Skill tag extraction, candidate phrase classification model training method and device

Abstract

The present disclosure provides a skill label extraction method and device of a candidate phrase classification model, and relates to the technical field of data processing, in particular to the technical field of artificial intelligence and deep learning. The specific implementation scheme is as follows: performing word segmentation processing on a to-be-extracted sentence to obtain a plurality of segmented words, and determining a multi-dimensional feature vector of each segmented word; extracting a candidate phrase from the to-be-extracted sentence; determining the multi-dimensional feature vector of each segmented word in the candidate phrase according to the multi-dimensional feature vector of each segmented word; generating a semantic representation vector of the candidate phrase according to the multi-dimensional feature vector of each segmented word in the candidate phrase; and extracting a skill label from the to-be-extracted sentence based on the semantic representation vector of the candidate phrase. The present scheme can effectively improve the accuracy of skill label extraction and provide necessary conditions for enterprises to realize intelligent talent management.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data processing, in particular to the technical field of artificial intelligence and deep learning, and specifically to skill label extraction, a training method and device of a candidate phrase classification model. BACKGROUND

[0002] There are massive multi-source heterogeneous data in enterprise recruitment and talent management scenarios, including resumes, job descriptions, talent evaluation data, etc. Talent insight can usually be achieved through labeling, which not only saves management costs for enterprises, but also is an important method for enterprises to realize intelligent talent management.

[0003] In related technologies, skill labels can be extracted from these multi-source heterogeneous data in a supervised or unsupervised manner, but the accuracy of extraction needs to be improved. SUMMARY

[0004] The present disclosure provides a skill label extraction method and a training method and device of a candidate phrase classification model.

[0005] According to an aspect of the present disclosure, a skill label extraction method is provided, comprising:

[0006] performing word segmentation processing on a to-be-extracted sentence to obtain a plurality of segmented words, and determining a multi-dimensional feature vector of each segmented word;

[0007] extracting a candidate phrase from the to-be-extracted sentence;

[0008] determining a multi-dimensional feature vector of each segmented word in the candidate phrase according to the multi-dimensional feature vector of each segmented word;

[0009] generating a semantic representation vector of the candidate phrase according to the multi-dimensional feature vector of each segmented word in the candidate phrase;

[0010] extracting a skill label from the to-be-extracted sentence based on the semantic representation vector of the candidate phrase.

[0011] According to a second aspect of the present disclosure, a training method of a candidate phrase classification model is provided, comprising:

[0012] obtaining a labeled training set and an unlabeled data set; wherein the labeled training set comprises first sentence samples and corresponding skill label samples thereof, and the unlabeled data set comprises second sentence samples and corresponding candidate phrase samples thereof;

[0013] training the candidate phrase classification model according to the first sentence samples and the corresponding skill label samples thereof, to obtain a trained candidate phrase classification model;

[0014] predict classification probabilities of the candidate phrase samples in the unlabeled data set based on the trained candidate phrase classification model;

[0015] update the labeled training set and the unlabeled data set based on the classification probabilities;

[0016] train the trained candidate phrase classification model based on the updated labeled training set.

[0017] According to a third aspect of the present disclosure, there is provided a skill label extraction device, comprising:

[0018] a first determination module configured to perform word segmentation on a to-be-extracted sentence to obtain a plurality of segmented words and determine a multi-dimensional feature vector of each segmented word;

[0019] a first extraction module configured to extract a candidate phrase from the to-be-extracted sentence;

[0020] a second determination module configured to determine a multi-dimensional feature vector of each segmented word in the candidate phrase based on the multi-dimensional feature vector of each segmented word;

[0021] a first generation module configured to generate a semantic representation vector of the candidate phrase based on the multi-dimensional feature vector of each segmented word in the candidate phrase;

[0022] a second extraction module configured to extract a skill label from the to-be-extracted sentence based on the semantic representation vector of the candidate phrase.

[0023] According to a fourth aspect of the present disclosure, there is provided a training device for a candidate phrase classification model, comprising:

[0024] an acquisition module configured to acquire a labeled training set and an unlabeled data set; wherein the labeled training set comprises first sentence samples and corresponding skill label samples, and the unlabeled data set comprises second sentence samples and corresponding candidate phrase samples;

[0025] a training module configured to train the candidate phrase classification model based on the first sentence samples and the corresponding skill label samples, to obtain a trained candidate phrase classification model;

[0026] a prediction module configured to predict classification probabilities of the candidate phrase samples in the unlabeled data set based on the trained candidate phrase classification model;

[0027] an update module configured to update the labeled training set and the unlabeled data set based on the classification probabilities;

[0028] The training module is further configured to train the trained candidate phrase classification model based on the updated labeled training set.

[0029] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising:

[0030] at least one processor; and

[0031] a memory connected with the at least one processor in communication; wherein,

[0032] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of the first aspect described above, and / or perform the method of the second aspect described above.

[0033] According to a sixth aspect of the present disclosure, a non-transitory computer readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to perform the method of the first aspect described above, and / or perform the method of the second aspect described above.

[0034] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method of the first aspect described above, and / or implements the method of the second aspect described above.

