CN116089614A

CN116089614A - Text marking method and device

Info

Publication number: CN116089614A
Application number: CN202310099070.6A
Authority: CN
Inventors: 袁堃平; 兰金鹤; 夏回美; 唐志慧; 张晋
Original assignee: Hangzhou Lingyang Intelligent Service Co ltd
Current assignee: Hangzhou Lingyang Intelligent Service Co ltd
Priority date: 2023-01-12
Filing date: 2023-01-12
Publication date: 2023-05-09
Anticipated expiration: 2043-01-12
Also published as: CN116089614B

Abstract

The embodiment of the specification provides a text marking method and a text marking device, wherein the method comprises the following steps: receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model; obtaining rule labels of rule dimensions corresponding to the output of the rule matching model, classification labels of classification dimensions corresponding to the output of the classification matching model, and semantic labels of semantic dimensions corresponding to the output of the semantic matching model; determining tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to the filtering result. Identifying the text to be processed from multiple dimensions, identifying rule labels of rule dimensions, classification labels of classification dimensions and semantic labels of semantic dimensions, filtering the labels in different dimensions, further improving the identification accuracy and obtaining more accurate identification results.

Description

Text marking method and device

Technical Field

The embodiment of the specification relates to the technical field of natural language processing, in particular to a text marking method and a strategy adjusting method. One or more embodiments of the present specification relate to a text marking apparatus, a policy adjustment apparatus, a computing device, and a computer-readable storage medium.

Background

With the rapid development of internet technology, people enter a big data age and generate massive data every day, so that analysis of the massive data and acquisition of valuable information become hot spots for people's common relations. In the news field, the label setting can be carried out on the news information text, and the auditing and the delivery of the news information text can be carried out according to the label in the later period; in the service field, the label setting can be carried out on the customer evaluation, and the customer intention can be known and the selling condition of the product can be analyzed according to the label in the later period.

In the prior art, most of the classified marking of texts is manually performed by industry-related personnel based on industry experience, and under the condition of large data volume, the marking efficiency is low and the high dependence on knowledge and experience is required. Or the text is marked through the pre-training model, but the label accuracy of model prediction is lower due to different text expression modes with the same meaning under different conditions, and the condition that labels corresponding to the texts with the same semantics are inconsistent exists. Therefore, how to accurately identify the label corresponding to the text is a problem to be solved.

Disclosure of Invention

In view of this, the embodiments of the present specification provide a text marking method and a policy adjustment method. One or more embodiments of the present specification relate to a text marking apparatus, a policy adjustment apparatus, a computing device, a computer-readable storage medium, and a computer program, which solve the technical drawbacks of the prior art.

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

receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model;

obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model;

determining tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively;

and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to filtering results.

According to a second aspect of embodiments of the present specification, there is provided a policy adjustment method, including:

determining a target object, a target operation strategy corresponding to the target object and a target text to be processed corresponding to the target object;

performing text marking processing of a text marking method on the target text to be processed to obtain a problem label corresponding to the target text to be processed;

and determining a target problem of the target object based on the problem label, and adjusting the target operation strategy based on the target problem.

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

determining a user evaluation text of a target object, and respectively inputting the user evaluation text into a rule matching model, a classification matching model and a semantic matching model;

And filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the user evaluation text according to filtering results.

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

the input module is configured to receive a text to be processed and input the text to be processed into a rule matching model, a classification matching model and a semantic matching model respectively;

the obtaining module is configured to obtain rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model and semantic labels of corresponding semantic dimensions output by the semantic matching model;

the determining module is configured to determine tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively;

and the filtering module is configured to filter the rule labels, the classification labels and the semantic labels based on the label attribute information, and determine a target label corresponding to the text to be processed according to a filtering result.

According to a fifth aspect of embodiments of the present specification, there is provided a policy adjustment device, comprising:

The determining module is configured to determine a target object, a target operation strategy corresponding to the target object and a target text to be processed corresponding to the target object;

the marking module is configured to perform text marking processing of a text marking method on the target text to be processed to obtain a problem label corresponding to the target text to be processed;

and the adjusting module is configured to determine a target problem of the target object based on the problem label and adjust the target operation strategy based on the target problem.

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

the input module is configured to determine a user evaluation text of a target object and input the user evaluation text into the rule matching model, the classification matching model and the semantic matching model respectively;

And the filtering module is configured to filter the rule labels, the classification labels and the semantic labels based on the label attribute information, and determine a target label corresponding to the user evaluation text according to a filtering result.

According to a seventh aspect of embodiments of the present specification, there is provided a computing device comprising a memory, a processor and computer instructions stored on the memory and executable on the processor, the processor implementing the steps of the text marking method, policy adjustment method when executing the computer instructions.

According to an eighth aspect of embodiments of the present specification, there is provided a computer readable storage medium storing computer instructions which, when executed by a processor, implement the steps of the text marking method, policy adjustment method.

According to a ninth 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 marking method, the policy adjustment method described above.

The text marking method provided by the specification comprises the steps of receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model; obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model; determining tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to filtering results.

According to the embodiment of the specification, the text to be processed is identified from multiple dimensions based on the rule matching model, the classification matching model and the semantic matching model, rule labels in the rule dimensions, classification labels in the classification dimensions and semantic labels in the semantic dimensions are identified, the labels in different dimensions are filtered, the identification accuracy is further improved, and more accurate identification results are obtained so that the target labels corresponding to the text to be processed are determined.

Drawings

FIG. 1 is a schematic diagram showing the effect of a text marking method according to an embodiment of the present disclosure

FIG. 2 is a flow chart of a text marking method provided in one embodiment of the present disclosure;

FIG. 3 is a flowchart of a policy adjustment method using text marking method according to one embodiment of the present disclosure;

FIG. 4 is a process flow diagram of a text marking method according to one embodiment of the present disclosure;

FIG. 5 is a flow chart of another text marking method provided by one embodiment of the present disclosure;

FIG. 6 is a schematic diagram of a text marking device according to an embodiment of the present disclosure;

fig. 7 is a schematic structural diagram of a policy adjustment device according to an embodiment of the present disclosure

FIG. 8 is a schematic diagram of another text marking device according to one embodiment of the present disclosure;

FIG. 9 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 description refers to any or all possible combinations including 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.

First, terms related to one or more embodiments of the present specification will be explained.

Text classification: text classification is a subtask of the natural language processing field aimed at classifying input text into one or more of preset categories by some techniques of the natural language processing field. Text classification is widely applied to scenes such as spam filtering, news topic classification, emotion recognition and the like.

Migration learning: the transfer learning is a machine learning method, i.e., a deep learning method that uses a model developed for task a as an initial point and reuses the model developed for task B.

Model distillation: and the knowledge learned by one large model or a plurality of models is migrated to another lightweight single model, so that the deployment is convenient. In short, a new small model is used for learning the prediction result of a large model, and an objective function is changed, so that the effect of compressing the model is achieved in the range of acceptable accuracy drop.

BERT model: BERT pre-trains deep bi-directional representations by jointly adjusting both left and right contexts at all layers, and further enhances understanding of length Cheng Yuyi by assembling long sentences as input. The BERT model can be finely tuned to be widely used for various tasks, and only one output layer is needed to be additionally added, so that model structure adjustment for the tasks is not needed, and a very good effect is achieved on some tasks such as text classification, semantic understanding and the like.

