011 and 1013 can collect data such as text, voice, image, etc. for communicating with the client in real time, and send the collected data to the server 102 for processing.

The server 102 may first extract the textual data via marketing training software on a regular basis (e.g., 0 o' clock per day).

Then, a plurality of target texts in the text data can be obtained through the target text screening logic. And determining the confidence corresponding to the target text as the emotion value corresponding to the emotional semantic information contained in the target text.

Then, clustering of the target texts may be completed through clustering logic, and the summary text 103 is generated to update the knowledge base. The summarized text 1 may show the classified and summarized target texts and emotion values corresponding to emotional semantic information included in each target text.

Thereafter, the server may periodically (e.g., 1 am each day) push the summary text to a responsible person (e.g., team leader, manager, etc.) responsible for marketing training, who trains the service personnel for the target text of the summary text record.

Therefore, the target text expressing the emotion semantics of the user can be captured by utilizing the neural network through an automatic means, and the knowledge base is formed according to the semantic classification summary represented by the target text without manual participation, so that the collection efficiency, timeliness, normalization and safety of the target text are remarkably improved, and better experience is brought to users.

Exemplary devices

Having described the method of the exemplary embodiment of the present disclosure, next, a text processing apparatus of the exemplary embodiment of the present disclosure will be described with reference to fig. 6. The text processing device is used for realizing the text processing method shown in any one of the preceding embodiments.

Referring to fig. 6, fig. 6 is a schematic structural diagram of a text processing apparatus according to an embodiment of the disclosure.

As shown in fig. 6, the apparatus 60 may include:

an obtaining module 61, configured to obtain multiple pieces of text data;

the recognition module 62 is configured to, based on the semantic information recognition network, obtain semantic information recognition results corresponding to the plurality of pieces of text data, respectively; the semantic information identification network comprises a neural network obtained by training based on a plurality of first training samples containing target semantic information labeling data;

a determining module 63, configured to determine, according to the semantic information recognition result, a target text including the target semantic information in the plurality of pieces of text data;

and the clustering and outputting module 64 is used for clustering the determined target texts according to the semantics of the target text representations and outputting clustering results.

In some embodiments, the identification module 62 is specifically configured to:

constructing input features respectively corresponding to the plurality of text data based on the semantics respectively represented by the plurality of text data and/or the behavior features of the user respectively corresponding to the plurality of text data;

and obtaining semantic information recognition results respectively corresponding to the plurality of text data by utilizing the semantic information recognition network based on the input characteristics.

for each text data of the plurality of text data, adding a first identifier in the input feature in response to the text data including a phrase that characterizes a target semantic; the first identification indicates that the text data represents target semantics; and/or the presence of a gas in the gas,

responding to the behavior characteristics of the user to represent that the user executes preset behaviors, and adding a second identifier in the input characteristics; the second identification indicates that the user performed a preset action.

In some embodiments, the clustering and outputting module 64 is specifically configured to:

clustering the determined target texts by using a preset clustering algorithm;

in the preset clustering algorithm, the method for determining the similarity between two target texts comprises the following steps:

acquiring target phrases which represent preset type semantics and are respectively contained in the two target texts;

and determining the similarity between the two target texts according to the similarity between the target phrases respectively contained in the two target texts.

acquiring a first phrase representing emotional semantics and a second phrase representing a theme, which are respectively contained in the two target texts;

determining the similarity between the two target texts according to the similarity between the target phrases respectively contained in the two target texts, including:

acquiring first similarity between first phrases respectively contained in the two target texts and second similarity between second phrases respectively contained in the two target texts;

and determining the similarity between the two target texts according to the first similarity and the second similarity.

In some embodiments, the apparatus 60 further comprises:

the third similarity determining module is used for utilizing the semantic information identification network to obtain the characteristic vectors respectively corresponding to the two target texts;

determining a third similarity between the two target texts according to the feature vector;

the clustering and outputting module 64 is specifically configured to:

and determining the similarity between the two target texts according to the first similarity, the second similarity and the third similarity.

In some embodiments, the apparatus 60 further comprises:

a weight determining module 63, configured to obtain a plurality of second training samples; the second training sample comprises similarity annotation information;

determining a network by utilizing the similarity to obtain the similarity corresponding to the plurality of second training samples respectively; wherein the similarity determination network comprises a neural network generated based on the weights;

based on similarity marking information respectively corresponding to the plurality of second training samples and the obtained similarity, adjusting network parameters of the similarity determination network to enable the similarity determination network to converge;

determining network parameters of the network according to the converged similarity, and determining the weight.

In some embodiments, the determining module 63 is specifically configured to:

determining a target text in the plurality of text data and a confidence corresponding to the target text according to the semantic information identification result; wherein the confidence level indicates a degree of trustworthiness of text data identified as containing the target semantic information;

and outputting the confidence level.

Exemplary Medium

Having described the method and apparatus of the exemplary embodiments of the present disclosure, a readable storage medium of the exemplary disclosure is described next with reference to fig. 7. The storage medium stores a computer program for causing a processor to execute a text processing method as in any one of the preceding embodiments.

Referring to fig. 7, fig. 7 is a flowchart illustrating a program product 70 applied to a text processing method according to an embodiment of the disclosure.

In some embodiments shown, the foregoing text processing method may be implemented by a program product 70, such as may employ a portable compact disc read only memory (CD-ROM) and include program code, and may be run on a device, such as a personal computer. However, the program product of the present disclosure is not limited thereto, and in this document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

A computer readable signal medium may include a propagated data signal with readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A readable signal medium may also be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RE, etc., or any suitable combination of the foregoing.