[0035] According to the technical solution of the present disclosure, the candidate phrases are extracted from the to-be-extracted sentence, and the semantic representation vector of the candidate phrase is generated according to the multi-dimensional feature vector of each word in the candidate phrase, and the skill label is extracted from the to-be-extracted sentence based on the semantic representation vector of the candidate phrase, thereby realizing the automatic extraction of the skill label. Wherein, the multi-dimensional feature vector of each word in the candidate phrase is used to extract the skill label, which can effectively improve the accuracy of skill label extraction, and can provide necessary conditions for intelligent talent management of enterprises.

[0036] It should be understood that the contents described in this part are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they used to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings are used to better understand the present scheme and do not limit the present disclosure. Among them:

[0038] Figure 1 is a schematic diagram according to the first embodiment of the present disclosure;

[0039] Figure 2 is a structural schematic diagram of a candidate phrase classification model in the embodiment of the present disclosure;

[0040] Figure 3is a schematic diagram according to a second embodiment of the present disclosure;

[0041] Figure 4 is a schematic diagram according to a third embodiment of the present disclosure;

[0042] Figure 5 is a structural schematic diagram of another candidate phrase classification model in an embodiment of the present disclosure;

[0043] Figure 6 is a schematic diagram according to a fourth embodiment of the present disclosure;

[0044] Figure 7 is a structural schematic diagram of yet another candidate phrase classification model in an embodiment of the present disclosure;

[0045] Figure 8 is a schematic diagram according to a fifth embodiment of the present disclosure;

[0046] Figure 9 is a schematic diagram according to a sixth embodiment of the present disclosure;

[0047] Figure 10 is a schematic diagram according to a seventh embodiment of the present disclosure;

[0048] Figure 11 is a schematic diagram according to an eighth embodiment of the present disclosure;

[0049] Figure 12 is a schematic diagram according to a ninth embodiment of the present disclosure;

[0050] Figure 13 is a schematic diagram according to a tenth embodiment of the present disclosure;

[0051] Figure 14 is a block diagram of an electronic device for implementing the method of the embodiments of the present disclosure. DETAILED DESCRIPTION

[0052] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, and should be considered as merely exemplary. Thus, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, for the sake of brevity and clarity, descriptions of well-known functions and constructions are omitted from the following description.

[0053] It should be noted that in the technical solutions of the present disclosure, the acquisition, storage and application of user personal information involved comply with relevant laws and regulations and do not violate public order and good customs. The user personal information involved is acquired, stored and applied with the consent of the user.

[0054] There are massive multi-source heterogeneous data in the enterprise recruitment and talent management scenarios, including resumes, job descriptions, talent evaluation data, etc. In order to facilitate talent management, talent insight can be achieved through tagging, which not only saves management costs for enterprises, but also is an important method for enterprises to realize intelligent talent management.

[0055] In the related art, skill tags can be extracted from these multi-source heterogeneous data in a supervised or unsupervised manner, but the accuracy of extraction needs to be improved.

[0056] Based on the above problems, the present disclosure provides a skill tag extraction method and a candidate phrase classification model training method and device.

[0057] Figure 1 A flowchart of a skill tag extraction method is provided in an embodiment of the present disclosure. It should be noted that the skill tag extraction method in the embodiment of the present disclosure can be used in the skill tag extraction device in the embodiment of the present disclosure, and the skill tag extraction device in the embodiment of the present disclosure can be configured in an electronic device. As shown in the figure, the method can include the following steps: Figure 1

[0058] Step 101, performing word segmentation processing on the to-be-extracted sentence to obtain a plurality of segmented words, and determining a multi-dimensional feature vector of each segmented word.

[0059] In some embodiments of the present disclosure, the to-be-extracted sentence can be a sentence in a resume, a sentence in a job function description, a sentence in talent evaluation data, or a sentence in other data containing skill descriptions. As an example, the way to obtain the to-be-extracted sentence can be: based on the interactive interface of the electronic device, obtaining the document selected by the user or the document uploaded by the user; based on a preset program, automatically reading the sentences in the document sentence by sentence, and each read sentence can be used as a to-be-extracted sentence. As another example, the user can input the sentence that needs to be extracted by skill tags through the interactive interface of the electronic device, so that the corresponding to-be-extracted sentence can be obtained according to the information submitted by the exchange interface.

[0060] In some embodiments of the present disclosure, the to-be-extracted sentence can be segmented by the word segmentation tool in the related art, or the to-be-extracted sentence can be segmented by the model constructed by the person skilled in the art to realize the word segmentation processing, and the present disclosure does not limit this.

[0061] ​wherein the multi-dimensional feature vector of each segmented word refers to a vector that can represent the semantic, part-of-speech, and other features of each segmented word. For example, the multi-dimensional feature vector of each segmented word can be composed based on a sequence of semantic features, part-of-speech features, and dependency syntax analysis features of the segmented word. As an example, the semantic features of each segmented word can be obtained by using a word vector model such as word2Vec; the part-of-speech features of each segmented word can be obtained by using a part-of-speech tagging model in the related art; and the dependency syntax analysis features of each segmented word can be obtained by using a dependency syntax analysis tool in the related art.

[0062] Step 102: extracting a candidate phrase from the sentence to be extracted.

[0063] In some embodiments of the present disclosure, the candidate phrase refers to a combination of segmented words that can be a skill tag.