Currently in the service industry, merchants receive a variety of different ratings or recommendations from customers. In the face of massive user original sound, merchants cannot acquire the meaning of all customers quickly, so that the problems of customer feedback cannot be changed in time, the subsequent use experience of the customers is poor, and bad influence is brought to the merchants. Therefore, the merchant can extract information aiming at massive user voices, obtain each evaluation or suggested keyword as a label, and determine the meaning of the customer according to the label. However, since the number of evaluations is large and the evaluation methods are different for each person, there are often many different expression methods for the same meaning, and thus the extraction result is not very accurate, and the label recognition is performed only according to the matching rule, so that the method has no generalization capability of text and low coverage rate of original sound.

Based on this, in the present specification, a text marking method and a policy adjustment method are provided, and the present specification relates to a text marking apparatus, a policy adjustment apparatus, a computing device, a computer-readable storage medium, and a computer program, which are described in detail in the following embodiments one by one.

Fig. 1 is an effect schematic diagram of a text marking method according to an embodiment of the present disclosure, where a text to be processed may be a user acoustic text such as a collected user evaluation, a user suggestion, and the like, and the text to be processed is input into a rule matching model, a classification matching model, and a semantic matching model, where the three models analyze the text to be processed from a rule dimension, a classification dimension, and a semantic dimension, and determine a label corresponding to the text to be processed, so as to obtain a rule label in the rule dimension, a classification label in the classification dimension, and a semantic label in the semantic dimension. After obtaining the labels in three dimensions, further filtering the rule labels, the classification labels and the semantic labels to obtain more accurate target labels serving as labels corresponding to the text to be processed. The label can be characterized as information such as keywords and categories of the text to be processed, the central thought of the user voice is reflected, and the user intention can be clearly known by the project party and the merchant, so that better service experience is provided for the user. It should be noted that, the user evaluation text, the user suggestion text or the user original sound text related to the application are all information and data authorized by the user or fully authorized by each party, wherein the user original sound text includes, but is not limited to, user data, user history behavior and the like, the user data includes, but is not limited to, data for analysis, stored data, displayed data and the like, and the collection, the use and the processing of the related data need to comply with the related laws and regulations and standards of related countries and regions, and are provided with corresponding operation entries for the user to select authorization or rejection.

Fig. 2 shows a flowchart of a text marking method according to an embodiment of the present disclosure, including steps 202 to 208.

Step 202: receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model.

The text to be processed can be understood as text to be subjected to marking processing, and the text to be processed can be user original sound text, including evaluation text of users in online shopping items, suggestion text of users in consumption items and the like. In practical application, the text to be processed can be the use experience of the user on the commodity in the online shopping project, or the evaluation of the project in the project process, etc. Such as evaluation data of related commodities in the links of items such as delivery, commodity quality, after-sales and the like in online shopping items. Because of the many acoustic text expressions that users focus on, direct aggregation cannot be categorized. And the user is marked in the original sound, the text unstructured data is labeled, and the association analysis can be further carried out. Such as analyzing how many users are "consulting a shipping question," and the user's acoustic text in such a question may be "how shipped", "my orders did not occur? "etc., these user acoustic texts all correspond to shipping problems, but cannot be directly categorized and aggregated. The specification therefore identifies the text to be processed from three dimensions by a rule matching model, a class matching model, a semantic matching model.

In practical application, the rule matching model is based on the thought of 'strict matching', and has corresponding keywords, excluding words and proper sentence patterns (positive sentences, question sentences, negative sentences and the like) for each initial tag in the tag library, and the model can be directly matched with corresponding tags for texts conforming to tag rules. The classification matching model classifies each sentence of user voice onto proper labels, and the processing process of the classification matching model is to input the user voice text into an encoder in the model to obtain high-dimensional sentence vectors, and then map the sentence vectors to the classification labels, so as to determine the labels corresponding to the text. And when the semantic matching model is used, carrying out semantic feature matching on the example sentences corresponding to the user acoustic text and the labels, namely, firstly extracting semantic features of the user acoustic text by the semantic matching model, then comparing the semantic features of the example sentences corresponding to the labels with the semantic features of the example sentences, selecting the example sentences with higher semantic feature similarity with the user acoustic text, and determining the labels corresponding to the example sentences as labels corresponding to the user acoustic text.

In specific implementation, the rule matching model, the classification matching model and the semantic matching model are all tag recognition models, and the rule matching model is constructed by the following steps: determining a training text and a training label corresponding to the training text; inputting the training text into an initial rule matching model to obtain a prediction rule label output by the initial rule matching model; calculating a rule matching model loss value according to the training label and the prediction rule label, and adjusting model parameters of the initial rule matching model based on the rule matching model loss value; and continuing training the initial rule matching model until the model reaches training conditions, and obtaining the rule matching model according to a training result.

The training text can be understood as a user original sound text for training the rule matching model, the training label corresponding to the training text can be understood as a label corresponding to training text matching, namely the training text is input into a pre-trained initial rule matching model, a model loss value is calculated by comparing the predicted rule label output by the initial rule matching model with the training label, so that model parameters of the initial rule matching model are adjusted, the training of the initial rule matching model is continued until the model reaches training conditions, wherein the training conditions can enable training rounds or the model loss value to reach a threshold value, and the trained rule matching model can be obtained after the training conditions are reached.

The construction of the classification matching model comprises the following steps: determining a training text and a training label corresponding to the training text; inputting the training text into an initial parent classification matching model to obtain a prediction classification label output by the initial parent classification matching model; calculating a classification matching model loss value according to the training label and the prediction classification label, and adjusting model parameters of the initial parent classification matching model based on the classification matching model loss value; continuing training the initial parent classification matching model until the model reaches training conditions, and obtaining a parent classification matching model according to training results; and training an initial sub-classification matching model based on model parameters of the parent classification matching model, and obtaining a classification matching model according to training results.

The training text is input into an initial parent classification matching model, a high-dimensional sentence vector is obtained by an encoder in the initial parent classification matching model, and then mapped to a classification label to determine a prediction classification label, a model loss value is calculated according to comparison of the prediction classification label and a training label, so that model parameters of the initial parent classification matching model are adjusted, and the initial parent classification matching model is continuously trained until training conditions are reached, so that a trained parent classification matching model is obtained. And then, in order to reduce the model volume and improve the model processing efficiency, model distillation can be performed on the parent classification matching model, firstly the weight of the parent classification matching model is saved, and then the parent classification matching model is distilled to the child classification matching model, so that the child classification matching model can ensure the accuracy of multiple classifications and can also exert the speed advantage of the lightweight model in large-scale calculation.

The construction of the semantic matching model comprises the following steps: determining a training text and a training label corresponding to the training text, and constructing a positive and negative sample pair according to the training text; inputting the positive and negative sample pairs into an initial semantic matching model to obtain a predicted semantic tag output by the initial semantic matching model; calculating a semantic matching loss value according to the training label and the prediction semantic label, and adjusting model parameters of the initial semantic matching model based on the semantic matching loss value; and continuing training the initial semantic matching model until the model reaches training conditions, and obtaining the semantic matching model according to training results.

The semantic matching problem can be abstracted into the feature matching degree of two sentences, the text S1 and the text S2 are subjected to feature interaction in different calculation modes, then the relation between the text S1 and the text S2 is learned through a deep learning model, namely score=f (S1, S2), F is a model of us, and the text S1 and the text S2 are two texts to be matched. Taking the BERT model as an example, the BERT model is regarded as a feature extractor, a text sequence S is input, a sequence of vectors [ C, T1, T2.] is output, and meanwhile, we can use the vector after the output vector passes through the connection layer as the sentence vector of the whole sentence. score=f (S1, S2) =cos (B (S1), B (S2)), where B generates sentence vectors for the BERT model, extracting semantic representations in the text. In practical application, we can construct positive and negative samples according to training texts, and the positive samples are constructed by combining different phones under the same label in pairs. Such as "this taste is good" i like this taste ", which is a set of positive samples, semantically similar and belonging to the same tag. The negative samples are a group of negative samples, which are formed by combining the different label-based techniques in pairs, such as "the taste is good and the make-up effect is poor". The negative samples can be combined in pairs by mutually exclusive samples under the labels, so that a high-difficulty negative sample is constructed, and the high-difficulty negative sample is a group of high-difficulty negative samples, such as ' good taste and ' bad taste '. The method comprises the steps of inputting positive and negative sample pairs into an initial semantic matching model, obtaining a predicted semantic label output by the initial semantic matching model, calculating a model loss value according to the predicted semantic label and a training label, adjusting training parameters of the initial semantic matching model, and continuing training the initial semantic matching model until the model reaches training conditions, so as to obtain a trained semantic matching model.