Program code for carrying out operations for the present disclosure may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the C language or similar programming languages. The program code may execute entirely on the consumer electronic device, partly on a remote electronic device, or entirely on the remote electronic device or server. In the case of remote electronic devices, the remote electronic devices may be connected to the consumer electronic device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external electronic device (e.g., through the internet using an internet service provider).

Exemplary electronic device

Having described the methods, apparatus and media of the exemplary embodiments of the present disclosure, an electronic device of the exemplary disclosure is next described with reference to fig. 8. The apparatus comprises: a processor; a memory for storing processor-executable instructions; wherein the processor executes the executable instructions to implement the text processing method as shown in any one of the foregoing embodiments.

Referring to fig. 8, fig. 8 is a schematic structural diagram of an electronic device shown in the embodiment of the present disclosure.

The electronic device 800 shown in fig. 8 is only an example and should not bring any limitations to the functionality and scope of use of the embodiments of the present disclosure.

As shown in fig. 8, the electronic device 800 is represented in the form of a general electronic device. The components of the electronic device 800 may include, but are not limited to: the aforementioned at least one processor 801, the aforementioned at least one storage processor 802, and a bus 803 connecting the various system components (including the processor 801 and the storage processor 802).

The bus 803 includes a data bus, a control bus, and an address bus.

The storage processor 802 may include readable media in the form of volatile memory, such as Random Access Memory (RAM)8021 and/or cache memory 8022, and may further include readable media in the form of non-volatile memory, such as Read Only Memory (ROM) 8023.

Storage processor 802 may also include a program/utility 8025 having a set (at least one) of program modules 8024, such program modules 8024 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which, or some combination thereof, may comprise an implementation of a network environment.

The electronic device 800 may also communicate with one or more external devices 804 (e.g., keyboard, pointing device, etc.).

Such communication may be through input/output (I/O) interfaces 805. Also, the electronic device 800 may communicate with one or more networks (e.g., a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network, such as the internet) via the network adapter 806. As shown in FIG. 8, the network adapter 806 communicates with the other modules of the electronic device 800 via the bus 803. It should be understood that although not shown in the figures, other hardware and/or software modules may be used in conjunction with electronic device 800, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.

It should be noted that although in the above detailed description several units/modules or sub-units/modules of the text processing apparatus are mentioned, such division is merely exemplary and not mandatory. Indeed, the features and functionality of two or more of the units/modules described above may be embodied in one unit/module, in accordance with embodiments of the present disclosure. Conversely, the features and functions of one unit/module described above may be further divided into embodiments by a plurality of units/modules.

Further, while the operations of the disclosed methods are depicted in the drawings in a particular order, this does not require or imply that these operations must be performed in this particular order, or that all of the illustrated operations must be performed, to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step execution, and/or one step broken down into multiple step executions.

While the spirit and principles of the present disclosure have been described with reference to several particular embodiments, it is to be understood that the present disclosure is not limited to the particular embodiments disclosed, nor is the division of aspects, which is for convenience only as the features in such aspects may not be combined to benefit. The disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A text processing method, comprising:

acquiring a plurality of pieces of text data;

obtaining semantic information recognition results respectively corresponding to the text data based on a semantic information recognition network; the semantic information identification network comprises a neural network obtained by training based on a plurality of first training samples containing target semantic information labeling data;

determining a target text containing the target semantic information in the plurality of pieces of text data according to the semantic information identification result;

and clustering the determined target texts according to the semantics of the target text representation, and outputting a clustering result.

2. The method according to claim 1, wherein obtaining semantic information recognition results corresponding to the plurality of pieces of text data respectively based on the semantic information recognition network comprises:

3. The method of claim 1, wherein clustering the determined target texts according to semantics of target text representations comprises:

clustering the determined target texts by using a preset clustering algorithm;

4. The method of claim 3, the preset type of semantics comprising emotional semantics and subject matter; the obtaining of the target phrases respectively contained in the two target texts and representing the preset type semantics includes:

5. The method of claim 4, further comprising:

acquiring feature vectors respectively corresponding to the two target texts by utilizing the semantic information recognition network;

determining the similarity between the two target texts according to the first similarity and the second similarity, including:

6. The method according to claim 5, wherein the determination method of the weight corresponding to each of the first similarity, the second similarity and the third similarity comprises:

obtaining a plurality of second training samples; the second training sample comprises similarity annotation information;

7. The method of claim 1, wherein the determining a target text containing the target semantic information from the plurality of pieces of text data according to the semantic information recognition result comprises:

and outputting the confidence level.

8. A text processing apparatus comprising:

the acquisition module is used for acquiring a plurality of pieces of text data;

the recognition module is used for recognizing the network based on the semantic information to obtain semantic information recognition results corresponding to the text data respectively; the semantic information identification network comprises a neural network obtained by training based on a plurality of first training samples containing target semantic information labeling data;

the determining module is used for determining a target text containing the target semantic information in the text data according to the semantic information recognition result;

and the clustering and outputting module is used for clustering the determined target texts according to the semantics of the target text representations and outputting clustering results.

9. An electronic device, the device comprising:

a processor;

a memory for storing processor-executable instructions;

wherein the processor implements the text processing method of any one of claims 1-7 by executing the executable instructions.

10. A computer-readable storage medium, which stores a computer program for causing a processor to execute the text processing method according to any one of claims 1 to 7.