[0064] As an implementation manner, the implementation manner of extracting the candidate phrase from the sentence to be extracted can be: according to the order of the segmented words in the segmented word processing result of the sentence to be extracted, sequentially extracting segmented word combinations, each of which is a candidate phrase, according to a preset window size. For example, if the segmented word processing result of the sentence to be extracted is {x1, x2, x3, x4, x5, x6}, and the preset window size is 2, then the extracted candidate phrases are {x1, x2}, {x2, x3}, {x3, x4}, {x4, x5}, and {x5, x6}.

[0065] As another implementation manner, since the skill tag is usually a combination of phrases of a certain part of speech, a candidate phrase template can be preset, wherein the candidate phrase template can be obtained based on a large amount of data and contains a plurality of part-of-speech combinations, that is, if a plurality of continuous segmented words in the segmented word processing result satisfy a part-of-speech combination in the candidate phrase template, then the plurality of continuous segmented words can be taken as a candidate phrase.

[0066] Step 103: determining the multi-dimensional feature vector of each segmented word in the candidate phrase according to the multi-dimensional feature vector of each segmented word.

[0067] It can be understood that since the candidate phrase can include at least one segmented word, the multi-dimensional feature vector of each segmented word in the candidate phrase can be determined according to the multi-dimensional feature vector of each segmented word.

[0068] In some embodiments of the present disclosure, the implementation process of this step can include: determining at least one target segmented word contained in the candidate phrase; and determining the multi-dimensional feature vector of each target segmented word according to the multi-dimensional feature vector of each segmented word. That is, the target segmented word contained in the candidate phrase can be determined first, and then the multi-dimensional feature vector of the target segmented word can be found from the multi-dimensional feature vector of each segmented word.

[0069] Step 104: Generate the semantic representation vector of the candidate phrase based on the multidimensional feature vectors of each word segment in the candidate phrase.

[0070] In other words, the features of the candidate phrase are characterized by the multi-dimensional feature vectors of each word segment in the candidate phrase, so as to obtain the semantic representation of the candidate phrase.

[0071] As one approach, the multidimensional feature vectors of each segment in the candidate phrase can be input into a model that can represent the semantics of the sequence, so as to extract features from the combined feature vectors and generate the semantic representation vector of the candidate phrase.

[0072] Step 105: Extract skill tags from the sentences to be extracted based on the semantic representation vectors of the candidate phrases.

[0073] In other words, based on the semantic representation vector of the candidate phrase, it is determined whether the candidate phrase is a skill tag, so as to achieve automatic extraction of skill tags.

[0074] As an example, the classification probability of a candidate phrase can be obtained based on its semantic representation vector; if the classification probability is greater than a preset threshold, the candidate phrase is identified as a skill tag. The classification probability of the candidate phrase can be obtained through classification calculation using a classifier function.

[0075] It should be noted that in some embodiments of this disclosure, after extracting skill tags from the statements to be extracted, the method may further include determining the category to which the skill tags belong. Typically, in talent management scenarios, skill tags can be categorized into soft skills, hard skills, project-related tags, etc., so after extracting the skill tags, it may be necessary to determine their category. As an example, a trained category determination model can be used to determine the category to which the extracted skill tags belong. The trained category determination model can be a BERT semantic model; for example, the skill tags can be masked and then input into the BERT semantic model for classification.

[0076] Furthermore, the skill tag extraction method in this embodiment can be implemented based on a candidate phrase classification model. Figure 2 This is a structural diagram of a candidate phrase classification model according to an embodiment of this disclosure. Figure 2 As shown, the candidate phrase classification model includes a first semantic representation layer, a classification layer, and multiple fully connected layers. The first semantic representation layer extracts features from the input vector to achieve semantic representation of the sequence. It can be a neural network such as a long short-term memory model used for sequence semantic representation. The multiple fully connected layers transform the features to transform the features from high-dimensional to low-dimensional. The classification layer performs classification calculations based on the features. For example, it can be a sigmoid binary classification function.

[0077] It should be noted that the candidate phrase classification model in this embodiment is a trained model, which can be trained based on a labeled training set or self-trained based on a labeled training set and an unlabeled dataset.

[0078] based on Figure 2 The candidate phrase classification model shown in this disclosure, and the skill tag extraction method of this embodiment, may include the following steps: performing word segmentation on the statement to be extracted to obtain multiple words, and determining the multidimensional feature vector of each word; extracting candidate phrases from the statement to be extracted; determining the multidimensional feature vector of each word in the candidate phrase based on the multidimensional feature vector of each word; inputting the multidimensional feature vector of each word in the candidate phrase into a first semantic representation layer to generate a semantic representation vector of the candidate phrase; inputting the semantic representation vector of the candidate phrase into a multi-layer fully connected layer to perform feature transformation on the semantic representation vector of the candidate phrase to obtain a transformed feature vector; inputting the transformed feature vector into a classification layer to perform classification calculation based on the transformed feature vector to obtain the classification probability of the candidate phrase; and determining the candidate phrase as a skill tag in response to the classification probability being greater than a preset threshold.

[0079] According to the skill tag extraction method of this disclosure, candidate phrases are extracted from the sentence to be extracted, and semantic representation vectors of the candidate phrases are generated based on the multidimensional feature vectors of each word segment in the candidate phrases. Skill tags are then extracted from the sentence based on the semantic representation vectors of the candidate phrases, thereby achieving automatic skill tag extraction. Utilizing the multidimensional feature vectors of each word segment in the candidate phrases for skill tag extraction can effectively improve the accuracy of skill tag extraction and provide necessary conditions for enterprises to achieve intelligent talent management.