In conclusion, a rule matching model, a classification matching model and a semantic matching model are obtained through training by different training methods, so that the three models can be utilized to conduct tag identification in terms of multiple dimensions in the follow-up process of tag identification, and the generalization capability and the accuracy of tag identification are improved.

Step 204: obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model.

The rule label can be understood as a label output by the rule matching model in a rule dimension, the classification label can be understood as a label output by the classification matching model in a classification dimension, and the semantic label can be understood as a label output by the semantic matching model in a semantic dimension.

In practical application, the rule matching model performs regular matching on the user acoustic text through the attributes of the tag, such as keywords, exclusion words, adaptation sentence patterns and the like, so as to obtain a rule tag with rule dimensions. And the classification matching model is filtered by combining a label confidence threshold value to obtain classification labels of classification dimensions. And carrying out semantic matching on the semantic matching model and the user original sound text through the sample sentences corresponding to the labels, and determining semantic labels of semantic dimensions.

In a specific embodiment of the present disclosure, a text to be processed is input into a rule matching model, the rule matching model performs regular matching with the text to be processed according to attributes such as keywords, excluded words, and adaptive sentence patterns corresponding to each tag, and a rule tag corresponding to the text to be processed is determined according to a matching result.

In another specific embodiment of the present disclosure, a text to be processed is input into a classification matching model, the classification matching model encodes the text to be processed, and a high-dimensional sentence vector corresponding to the text to be processed is obtained and mapped onto a classification label to determine a classification label corresponding to the text to be processed.

In another specific embodiment of the present disclosure, a text to be processed is input into a semantic matching model, the semantic matching model determines text features corresponding to the text to be processed, and matches the text features of example sentences of each tag, thereby determining semantic tags corresponding to the text to be processed.

Further, in order to prevent the text to be processed from matching a plurality of texts to be processed to cause inaccurate tag recognition results, regular matching needs to be performed according to a plurality of matching information corresponding to each tag, and specifically, obtaining a rule tag with a rule dimension corresponding to the rule matching model output includes: matching the text to be processed with the initial tag based on tag matching information corresponding to the initial tag through the rule matching model; and determining rule labels corresponding to the rule dimensions according to the matching result and outputting the rule labels.

The initial labels can be understood as labels defined by the item side according to some user original sounds in advance, corresponding label matching information is set for each initial label, the label matching information can be understood as attribute information of the labels, and when a certain sentence of original sound text accords with the matching information of the labels, the original sound text is matched with the labels. If the keyword of a certain tag is A and the adaptive sentence pattern is a statement sentence, when a certain original sound text is the statement sentence and the content contains the keyword A, the tag is described as the tag of the original sound text.

In practical application, the tag matching information may further include a tag exclusion word, and when the tag exclusion word exists in a certain acoustic text, it is indicated that the tag does not correspond to the acoustic text.

In sum, through the matching rules in the rule matching model, regular matching can be carried out on the original sound text according to the label matching information of the labels, so that an accurate label identification result is obtained.

Further, avoiding that the same semantic but different acoustic texts are not recognized by the same label, the label recognition can be performed on the text classification, specifically obtaining the classification label of the classification matching model output corresponding classification dimension includes: coding the text to be processed through the classification matching model to obtain a coding vector corresponding to the text to be processed; and determining and outputting the classification label corresponding to the classification dimension based on the coding vector and the vector label mapping relation.

The text to be processed is encoded by inputting a batch of text corpus into an encoder to obtain high-dimensional sentence vectors, and mapping the sentence vectors to classification labels, so as to determine the classification labels corresponding to the text corpus.

In practical application, since different dialects may exist for expressing the same semantics among different users, a classification matching model may be used in order to be able to identify the same tag in the text of the different dialects. If the existing text corpus can help to make an invoice, does not make an invoice first, and does not make an electronic invoice first is input into a classification matching model, the classification matching model converts each sentence into a text vector, and can find out that the text vector is mapped to the same classification label of the invoice, and then the invoice can be determined to be a label shared by the sentences.

Based on the method, the same label can be determined by the classification matching model through the acoustic texts of different phones, so that the generalization capability of label identification is expanded, and the coverage rate of label identification is improved.

Further, although the classification matching model can identify the same label in the original sound text of different vocabularies, there may be a situation that the identification is missed or even is wrong, in order to avoid the situation, label identification may be performed from text semantics, and specifically, obtaining the semantic matching model outputs a semantic label corresponding to a semantic dimension includes: obtaining semantic information of the text to be processed through the semantic matching model analysis; and matching the semantic information with the reference semantic information corresponding to the initial label, and determining the semantic label corresponding to the semantic dimension according to a matching result.

The semantic information of the text to be processed can be understood as semantic features extracted from the text to be processed, the reference semantic information can be understood as semantic feature information extracted from example sentences of the initial tag, and by comparing the semantic information with the reference semantic information, whether the semantics of the text to be processed and the semantics of example sentences of the initial tag are identical can be determined, so that the initial tag of the example sentences with identical semantics of the text to be processed is determined as the semantic tag of the text to be processed.

In practical application, each initial tag is provided with a plurality of example sentences, for example, the example sentences with good tag taste can be 'good taste', 'i like eating the taste', and when the text to be processed is 'like the taste' well, the semantic features of the example sentences with good tag taste are determined to be similar to the semantic features of the text to be processed, so that the tag with good taste can be determined to be the text to be processed.

Based on the method, the label can be identified from the dimension of the text semantics through the semantic matching model, and the accuracy of label identification is further improved.

Step 206: and determining label attribute information corresponding to the rule labels, the classification labels and the semantic labels respectively.

The rule labels, the classification labels and the semantic labels are labels identified by the rule matching model, the classification matching model and the semantic matching model respectively, a plurality of labels are arranged in each type of labels, and each label has label attribute information corresponding to each label. The tag attribute information includes tag sentence pattern information corresponding to the tag, such as a positive sentence, a question sentence, etc., and priority information corresponding to the tag, etc. The labels output by all the models can be further screened according to the label attribute information, so that the labels more conforming to the text to be processed are determined, and the accuracy of label identification is improved.

In practical application, the rule matching model outputs a plurality of rule labels, the classification matching model outputs a plurality of classification labels, the semantic matching model outputs a plurality of semantic labels, all labels have label attribute information corresponding to the labels, and the labels can be further screened according to the label attribute information, so that the labels more conforming to the text to be processed are selected.

Step 208: and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to filtering results.

The filtering processing of the rule labels, the classification labels and the semantic labels according to the label attribute information comprises confidence threshold filtering processing, sentence filtering processing, mutual exclusion type filtering processing and priority filtering processing, wherein the confidence threshold filtering processing can be understood as removing labels with confidence less than a confidence threshold; sentence pattern filtering processing can be understood as removing tags which do not meet sentence pattern requirements; the mutual exclusion type filtering processing can be understood as removing the label with lower confidence in the mutual exclusion label; the priority filtering process may be understood as the elimination of lower priority tags.