[0080] It is understood that the quality of the extracted candidate phrases can directly affect the computational load and accuracy of skill tag extraction. Therefore, in order to improve the efficiency and accuracy of skill tag extraction, this disclosure provides another embodiment.

[0081] Figure 3 This is a flowchart illustrating the process of extracting candidate phrases from the statements to be extracted, as described in an embodiment of this disclosure. Figure 3 As shown, the implementation of the step of extracting candidate phrases from the statement to be extracted can include:

[0082] Step 301: Obtain the part-of-speech tag for each word from the multi-dimensional feature vector of each word segment.

[0083] It is understandable that, since the multidimensional feature vector of each word segment contains the part-of-speech (POS) features of each word segment, the POS features of each word segment can be obtained from the multidimensional feature vector of each word segment, and the POS tag of each word segment can be determined based on the POS features of each word segment. The POS tag is a label used to characterize a word segment as a verb, noun, etc.

[0084] As an example, the part-of-speech features of each word can be obtained from the multi-dimensional feature vector of each word segment, and the part-of-speech tags of each word can be determined according to the mapping relationship between the part-of-speech features and the part-of-speech tags.

[0085] Step 302: Based on the preset candidate phrase template and the part-of-speech tag of each word segment, extract candidate phrases from the sentence to be extracted.

[0086] Through research on skill tags, the inventors of this disclosure discovered that skill tags are typically phrase combinations with certain parts of speech, such as noun phrases. Therefore, candidate phrase templates can be pre-configured according to actual application scenarios. These templates are collections of various parts of speech combinations, such as {noun, noun 1 + noun 2, verb + noun, ...}. If the part-of-speech tags of consecutive words in the sentence to be extracted match the part-of-speech combinations in the candidate phrase template, then the corresponding phrases can be extracted as candidate phrases. In other words, by limiting the selection through candidate phrase templates, the quality of the extracted candidate phrases can be effectively improved, increasing the likelihood that the extracted candidate phrases are skill tags. This avoids resource waste caused by low-quality candidate phrases and also prevents the impact on the accuracy of skill tag extraction from the failure to extract high-quality candidate phrases.

[0087] In embodiments of this disclosure, the implementation may include: comparing the part-of-speech tags of each consecutive word segment in the sentence to be extracted with the part-of-speech combinations in the candidate phrase template; if the word segment tags contained in a certain phrase in the sentence to be extracted are consistent with a certain part-of-speech combination in the candidate phrase template, then the phrase is extracted as a candidate phrase.

[0088] According to the skill tag extraction method of this disclosure, based on the part-of-speech tag of each word segment and a preset candidate phrase model, candidate phrases are extracted from the sentence to be extracted in order to improve the quality of the extracted candidate phrases, that is, to increase the probability that the extracted candidate phrases are skill tags. This can avoid the waste of resources caused by low-quality candidate phrases, and can also avoid the impact on the accuracy of skill tag extraction caused by the failure to extract high-quality candidate phrases.

[0089] Since the semantic representation of candidate phrases is usually also related to the context of the candidate phrases, this disclosure provides yet another embodiment in order to further improve the accuracy of skill tag extraction.

[0090] Figure 4 A flowchart illustrating another method for extracting skill tags provided in this embodiment of the disclosure. Figure 4 As shown, the method includes:

[0091] Step 401: Perform word segmentation on the sentence to be extracted to obtain multiple words, and determine the multidimensional feature vector of each word.

[0092] Step 402: Extract candidate phrases from the sentences to be extracted.

[0093] Step 403: Determine the multidimensional feature vectors of each word in the candidate phrase based on the multidimensional feature vectors of each word segment.

[0094] Step 404: Generate the semantic representation vector of the candidate phrase based on the multidimensional feature vectors of each word segment in the candidate phrase.

[0095] Step 405: Determine the multidimensional feature vector of each word in the candidate phrase context based on the multidimensional feature vector of each word and the preset window size; wherein, the candidate phrase context contains the candidate phrase.

[0096] As an example, suppose the sentence to be extracted is processed into multiple segments {x1,x2,x3,x4,x5,x6,x7,x8,x9}, where the candidate phrase is {x4,x5}, and the preset window size is 2. Then the context of the candidate phrase contains the segments {x2,x3,x4,x5,x6,x7}. That is to say, the multi-dimensional feature vectors of x2,x3,x4,x5,x6,x7,x8,x7 in the context of the candidate phrase can be determined from the multi-dimensional feature vectors of x1,x2,x3,x4,x5,x6,x7,x8,x7.

[0097] Step 406: Generate the semantic representation vector of the candidate phrase context based on the multidimensional feature vectors of each word segment in the candidate phrase context.

[0098] It is understandable that the context of a candidate phrase usually has a certain influence on its semantic representation. Therefore, in order to improve the accuracy of skill tag extraction, we can combine the semantic representation vector of the candidate phrase context with the semantic representation vector of the candidate phrase to generate the semantic representation vector of the candidate phrase context.

[0099] As one approach, the multidimensional feature vectors of each word segment in the candidate phrase context can be input into a model that can represent the semantics of the sequence, so as to extract features from the combined feature vectors and generate the semantic representation vector of the candidate phrase context.

[0100] Step 407: Extract skill tags from the sentences to be extracted based on the semantic representation vectors of the candidate phrases and the semantic representation vectors of the candidate phrase contexts.