In practical application, the filtering modes include multiple types, one or more types can be selected in specific implementation, and the filtering sequence commonly used in selecting multiple types is confidence filtering, sentence pattern filtering, mutual exclusion type filtering and priority filtering.

Further, since the label identified by the error in the model may have an error, and the label output by the model may carry a confidence coefficient corresponding to each label, the confidence coefficient of each label may be screened, and the rule label, the classification label and the semantic label are specifically filtered based on the label attribute information, and the target label corresponding to the text to be processed is determined according to the filtering result, including: determining label confidence degrees respectively corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information; determining a label confidence coefficient threshold value, and comparing the label confidence coefficient threshold value with label confidence coefficients respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the comparison result.

The label confidence level can be understood as the confidence level carried by each label when the label is output, the label confidence level is used for representing the credibility of the model for outputting the label, and the higher the label confidence level is, the more trust the model is to the output result. And the label with low confidence coefficient can be filtered by setting the label confidence coefficient threshold value, so that the label identification accuracy is improved.

In practical application, the label attribute information comprises label confidence corresponding to the label, and the label confidence is compared with a label confidence threshold, so that labels with low confidence thresholds can be filtered out of all labels, and the rest labels are used as target labels corresponding to the text to be processed.

In sum, the labels are filtered by setting the label confidence threshold, and labels with confidence less than the confidence threshold are filtered, so that the accuracy of label identification is further improved.

Further, since sentence pattern judgment is performed in the rule matching model, but sentence pattern judgment is not performed in the classification matching model and the semantic matching model, sentence pattern errors may exist in the classification label and the semantic label, and at this time, sentence pattern judgment of the label may be performed, so that labels with inconsistent sentence patterns are filtered, specifically, filtering processing is performed on the rule label, the classification label and the semantic label based on the label attribute information, and a target label corresponding to the text to be processed is determined according to a filtering processing result, including: determining label sentence pattern information corresponding to the rule labels, the classification labels and the semantic labels respectively according to the label attribute information; determining target sentence pattern information corresponding to the text to be processed, and judging based on the target sentence pattern information and label sentence pattern information respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the judging result.

The tag sentence pattern information may be understood as sentence pattern information adapted to a tag, for example, an a tag is adapted to a statement sentence, a positive sentence, etc., and if the tag of the text to be processed is determined to be an a tag, the a tag may be filtered through sentence pattern judgment. The target sentence pattern information corresponding to the text to be processed can be understood as the sentence pattern corresponding to the text to be processed, for example, the target sentence pattern information of 'I feel that the object is better to be used' is a positive sentence and a statement sentence; the target sentence pattern information of the express delivery when to deliver is an question sentence. And comparing the target sentence pattern information corresponding to the text to be processed with the label sentence pattern information corresponding to each label, and eliminating labels which do not accord with the target sentence pattern information.

In practical application, only the rule matching model determines the tag through sentence pattern judgment when determining the tag, but sentence pattern judgment is not performed in the other two classification matching models and the semantic matching model, so that sentence pattern judgment is performed again after determining the rule tag, the classification tag and the semantic tag in order to avoid the occurrence of sentence patterns of the classification tag and the semantic tag which are not in accordance with the text to be processed, and sentence pattern information which is not in accordance with the text to be processed is filtered. For example, the label 'no powder blocking' in the cosmetic industry should be adapted to the negative sentence pattern, and the label 'size consultation' in the clothing industry should be adapted to the question sentence pattern.

In a specific embodiment of the present disclosure, determining target sentence pattern information corresponding to a text to be processed as question sentences, determining tag attribute information corresponding to each tag of rule tags, classification tags and semantic tags according to tag attribute information, removing tags unsuitable for the question sentences, and determining target tags corresponding to the text to be processed according to a removing result.

In sum, the label which does not accord with the target sentence pattern information can be filtered through comparing and judging the label sentence pattern information corresponding to the label and the target sentence pattern information corresponding to the text to be processed, so that the label which accords with the text to be processed better is selected, and the label identification accuracy is further improved.

Further, the rule labels, the classification labels and the semantic labels may have labels with opposite meanings, but a text may not have labels with opposite meanings, so that filtering needs to be performed on the labels, specifically, filtering processing is performed on the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining the target labels corresponding to the text to be processed according to the filtering processing result includes: determining label category information corresponding to the rule labels, the classification labels and the semantic labels respectively according to the label attribute information; determining a mutual exclusion tag and a sharing tag according to tag category information corresponding to the rule tag, the classification tag and the semantic tag respectively; determining the mutual exclusion label confidence of the mutual exclusion label, and selecting a target mutual exclusion label from the mutual exclusion labels based on the mutual exclusion label confidence; and taking the target mutual exclusion tag and the shared tag as target tags corresponding to the text to be processed.

The label category information can be understood as category information of labels, each label belongs to one category, opposite, namely mutual exclusion labels, possibly exist among different categories, and labels without opposite relation are shared labels. If the label A and the label B are labels in the same industry, but the label A belongs to a good evaluation category and the label B belongs to a bad evaluation category, the label A and the label B are mutually exclusive labels, the shared label can be understood as labels without mutually exclusive attributes, the mutually exclusive labels are labels with opposite meanings and describing the same things, and if the labels aiming at the problem of delivery comprise two categories of quick delivery and slow delivery, the labels respectively belonging to the two categories are mutually exclusive labels.

In practical application, according to the label category information corresponding to the labels, it can be determined which labels are mutually exclusive labels and which labels are shared labels in all the labels. After the mutual exclusion label and the shared label are determined, a certain label in the mutual exclusion label needs to be removed, only one label can be reserved because the meanings of the two labels in the mutual exclusion label are opposite, and the selection of which label to reserve is determined according to the confidence of the label.

In a specific embodiment of the present disclosure, tag category information corresponding to a rule tag, a classification tag, and a semantic tag is determined according to tag attribute information, a mutual exclusion tag and a sharing tag are determined according to each tag category information, the mutual exclusion tag includes "service satisfaction" and "service dissatisfaction", tag confidence degrees of the two are compared, a tag with higher retention confidence degree is selected, and the retained target mutual exclusion tag and other sharing tags are used as target tags corresponding to a text to be processed.

In sum, the mutual exclusion labels and the sharing labels can be determined in all the labels through the label category information, and the labels with lower confidence in the mutual exclusion labels are removed, so that the mutual exclusion labels are not existed in all the labels, and the label identification result is more accurate.

Performing priority filtering processing on the rule tag, the classification tag and the semantic tag according to tag priority information in the tag attribute information, and determining a target tag corresponding to the text to be processed according to a filtering processing result, wherein the method comprises the following steps:

further, for the fact that the details of possible descriptions of the labels describing the same dimension are different, in order to reduce repeated descriptions and description of labels not specific, filtering may be performed according to label priorities, including label priority information corresponding to the rule labels, the classification labels and the semantic labels respectively according to the label attribute information; sorting according to label priority information respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the sequencing result.

The tag priority information is understood as the accurate and detailed level information of tag description. Among tags describing the same dimension, the tags of different priorities have different detailed descriptions, such as "refund or return (uncomfortable to wear)" > "refund or return (unwanted)" > "refund or return (undefined reason)", and because "refund or return (uncomfortable to wear)", giving more information, are more interesting contents to customers, the tags of higher priority are transmitted to customers. Therefore, the labels with higher priority can be reserved, and the labels with low priority can be filtered.

In practical application, the label priority information of each label can be determined according to the label attribute information, the labels in the same dimension are subjected to priority sorting, and the label with higher priority is determined according to the sorting result to be used as the target label corresponding to the text to be processed.