[0101] As an example, the semantic representation vector of the candidate phrase and the semantic representation vector of the candidate phrase's context can be concatenated to obtain a concatenated feature vector. Based on this concatenated feature vector, the classification probability of the candidate phrase is obtained. If the classification probability is greater than a preset threshold, the candidate phrase is identified as a skill tag. The classification probability of the candidate phrase can be obtained through a classifier function.

[0102] It should be noted that the skill tag extraction method in this embodiment can be implemented based on a candidate phrase classification model. Figure 5 This is a structural diagram of another candidate phrase classification model in an embodiment of this disclosure. Figure 5 As shown, the candidate phrase classification model can include a first semantic representation layer, a second semantic representation layer, a classification layer, and a multi-layer connection layer. The first semantic representation layer extracts features from the multi-dimensional feature vectors of each word segment in the candidate phrase to achieve semantic representation of the sequence; this can be a neural network such as a long short-term memory model used for sequence semantic representation. The second semantic representation layer extracts features from the multi-dimensional feature vectors of each word segment in the context of the candidate phrase; this can also be a neural network such as a long short-term memory model used for sequence semantic representation. The multi-layer fully connected layer transforms the features to reduce their dimension from high to low. The classification layer performs classification calculations based on the features, such as using a sigmoid binary classification function.

[0103] It should be noted that the candidate phrase classification model in this embodiment is a trained model, which can be trained based on a labeled training set or self-trained based on a labeled training set and an unlabeled dataset.

[0104] based on Figure 5The candidate phrase classification model shown in this disclosure, and the skill tag extraction method implemented herein, may include the following steps: segmenting the sentence to be extracted to obtain multiple segments and determining the multidimensional feature vector of each segment; extracting candidate phrases from the sentence to be extracted; determining the multidimensional feature vector of each segment in the candidate phrase based on the multidimensional feature vector of each segment; inputting the multidimensional feature vector of each segment in the candidate phrase into the first semantic representation layer to generate the semantic representation vector of the candidate phrase; inputting the multidimensional feature vector of each segment in the context of the candidate phrase into the second semantic representation layer to generate the semantic representation vector of the context of the candidate phrase; concatenating the semantic representation vector of the candidate phrase and the semantic representation vector of the context of the candidate phrase to obtain the concatenated feature vector; inputting the concatenated feature vector into a multi-layer fully connected layer to perform feature transformation on the concatenated feature vector to obtain the transformed feature vector; inputting the transformed feature vector into a classification layer to perform classification calculation based on the transformed feature vector to obtain the classification probability of the candidate phrase; and determining the candidate phrase as a skill tag if the classification probability is greater than a preset threshold.

[0105] The skill tag extraction method according to the embodiments of this disclosure adds a multi-dimensional feature vector based on each word segmentation in the candidate phrase context to generate a semantic representation vector of the candidate phrase context, and extracts skill tags from the sentence to be extracted based on the semantic representation vector of the candidate phrase and the semantic representation vector of the candidate phrase context. In other words, the skill tag extraction process can take into account the semantic representation of the candidate phrase context, thereby further improving the accuracy of skill tag extraction.

[0106] To improve the accuracy of skill tag extraction, this disclosure proposes yet another embodiment.

[0107] Figure 6 This is a flowchart illustrating another method for extracting skill tags provided in this embodiment of the disclosure. Figure 6 As shown, the implementation method may include the following steps:

[0108] Step 601: Perform word segmentation on the sentence to be extracted to obtain multiple words, and determine the multidimensional feature vector of each word.

[0109] Step 602: Extract candidate phrases from the sentences to be extracted.

[0110] Step 603: Determine the multidimensional feature vectors of each word segment in the candidate phrase based on the multidimensional feature vectors of each word segment.

[0111] Step 604: Generate the semantic representation vector of the candidate phrase based on the multidimensional feature vectors of each word segment in the candidate phrase.

[0112] Step 605: Determine the multidimensional feature vector of each word in the candidate phrase context based on the multidimensional feature vector of each word and the preset window size; wherein, the candidate phrase context contains the candidate phrase.

[0113] Step 606: Generate the semantic representation vector of the candidate phrase context based on the multidimensional feature vectors of each word segment in the candidate phrase context.

[0114] Step 607: Generate other feature representation vectors of candidate phrases based on the preset candidate phrase feature engineering.

[0115] In some embodiments of this disclosure, candidate phrase feature engineering aims to extract additional features from the raw data to improve the accuracy of skill tag extraction. As an example, candidate phrase feature engineering may include features such as the number of characters in the candidate phrase, the position of the candidate phrase in the sentence to be extracted, and whether the candidate phrase consists of English words.

[0116] As one implementation method, the information of each feature in the candidate phrase feature engineering can be determined based on the candidate phrase feature engineering, thereby generating other feature representation vectors of the candidate phrase based on the expression of feature information.

[0117] Step 608: Extract skill tags from the sentences to be extracted based on the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase.

[0118] As an example, the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase's context, and other feature representation vectors of the candidate phrase can be concatenated to obtain a concatenated feature vector. Based on this concatenated feature vector, the classification probability of the candidate phrase is obtained. If the classification probability is greater than a preset threshold, the candidate phrase is identified as a skill tag. The classification probability of the candidate phrase can be obtained through classification calculation using a classifier function.