In a specific embodiment of the present disclosure, tag priority information corresponding to each tag is determined according to tag attribute information, and the tags in each dimension are ordered according to the tag priority information. A certain sorting result is "good taste (good taste, sweet and sour taste) > good taste (good taste)" in the taste dimension, and a label of "good taste (good taste, sweet and sour taste)" is selectively reserved according to the sorting result.

In summary, the label with more detailed description can be selected and reserved through the label priority, so that the customer can know the essence of things in more detail, and the identification accuracy of the label can be further improved through identifying and filtering the label through the label priority.

Furthermore, in order to enable the model to recognize multiple labels, an initial label may be created in advance according to the original sound text of the user, and a label library may be constructed, where the method further includes: acquiring a user acoustic text set, and determining key texts corresponding to the user acoustic text in the user acoustic text set; determining a target user acoustic text in the user acoustic text set based on the key text corresponding to the user acoustic text; and generating a corresponding initial label according to the original sound text of the target user.

The method comprises the steps that a user acoustic text set comprises a large number of user acoustic texts, the user acoustic texts are identified, key texts in the user acoustic text set are extracted through an algorithm, the key texts can be understood as key words in a sentence, the user acoustic texts with higher key text occurrence frequency are selected to be used as target user acoustic texts, the target user acoustic texts can be understood as typical user acoustic, and business labels are summarized aiming at the typical user acoustic and are used as initial labels. The initial label determined by the method can cover most of original sound, and most of demands of users are met.

In practical application, the initial label is divided into three layers of results, such as a first layer is a transaction, a second layer is a price consultation, a coupon, a goods return guarantee, a goods return freight risk, a payment certificate and a payment red packet, and a dimension is specifically defined through the multi-layer label, so that the label can be more refined to embody the user original sound text, the coverage of the label is wider, and the label can be applied to various different project environments.

Further, when the project side wants to customize the tag, the project side may customize the tag according to its own requirement and add the tag to a tag library of the initial tag, and specifically after determining the target user acoustic text in the user acoustic text set based on the key text corresponding to the user acoustic text, the method further includes: generating a corresponding first initial tag according to the original sound text of the target user; and receiving a custom tag, and determining an initial tag according to the custom tag and the first initial tag.

After a first initial tag is summarized according to original sound text of a target user, customizing a custom tag according to the self-requirements of an item party, and forming the custom tag and the first initial tag into an initial tag together, so that an initial tag library is generated. In practical application, example sentences are required to be set for the custom tags, so that the custom tags can be identified based on the semantic matching model later.

Based on the method, the user-defined label is supported, so that the label can be customized according to the requirements of the project side, and the label can conform to more use scenes.

The text marking method comprises the steps of receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model; obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model; determining tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to filtering results. Based on the rule matching model, the classification matching model and the semantic matching model, the text to be processed is identified from multiple dimensions, rule labels of the rule dimensions, classification labels of the classification dimensions and semantic labels of the semantic dimensions are identified, filtering processing is carried out on the labels of different dimensions, identification accuracy is further improved, and more accurate identification results are obtained so that target labels corresponding to the text to be processed are determined.

Fig. 3 shows a flowchart of a policy adjustment method for applying a text marking method according to an embodiment of the present disclosure, including steps 302 to 306.

Step 302: and determining a target object, a target operation strategy corresponding to the target object and a target text to be processed corresponding to the target object.

The target object may be understood as an object in the project service provided by the project party, for example, in the online shopping project, and the target object may be a commodity or a service. As in the travel industry, the target object may be a sight spot environment or the like. The target operation policy corresponding to the target object can be understood as a policy of operating the target object by the item party, and when the target object is a commodity, the target operation policy can be a policy of selling price, selling time, selling condition and the like of the commodity by taking the online shopping item as an example. The target text to be processed corresponding to the target object can be understood as user original sound text related to the target object, and the target text to be processed can be evaluation of the target object by a user, suggestion text, such as use experience evaluation, after-sales service evaluation and the like aiming at the target object.

In practical application, after the project side operates the target object for a period of time, the evaluation of the user on the operation condition of the target object is acquired, so that the project side can know a opinion of the public on the current target object, the project side can timely adjust the current operation strategy based on the opinion, the target object is more in line with the user's expectations, and the user experience is improved.

Step 304: and performing text marking processing of a text marking method on the target text to be processed to obtain a problem label corresponding to the target text to be processed.

The text marking process can be understood as marking a target text to be processed of a target object, and determining evaluation of an item party on the target object, namely, a problem label. A question label may be understood as a user's evaluation, suggestion for a target object, including positive or negative evaluation. So that a deficiency of the current target object or a place where improvement is required can be determined for the problem tag.

In practical application, taking a commodity sold by a target object as an item party as an example, if it is determined that the problem label includes labels such as "price is too high", "quality is good", "express delivery is too slow", the item party can adjust the selling price of the target object based on the problem label and adjust in terms of logistics, so as to meet the expectations of users.

Step 306: and determining a target problem of the target object based on the problem label, and adjusting the target operation strategy based on the target problem.

The target problem can be understood as a problem that the project side needs to be adjusted and modified according to the problem label, and the target operation strategy is modified according to the target problem, so that the target object can be more in line with the user's expectations, the negative evaluation of the user on the target object is reduced, and the use experience of the user is improved.

In practical applications, the target problem may be to adjust the target object itself, or may need to adjust the operation policy of the target object, if the target problem is "pricing too high", the pricing in the operation policy may be adjusted, if the target problem is "inaccurate", the size of the target object may be adjusted.

The strategy adjustment method comprises the steps of determining a target object, a target operation strategy corresponding to the target object and a target text to be processed corresponding to the target object; performing text marking processing of a text marking method on the target text to be processed to obtain a problem label corresponding to the target text to be processed; and determining a target problem of the target object based on the problem label, and adjusting the target operation strategy based on the target problem. Through marking the target text to be processed of the target object, the project side can quickly acquire the problem label of the target object, so that the evaluation of the user on the target object can be accurately acquired, the target object can be subjected to strategy adjustment based on the target problem, the target object is more in line with the user expectation, and the user experience is improved.

The text marking method provided in the present specification is further described below with reference to fig. 4 by taking an application of the text marking method to online shopping items as an example. Fig. 4 is a flowchart of a process of a text marking method according to an embodiment of the present disclosure, and specific steps include steps 402 to 410.

Step 402: receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model.

In one implementation manner, the text to be processed is an evaluation text of a certain commodity, and the evaluation text is respectively input into a rule matching model, a classification matching model and a semantic matching model.

In one implementation manner, a user acoustic text set including user acoustic texts can be obtained first, key texts of each user acoustic text are extracted through an algorithm, user acoustic texts with higher occurrence frequency of the key texts are selected to be typical user acoustic texts, a first initial tag is summarized based on typical user acoustic texts, the first initial tag and a custom tag are used as initial tags, an initial tag library is generated, and a subsequent model can select the initial tag from the initial tag library as a tag corresponding to a text to be processed.

Step 404: and matching the text to be processed with the initial label based on label matching information corresponding to the initial label through a rule matching model, determining rule labels corresponding to rule dimensions according to matching results, and outputting the rule labels.

In one implementation, the rule matching model performs regular matching on the text to be processed and the initial tag according to tag matching information corresponding to the initial tag, and determines a rule tag under a rule dimension according to a matching result.

Step 406: and encoding the text to be processed through the classification matching model to obtain an encoding vector corresponding to the text to be processed, determining a classification label corresponding to the classification dimension based on the mapping relation of the encoding vector and the vector label, and outputting the classification label.

In one implementation, the classification matching model encodes the text to be processed to obtain a high-dimensional encoding vector corresponding to the text to be processed, and determines a classification label in a classification dimension according to the encoding vector and a vector label mapping key.

Step 408: semantic information of a text to be processed is obtained through semantic matching model analysis, matching is conducted on the basis of the semantic information and reference semantic information corresponding to the initial tag, and semantic tags corresponding to semantic dimensions are determined according to matching results.

In one implementation, the semantic matching model analyzes and extracts semantic information of the text to be processed, matches the semantic information with reference semantic information of each initial tag example sentence, and determines semantic tags in semantic dimensions according to matching results.

Step 410: and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to filtering results.

In one implementation, the label confidence of each label in each dimension is determined according to the label attribute information, the label confidence is compared with a label confidence threshold, labels with label confidence less than the label confidence threshold are removed, and the target label corresponding to the text to be processed is determined according to the removal result.

In another implementation manner, the label sentence pattern information of each label in each dimension is determined according to the label attribute information, sentence pattern judgment is carried out according to the label sentence pattern information of each label and the target sentence pattern information corresponding to the text to be processed, labels with inconsistent sentence pattern information are removed according to the judgment result, and the target label corresponding to the text to be processed is determined according to the removal result.

In another implementation manner, the label category information of each label in each dimension is determined according to the label attribute information, the mutual exclusion labels and the sharing labels are determined in all labels according to the label category information of each label, the labels with low label confidence are removed from the mutual exclusion labels, and the reserved target mutual exclusion labels and the target sharing labels are used as target labels corresponding to the text to be processed.

In another implementation manner, the label priority information of each label in each dimension is determined according to the label attribute information, the label priority information of each label in the same dimension is subjected to priority sorting, labels with higher retention priority are selected, and the target label corresponding to the text to be processed is determined according to the sorting result.

The text marking method provided by the specification comprises the steps of receiving a text to be processed, and respectively inputting the text to be processed into a rule matching model, a classification matching model and a semantic matching model; obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model; determining tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the text to be processed according to filtering results. Based on the rule matching model, the classification matching model and the semantic matching model, the text to be processed is identified from multiple dimensions, rule labels of the rule dimensions, classification labels of the classification dimensions and semantic labels of the semantic dimensions are identified, filtering processing is carried out on the labels of different dimensions, identification accuracy is further improved, and more accurate identification results are obtained so that target labels corresponding to the text to be processed are determined.

The text marking method provided in the present specification is further described below with reference to fig. 5 by taking application of the text marking method in an exhibition scene as an example. Fig. 5 is a flowchart illustrating a processing procedure of another text marking method according to an embodiment of the present disclosure, and specific steps include steps 502 to 508.

Step 502: and determining a user evaluation text of the target object, and respectively inputting the user evaluation text into the rule matching model, the classification matching model and the semantic matching model.

The target object may be understood as a target exhibition object in an exhibition scene, for example, a user participates in an exhibition activity of a museum, and the target object is the museum. The user evaluation text is the evaluation of the user on the target object, such as the evaluation of the user on the exhibition activity of the museum today, which has a history value, or the evaluation of the related museum exhibition activity, which is very good in the exhibition of the museum, and the like, so that the meaning expressed by the user can be aggregated in detail according to the evaluation text of the user, and the user evaluation text can be marked based on the text marking method provided by the specification.

In practical applications, the number of users participating in the exhibition is large, and the number of received user evaluation texts is also large. In the face of massive user evaluation texts, an active party cannot accurately acquire user feedback, so that corresponding adjustment cannot be timely made according to the problem of user feedback, and poor use experience can be brought to a user in the follow-up process. Therefore, marking processing can be carried out on the user evaluation texts, the labels of each user evaluation text are obtained, the central thought expressed by the user can be accurately known according to the labels, the user evaluation texts can be subsequently classified according to the labels, the problems existing in the activities are determined, corresponding adjustment is timely carried out, and better activity experience is brought to the user.

In one implementation, a user evaluation text of a user for a museum exhibition activity is determined, text marking is performed on the user evaluation text, and the user evaluation text is respectively input into a rule matching model, a classification matching model and a semantic matching model.

Step 504: obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model.

In one implementation manner, a rule matching model is input into a user evaluation text, the rule matching model performs regular matching with the user evaluation text according to attributes such as keywords, exclusion words, adaptation sentence patterns and the like corresponding to each tag, and the rule tag corresponding to the user evaluation text is determined according to a matching result. Inputting the user evaluation text into a classification matching model, encoding the user evaluation text by the classification matching model to obtain a high-dimensional sentence vector corresponding to the user evaluation text, mapping the high-dimensional sentence vector onto a classification label, and determining the classification label corresponding to the user evaluation text. And inputting the user evaluation text into a semantic matching model, determining text features corresponding to the user evaluation text by the semantic matching model, and matching the text features of example sentences of each label, thereby determining semantic labels corresponding to the user evaluation text.

Step 506: and determining label attribute information corresponding to the rule labels, the classification labels and the semantic labels respectively.

The rule labels, the classification labels and the semantic labels are labels identified by the rule matching model, the classification matching model and the semantic matching model respectively, a plurality of labels are arranged in each type of labels, and each label has label attribute information corresponding to each label. The tag attribute information includes tag sentence pattern information corresponding to the tag, such as a positive sentence, a question sentence, etc., and priority information corresponding to the tag, etc. The labels output by all the models can be further screened according to the label attribute information, so that the labels more in line with the user evaluation text are determined, and the accuracy of label identification is improved.

In practical application, the rule matching model outputs a plurality of rule labels, the classification matching model outputs a plurality of classification labels, the semantic matching model outputs a plurality of semantic labels, all labels have label attribute information corresponding to the labels, and the labels can be further screened according to the label attribute information, so that the labels more in line with user evaluation texts are selected.

In one implementation manner, tag sentence pattern information and tag priority information corresponding to rule tags are determined, tag sentence pattern information and tag priority information corresponding to tags are classified, corresponding tag sentence pattern information and tag priority information corresponding to semantic tags are obtained, tag screening is performed based on tag attribute information corresponding to each type of tags, and tags which are more in line with the meaning of the user evaluation text to be expressed are determined.

Step 508: and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the user evaluation text according to filtering results.

In one implementation manner, according to label attribute information corresponding to each type of labels, label rules are carried out according to a preset rule strategy, labels which do not accord with user evaluation text sentence pattern information are removed, labels with low label priority are removed, and target labels corresponding to user evaluation texts are determined according to filtering results.

The text marking method provided by the specification comprises the following steps: determining a user evaluation text of a target object, and respectively inputting the user evaluation text into a rule matching model, a classification matching model and a semantic matching model; obtaining rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model, and semantic labels of corresponding semantic dimensions output by the semantic matching model; determining tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and filtering the rule labels, the classification labels and the semantic labels based on the label attribute information, and determining target labels corresponding to the user evaluation text according to filtering results. The user evaluation text is identified through multiple identification dimensions, so that the labels of the user evaluation text in multiple dimensions can be determined, and the labels in multiple dimensions can be filtered based on label attribute information to obtain labels which are more in line with the content expressed by the user evaluation text. Based on the target label corresponding to the user evaluation text, the active party can accurately know the content which the user wants to express, and can make an activity policy adjustment based on the target label in time, so that better experience is provided for the user.