[0119] It should be noted that the skill tag extraction method of this disclosure embodiment can also be implemented based on a candidate phrase classification model, and its model structure is similar to... Figure 5The candidate phrase classification model is consistent with that in the text, and its implementation may include: segmenting the sentence to be extracted to obtain multiple segments and determining the multidimensional feature vector of each segment; extracting candidate phrases from the sentence to be extracted; determining the multidimensional feature vector of each segment in the candidate phrase based on the multidimensional feature vector of each segment; inputting the multidimensional feature vector of each segment in the candidate phrase into the first semantic representation layer to generate the semantic representation vector of the candidate phrase; inputting the multidimensional feature vector of each segment in the context of the candidate phrase into the second semantic representation layer to generate the semantic representation vector of the context of the candidate phrase; and according to the preset candidate... The phrase feature engineering process generates additional feature representation vectors for candidate phrases. The semantic representation vectors of the candidate phrases, their context, and other feature representation vectors are concatenated to obtain a concatenated feature vector. This concatenated feature vector is then input into a multi-layer fully connected layer for feature transformation, yielding a transformed feature vector. The transformed feature vector is then input into a classification layer for classification calculation, obtaining the classification probability of the candidate phrase. If the classification probability exceeds a preset threshold, the candidate phrase is identified as a skill tag.

[0120] It should be noted that, since the semantic representation vectors of candidate phrases, the semantic representation vectors of candidate phrase contexts, and other representation vectors of candidate phrases may have relatively high dimensions, the semantic representation vectors of candidate phrases, the semantic representation vectors of candidate phrase contexts, and other representation vectors of candidate phrases can be first transformed through fully connected layers, and then the feature vectors obtained after feature transformation can be concatenated. As an example, the structure of the candidate phrase classification model in this embodiment can also be as follows. Figure 7 As shown, the outputs of the first and second semantic representation layers can be transformed through fully connected layers. Other feature representation vectors of candidate phrases can also be transformed through fully connected layers first, and then the transformed features are concatenated and input into multiple fully connected layers.

[0121] The skill tag extraction method according to the embodiments of this disclosure adds a step of generating other feature representation vectors of candidate phrases based on candidate phrase feature engineering, and extracts skill tags from the sentences to be extracted based on the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase. This can make the feature expression of the candidate phrase more comprehensive, thereby further improving the accuracy of skill tag extraction.

[0122] To implement the above embodiments, this disclosure provides a training method for a candidate phrase classification model.

[0123] Figure 8 This is a flowchart illustrating a training method for a candidate phrase classification model provided in an embodiment of this disclosure.Figure 8 As shown, the method may include:

[0124] Step 801: Obtain the labeled training set and the unlabeled dataset; wherein, the labeled training set includes the first sentence sample and its corresponding skill label sample, and the unlabeled dataset includes the second sentence sample and its corresponding candidate phrase sample.

[0125] In some embodiments of this disclosure, the labeled training set is labeled data, including a first statement sample and its corresponding skill tag sample. The skill tag sample is the skill tag annotated in the first statement sample. The first statement sample may come from resume text, job description text, talent evaluation data, etc. Since each first statement sample may contain multiple skill tag samples, for ease of data use, each skill tag sample and its corresponding first statement sample in the labeled training set can be treated as a set of sample data. For example, if the skill tag samples corresponding to the first statement sample s are skill tag sample L1, skill tag sample L2, and skill tag sample L3, then the data format in the labeled training set is: {first statement sample s, skill tag sample L1}, {first statement sample s, skill tag sample L2}, {first statement sample s, skill tag sample L3}.

[0126] Furthermore, unlabeled datasets refer to data that has not been labeled, including second-statement samples and their corresponding candidate phrase samples. Candidate phrase samples are extracted from first-statement samples. Second-statement samples can come from resume text, job description text, talent evaluation data, etc. Moreover, the composition of samples in the unlabeled dataset can be consistent with the composition of samples in the labeled training set; that is, each candidate phrase sample and its corresponding second-statement sample form a set of sample data. As an example, obtaining candidate phrase samples can be implemented by: segmenting the second-statement samples to obtain multiple segments, and obtaining the part-of-speech tags for each segment using a part-of-speech tagging model; based on a preset candidate phrase template and the part-of-speech tags of each segment, extracting candidate phrases from the second-statement samples to obtain candidate phrase samples corresponding to each second-statement sample.

[0127] It should be noted that, since manual annotation is costly, the number of samples in the unlabeled dataset in this embodiment can be much larger than the number of samples in the labeled training set, in order to reduce the cost of obtaining training samples.

[0128] Step 802: Train the candidate phrase classification model based on the first sentence sample and its corresponding skill tag sample to obtain the trained candidate phrase classification model.

[0129] In some embodiments of this disclosure, the candidate phrase classification model is used to classify candidate phrases to determine whether a candidate phrase is a skill tag. The candidate phrase classification model may include a semantic representation layer and a classification layer. Feature information of the candidate phrase is input into the semantic representation layer of the candidate phrase classification model to extract features of the candidate phrase, obtaining a semantic representation vector of the candidate phrase. The obtained semantic representation vector is then input into the classification layer for classification calculation, outputting the classification probability of the candidate phrase.