Corresponding to the above method embodiment, the present disclosure further provides an embodiment of a text marking device, and fig. 6 shows a schematic structural diagram of a text marking device according to an embodiment of the present disclosure. As shown in fig. 6, the apparatus includes:

an input module 602 configured to receive a text to be processed, and input the text to be processed into a rule matching model, a classification matching model and a semantic matching model, respectively;

an obtaining module 604 configured to obtain a rule tag of the rule matching model output corresponding rule dimension, a classification tag of the classification matching model output corresponding classification dimension, and a semantic tag of the semantic matching model output corresponding semantic dimension;

a determining module 606 configured to determine tag attribute information corresponding to the rule tag, the classification tag, and the semantic tag, respectively;

and the filtering module 608 is configured to perform filtering processing on the rule tag, the classification tag and the semantic tag based on the tag attribute information, and determine a target tag corresponding to the text to be processed according to a filtering processing result.

Optionally, the obtaining module 604 is further configured to: matching the text to be processed with the initial tag based on tag matching information corresponding to the initial tag through the rule matching model; and determining rule labels corresponding to the rule dimensions according to the matching result and outputting the rule labels.

Optionally, the obtaining module 604 is further configured to: coding the text to be processed through the classification matching model to obtain a coding vector corresponding to the text to be processed; and determining and outputting the classification label corresponding to the classification dimension based on the coding vector and the vector label mapping relation.

Optionally, the obtaining module 604 is further configured to: obtaining semantic information of the text to be processed through the semantic matching model analysis; and matching the semantic information with the reference semantic information corresponding to the initial label, and determining the semantic label corresponding to the semantic dimension according to a matching result.

Optionally, the filtering module 608 is further configured to: determining label confidence degrees respectively corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information; determining a label confidence coefficient threshold value, and comparing the label confidence coefficient threshold value with label confidence coefficients respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the comparison result.

Optionally, the filtering module 608 is further configured to: determining label sentence pattern information corresponding to the rule labels, the classification labels and the semantic labels respectively according to the label attribute information; determining target sentence pattern information corresponding to the text to be processed, and judging based on the target sentence pattern information and label sentence pattern information respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the judging result.

Optionally, the filtering module 608 is further configured to: determining label category information corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information, and determining mutual exclusion labels and sharing labels according to the label category information corresponding to the rule labels, the classification labels and the semantic labels; determining the mutual exclusion label confidence of the mutual exclusion label, and selecting a target mutual exclusion label from the mutual exclusion labels based on the mutual exclusion label confidence; and taking the target mutual exclusion tag and the shared tag as target tags corresponding to the text to be processed.

Optionally, the filtering module 608 is further configured to: tag priority information corresponding to the rule tag, the classification tag and the semantic tag respectively according to the tag attribute information; sorting according to label priority information respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the sequencing result.

Optionally, the apparatus further comprises a generating module configured to: acquiring a user acoustic text set, and determining key texts corresponding to the user acoustic text in the user acoustic text set; determining a target user acoustic text in the user acoustic text set based on the key text corresponding to the user acoustic text; and generating a corresponding initial label according to the original sound text of the target user.

Optionally, the generating module is further configured to: generating a corresponding first initial tag according to the original sound text of the target user; and receiving a custom tag, and determining an initial tag according to the custom tag and the first initial tag.

The text marking device comprises an input module, a rule matching model, a classification matching model and a semantic matching model, wherein the input module is configured to receive a text to be processed and input the text to be processed into the rule matching model, the classification matching model and the semantic matching model respectively; the obtaining module is configured to obtain rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model and semantic labels of corresponding semantic dimensions output by the semantic matching model; the determining module is configured to determine tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and the filtering module is configured to filter the rule labels, the classification labels and the semantic labels based on the label attribute information, and determine a target label corresponding to the text to be processed according to a filtering result. Based on the rule matching model, the classification matching model and the semantic matching model, the text to be processed is identified from multiple dimensions, rule labels of the rule dimensions, classification labels of the classification dimensions and semantic labels of the semantic dimensions are identified, filtering processing is carried out on the labels of different dimensions, identification accuracy is further improved, and more accurate identification results are obtained so that target labels corresponding to the text to be processed are determined.

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

Corresponding to the method embodiment, the present disclosure further provides an embodiment of a policy adjustment device, and fig. 7 shows a schematic structural diagram of a policy adjustment device according to an embodiment of the present disclosure. As shown in fig. 7, the apparatus includes:

a determining module 702, configured to determine a target object, a target operation policy corresponding to the target object, and a target text to be processed corresponding to the target object;

the marking module 704 is configured to perform text marking processing of a text marking method on the target text to be processed to obtain a problem label corresponding to the target text to be processed;

an adjustment module 706 configured to determine a target issue for the target object based on the issue tag and adjust the target operation policy based on the target issue.

The policy adjustment device provided by the specification comprises a determination module, a processing module and a processing module, wherein the determination module is configured to determine a target object, a target operation policy corresponding to the target object and a target text to be processed corresponding to the target object; the marking module is configured to perform text marking processing of a text marking method on the target text to be processed to obtain a problem label corresponding to the target text to be processed; and the adjusting module is configured to determine a target problem of the target object based on the problem label and adjust the target operation strategy based on the target problem. Through marking the target text to be processed of the target object, the project side can quickly acquire the problem label of the target object, so that the evaluation of the user on the target object can be accurately acquired, the target object can be subjected to strategy adjustment based on the target problem, the target object is more in line with the user expectation, and the user experience is improved.

The above is a schematic scheme of a policy adjustment device of the present embodiment. It should be noted that, the technical solution of the policy adjustment device and the technical solution of the policy adjustment method belong to the same concept, and details of the technical solution of the policy adjustment device, which are not described in detail, can be referred to the description of the technical solution of the text marking method.

Corresponding to the method embodiment, the present disclosure further provides an embodiment of a text marking device, and fig. 8 shows a schematic structural diagram of a text marking device according to an embodiment of the present disclosure. As shown in fig. 8, the apparatus includes:

an input module 802 configured to determine a user evaluation text of a target object and input the user evaluation text to a rule matching model, a classification matching model, and a semantic matching model, respectively;

an obtaining module 804 configured to obtain a rule tag of the rule matching model output corresponding rule dimension, a classification tag of the classification matching model output corresponding classification dimension, and a semantic tag of the semantic matching model output corresponding semantic dimension;

a determining module 806 configured to determine tag attribute information corresponding to the rule tag, the classification tag, and the semantic tag, respectively;

And the filtering module 808 is configured to perform filtering processing on the rule tag, the classification tag and the semantic tag based on the tag attribute information, and determine a target tag corresponding to the user evaluation text according to a filtering processing result.

Optionally, the obtaining module 804 is further configured to: matching the text to be processed with the initial tag based on tag matching information corresponding to the initial tag through the rule matching model; and determining rule labels corresponding to the rule dimensions according to the matching result and outputting the rule labels.

Optionally, the obtaining module 804 is further configured to: coding the text to be processed through the classification matching model to obtain a coding vector corresponding to the text to be processed; and determining and outputting the classification label corresponding to the classification dimension based on the coding vector and the vector label mapping relation.

Optionally, the obtaining module 804 is further configured to: obtaining semantic information of the text to be processed through the semantic matching model analysis; and matching the semantic information with the reference semantic information corresponding to the initial label, and determining the semantic label corresponding to the semantic dimension according to a matching result.

Optionally, the filtering module 808 is further configured to: determining label confidence degrees respectively corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information; determining a label confidence coefficient threshold value, and comparing the label confidence coefficient threshold value with label confidence coefficients respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the comparison result.

Optionally, the filtering module 808 is further configured to: determining label sentence pattern information corresponding to the rule labels, the classification labels and the semantic labels respectively according to the label attribute information; determining target sentence pattern information corresponding to the text to be processed, and judging based on the target sentence pattern information and label sentence pattern information respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the judging result.

Optionally, the filtering module 808 is further configured to: determining label category information corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information, and determining mutual exclusion labels and sharing labels according to the label category information corresponding to the rule labels, the classification labels and the semantic labels; determining the mutual exclusion label confidence of the mutual exclusion label, and selecting a target mutual exclusion label from the mutual exclusion labels based on the mutual exclusion label confidence; and taking the target mutual exclusion tag and the shared tag as target tags corresponding to the text to be processed.

Optionally, the filtering module 808 is further configured to: tag priority information corresponding to the rule tag, the classification tag and the semantic tag respectively according to the tag attribute information; sorting according to label priority information respectively corresponding to the rule labels, the classification labels and the semantic labels; and determining the target label corresponding to the text to be processed according to the sequencing result.

The text marking device comprises an input module, a rule matching model, a classification matching model and a semantic matching model, wherein the input module is configured to determine a user evaluation text of a target object, and input the user evaluation text into the rule matching model, the classification matching model and the semantic matching model respectively; the obtaining module is configured to obtain rule labels of corresponding rule dimensions output by the rule matching model, classification labels of corresponding classification dimensions output by the classification matching model and semantic labels of corresponding semantic dimensions output by the semantic matching model; the determining module is configured to determine tag attribute information corresponding to the rule tag, the classification tag and the semantic tag respectively; and the filtering module is configured to filter the rule labels, the classification labels and the semantic labels based on the label attribute information, and determine a target label corresponding to the user evaluation text according to a filtering result. The user evaluation text is identified through multiple identification dimensions, so that the labels of the user evaluation text in multiple dimensions can be determined, and the labels in multiple dimensions can be filtered based on label attribute information to obtain labels which are more in line with the content expressed by the user evaluation text. Based on the target label corresponding to the user evaluation text, the active party can accurately know the content which the user wants to express, and can make an activity policy adjustment based on the target label in time, so that better experience is provided for the user.

Fig. 9 illustrates a block diagram of a computing device 900 provided in accordance with an embodiment of the present specification. The components of computing device 900 include, but are not limited to, memory 910 and processor 920. Processor 920 is coupled to memory 910 via bus 930 with database 950 configured to hold data.

Computing device 900 also includes an access device 940, access device 940 enabling computing device 900 to communicate via one or more networks 960. Examples of such networks include the Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the internet. Access device 940 may include one or more of any type of network interface, wired or wireless (e.g., a Network Interface Card (NIC)), such as an IEEE802.11 Wireless Local Area Network (WLAN) wireless interface, a worldwide interoperability for microwave access (Wi-MAX) interface, an ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a bluetooth interface, a Near Field Communication (NFC) interface, and so forth.

In one embodiment of the present description, the above-described components of computing device 900 and other components not shown in FIG. 9 may also be connected to each other, for example, by a bus. It should be understood that the block diagram of the computing device illustrated in FIG. 9 is for exemplary purposes only and is not intended to limit the scope of the present description. Those skilled in the art may add or replace other components as desired.

Computing device 900 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 PC. Computing device 900 may also be a mobile or stationary server.

Wherein the processor 920 performs the steps of the text marking method when executing the computer instructions.

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 marking 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 marking method.

An embodiment of the present disclosure also provides a computer readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the text marking method and the policy adjustment method as 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 solutions of the text marking method and the policy adjustment 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 solutions of the text marking method and the policy adjustment method.

An embodiment of the present disclosure further provides a computer program, where the computer program when executed in a computer causes the computer to execute the steps of the text marking method and the policy adjustment 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 solutions of the text marking method and the policy adjustment method belong to the same concept, 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 solutions of the text marking method and the policy adjustment 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, randomAccessMemory), an electrical carrier signal, a telecommunication 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 marking method, comprising:

2. The method of claim 1, obtaining a rule tag for the rule matching model output corresponding to a rule dimension, comprising:

matching the text to be processed with the initial tag based on tag matching information corresponding to the initial tag through the rule matching model;

and determining rule labels corresponding to the rule dimensions according to the matching result and outputting the rule labels.

3. The method of claim 1, obtaining a classification label for the classification matching model output corresponding classification dimensions, comprising:

Coding the text to be processed through the classification matching model to obtain a coding vector corresponding to the text to be processed;

and determining and outputting the classification label corresponding to the classification dimension based on the coding vector and the vector label mapping relation.

4. The method of claim 1, obtaining the semantic tags of the semantic matching model output corresponding semantic dimensions, comprising:

obtaining semantic information of the text to be processed through the semantic matching model analysis;

and matching the semantic information with the reference semantic information corresponding to the initial label, and determining the semantic label corresponding to the semantic dimension according to a matching result.

5. The method of claim 1, wherein filtering the rule tag, the classification tag and the semantic tag based on the tag attribute information, and determining the target tag corresponding to the text to be processed according to a filtering result comprises:

determining label confidence degrees respectively corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information;

determining a label confidence coefficient threshold value, and comparing the label confidence coefficient threshold value with label confidence coefficients respectively corresponding to the rule labels, the classification labels and the semantic labels;

And determining the target label corresponding to the text to be processed according to the comparison result.

6. The method of claim 1, wherein filtering the rule tag, the classification tag and the semantic tag based on the tag attribute information, and determining the target tag corresponding to the text to be processed according to a filtering result comprises:

determining label sentence pattern information corresponding to the rule labels, the classification labels and the semantic labels respectively according to the label attribute information;

determining target sentence pattern information corresponding to the text to be processed, and judging based on the target sentence pattern information and label sentence pattern information respectively corresponding to the rule labels, the classification labels and the semantic labels;

and determining the target label corresponding to the text to be processed according to the judging result.

7. The method of claim 1, wherein filtering the rule tag, the classification tag and the semantic tag based on the tag attribute information, and determining the target tag corresponding to the text to be processed according to a filtering result comprises:

determining label category information corresponding to the rule labels, the classification labels and the semantic labels according to the label attribute information

Determining a mutual exclusion tag and a sharing tag according to tag category information corresponding to the rule tag, the classification tag and the semantic tag respectively;

determining the mutual exclusion label confidence of the mutual exclusion label, and selecting a target mutual exclusion label from the mutual exclusion labels based on the mutual exclusion label confidence;

and taking the target mutual exclusion tag and the shared tag as target tags corresponding to the text to be processed.

8. The method of claim 1, wherein the performing priority filtering processing on the rule tag, the classification tag and the semantic tag according to tag priority information in the tag attribute information, and determining the target tag corresponding to the text to be processed according to a filtering processing result includes:

tag priority information corresponding to the rule tag, the classification tag and the semantic tag respectively according to the tag attribute information;

sorting according to label priority information respectively corresponding to the rule labels, the classification labels and the semantic labels;

and determining the target label corresponding to the text to be processed according to the sequencing result.

9. The method of claim 1, the method further comprising:

acquiring a user acoustic text set, and determining key texts corresponding to the user acoustic text in the user acoustic text set;

Determining a target user acoustic text in the user acoustic text set based on the key text corresponding to the user acoustic text;

and generating a corresponding initial label according to the original sound text of the target user.

10. The method of claim 9, after determining target user acoustic text in the set of user acoustic text based on key text corresponding to the user acoustic text, the method further comprising:

generating a corresponding first initial tag according to the original sound text of the target user;

and receiving a custom tag, and determining an initial tag according to the custom tag and the first initial tag.

11. A policy adjustment method, comprising:

performing text marking processing of the text marking method according to any one of claims 1-10 on the target text to be processed to obtain a question label corresponding to the target text to be processed;

12. A text marking method, comprising:

13. A computing device comprising a memory, a processor, and computer instructions stored on the memory and executable on the processor, when executing the computer instructions, implementing the steps of the method of any one of claims 1-12.

14. A computer readable storage medium storing computer executable instructions which when executed by a processor perform the steps of the method of any one of claims 1 to 12.