[0130] As an example, the candidate phrase classification model can be trained as follows: The first sentence sample is segmented into words, and the semantic features, part-of-speech features, and syntactic dependency segmentation features of each word are obtained to determine the feature information of each word in each skill tag sample. This feature information is then input into the candidate phrase classification model to obtain the predicted classification probability. Based on the predicted classification probability and the true classification value of the skill tag sample, the loss value is calculated. The model parameters are continuously adjusted based on the loss value to train the candidate phrase classification model.

[0131] Step 803: Based on the trained candidate phrase classification model, predict the classification probability of candidate phrase samples in the unlabeled dataset.

[0132] In other words, the trained candidate phrase classification model has learned the ability to represent sequence semantics and to perform classification calculations based on the semantic representation results. Therefore, it can predict the classification probability of candidate phrase samples in a dataset without standard based on the trained candidate phrase classification model.

[0133] As one implementation method, the process of predicting the classification probability of candidate phrase samples in an unlabeled dataset may include: segmenting the second sentence sample into words to obtain multiple second words, and obtaining the multidimensional feature vector of each second word, wherein the multidimensional feature vector may include semantic features, part-of-speech features, dependency parsing features, etc.; determining the second target word contained in each candidate phrase sample, and determining the multidimensional feature vector of each second word in each candidate phrase sample based on the multidimensional feature vector of each second word in the second sentence sample; and inputting the multidimensional feature vector of each second word in the candidate phrase into the trained candidate phrase classification model to predict the classification probability of candidate phrase samples in the unlabeled dataset.

[0134] It should be noted that when this step is executed for the first time, the second sentence sample can be segmented and the multidimensional feature vector of each segment can be obtained. When this step is executed again in a loop, the segmentation process can be skipped, and the multidimensional feature vector of each segment can be obtained directly using the multidimensional feature vector of each segment obtained when this step is executed for the first time to determine the multidimensional feature vector of each segment in the candidate phrase sample.

[0135] Furthermore, the candidate phrase classification model trained in step 803 refers to the candidate phrase classification model obtained after the latest training at the time of executing this step. Here, the unlabeled dataset refers to the latest unlabeled dataset.

[0136] Step 804: Update the labeled training set and the unlabeled dataset based on the classification probability.

[0137] It is understandable that, due to the small number of samples in the initial labeled training set, the predictive ability of the trained candidate phrase classification model is limited, although the accuracy is higher for more obvious candidate phrases (those that are easier to identify as skill tags). Therefore, based on the classification probability, a portion of the unlabeled dataset can be obtained and used as labeled data to participate in the model training, thereby gradually increasing the sample size in the labeled training set and achieving the goal of model self-training.

[0138] In some embodiments of this disclosure, updating the labeled training set and the unlabeled dataset based on classification probability may include: obtaining target candidate phrase samples with classification probabilities greater than a probability threshold from candidate phrase samples; adding the target candidate phrase samples and their corresponding second sentence samples to the labeled training set, wherein the target candidate phrase samples serve as skill label samples of their corresponding second sentence samples; and deleting the target candidate phrase samples and their corresponding second sentence samples from the unlabeled dataset.

[0139] It should be noted that the update operations for the labeled training set and the unlabeled dataset are based on the latest labeled training set and the unlabeled dataset at the time of execution of this step.

[0140] Step 805: Train the candidate phrase classification model based on the updated labeled training set.

[0141] In other words, based on the updated labeled training set, the trained candidate phrase classification model is trained, and based on the new trained candidate phrase classification model, the classification probability of candidate phrase samples in the updated unlabeled dataset is predicted. Based on the classification probability, the labeled training set and the unlabeled dataset are updated again, and this process is repeated.

[0142] It is understandable that since all sentence samples in the updated labeled training set (including the first sentence sample and the second sentence sample added to the labeled training set) have been segmented, and the feature information of each segmented word has been obtained, the feature information of the skill tag samples (including the skill tag samples corresponding to the first sentence sample and the candidate tags corresponding to the second sentence samples added to the labeled training set) can be directly input into the trained candidate phrase classification model to obtain the classification probability of the skill tag samples, and then the loss value is calculated. The model is then trained based on the loss value.

[0143] In this embodiment of the disclosure, after executing step 805, the process returns to step 803 to repeatedly train the candidate classification model until a preset condition is met. For example, the number of training iterations can be preset, and training ends when the number of iterations reaches the preset maximum; alternatively, training can end based on the updates to the labeled training set and the unlabeled dataset, if the updated labeled training set and the unlabeled dataset show no changes.

[0144] According to the training method of the candidate phrase classification model in this disclosure, the candidate phrase classification model is trained on a labeled training set, and the labeled training set and the unlabeled dataset are continuously updated based on the prediction results of the classification probability of candidate phrase samples in the unlabeled dataset by the trained candidate phrase classification model, so as to iteratively train the candidate phrase classification model. In other words, the model can be trained using only a small amount of labeled data, which not only improves the training effect of the model through self-training, but also reduces the cost of training data.

[0145] The following section will provide a detailed introduction to the training process of the candidate phrase classification model, based on its model structure.

[0146] Figure 9 A flowchart illustrating a training method for another candidate phrase classification model provided in this embodiment. In this embodiment, the candidate phrase classification model includes a first semantic representation layer, a classification layer, and multiple fully connected layers. Figure 2 The model structure shown is consistent. Based on the above embodiments, Figure 8 Step 802 in the process involves training the candidate phrase classification model based on the first sentence sample and its corresponding skill tag sample, obtaining the trained candidate phrase classification model as follows: Figure 9 As shown:

[0147] Step 901: Perform word segmentation on the first sentence sample to obtain multiple first words, and determine the multidimensional feature vector of each first word.

[0148] In some embodiments of this disclosure, the first sentence sample can be segmented using word segmentation tools in related technologies, or it can be segmented using a word segmentation model built by those skilled in the art. This disclosure does not limit the scope of the first sentence sample.

[0149] Here, the multidimensional feature vector of each first word segment refers to a vector that can characterize the semantic, part-of-speech, and other features of each first word segment. For example, the multidimensional feature vector of each first word segment can be based on a sequence composed of the semantic features, part-of-speech features, and dependency parsing features of that first word segment. As an example, the semantic features of each first word segment can be obtained using word vector models such as word2Vec; the part-of-speech features of each first word segment can be obtained using part-of-speech tagging models in related technologies; and the dependency parsing features of each first word segment can be obtained using dependency parsing tools in related technologies.

[0150] Step 902: Determine the multidimensional feature vector of each first word in the skill tag sample based on the multidimensional feature vector of each first word segment.

[0151] It is understandable that, since the skill tag sample can include at least one first segment, the multidimensional feature vector of each first segment can be determined based on the multidimensional feature vector of each first segment.

[0152] In some embodiments of this disclosure, the implementation process of this step may include: determining at least one first target word contained in each skill tag sample; and determining the multidimensional feature vector of each first target word based on the multidimensional feature vector of each first word. That is, the first target word contained in the skill tag sample can be determined first, and then the multidimensional feature vector of the first target word can be found from the multidimensional feature vector of each first word.

[0153] Step 903: Input the multidimensional feature vectors of each first word segment in the skill tag sample into the first semantic representation layer to generate the semantic representation vector of the skill tag sample.

[0154] Step 904: Based on a multi-layer fully connected layer, perform feature transformation on the semantic representation vector of the skill tag sample to obtain the transformed feature vector.

[0155] It should be noted that the number of fully connected layers here can be determined based on the dimension of the feature vector, and this disclosure does not impose any limitation on it.

[0156] Step 905: Based on the classification layer, perform classification calculations on the transformed feature vectors to obtain the classification probability of the skill tag samples.

[0157] Step 906: Train the candidate phrase classification model based on the classification probability of the skill tag samples.

[0158] In this embodiment of the disclosure, the process of training a candidate phrase classification model based on the classification probability of skill tag samples may include: obtaining the true classification value of the skill tag samples; obtaining a loss value based on the true classification value and the classification probability of the skill tag samples; and training the candidate phrase classification model based on the loss value.

[0159] First, it should be noted that, regarding Figure 2 During the training process of the candidate phrase classification model shown, Figure 8 The implementation of step 805 can include only steps 903-905, that is, there is no need to perform word segmentation and multi-dimensional feature vector acquisition again.

[0160] It should also be noted that, in Figure 2 During the training of the candidate phrase classification model shown, steps 901-905 above can also be used. Figure 8 In step 803, if step 803 is being executed for the first time, its implementation may include: segmenting the second sentence in the unlabeled dataset to obtain multiple second segments, and determining the multidimensional feature vector of each second segment; determining the multidimensional feature vector of each second segment in the candidate phrase samples in the unlabeled dataset based on the multidimensional feature vector of each second segment; inputting the multidimensional feature vector of each second segment in the candidate phrase samples into the first semantic representation layer of the candidate phrase classification model to generate the semantic representation vector of the candidate phrase samples; performing feature transformation on the semantic representation vector of the candidate phrase samples based on multiple fully connected layers to obtain the transformed feature vector; and performing classification calculation on the transformed feature vector based on the classification layer to predict the classification probability of the candidate phrase samples.

[0161] For non-first executions of step 803, the implementation may include: inputting the multidimensional feature vectors of each second word segment in the candidate phrase samples in the updated unlabeled dataset into the first semantic representation layer of the candidate phrase classification model to generate the semantic representation vector of the candidate phrase samples; performing feature transformation on the semantic representation vector of the candidate phrase samples based on multiple fully connected layers to obtain the transformed feature vector; and performing classification calculation on the transformed feature vector based on the classification layer to predict the classification probability of the candidate phrase samples.

[0162] According to the training method of the candidate phrase classification model in this disclosure, the candidate phrase classification model is trained on a labeled training set, and the labeled training set and the unlabeled dataset are continuously updated based on the prediction results of the classification probability of candidate phrase samples in the unlabeled dataset by the trained candidate phrase classification model, so as to iteratively train the candidate phrase classification model. In other words, the model can be trained using only a small amount of labeled data, which not only improves the training effect of the model through self-training, but also reduces the cost of training data.

[0163] To further improve the training performance of the model, this disclosure proposes yet another embodiment.

[0164] Figure 10 This is a flowchart illustrating a training method for another candidate phrase classification model provided in this embodiment. In this embodiment, the candidate phrase classification model includes a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers, and its structure is similar to... Figure 5 The model structure shown is consistent. Based on the above embodiments, Figure 8 Step 802 in the process involves training the candidate phrase classification model based on the first sentence sample and its corresponding skill tag sample, obtaining the trained candidate phrase classification model as follows: Figure 10 As shown:

[0165] Step 1001: Perform word segmentation on the first sentence sample to obtain multiple first words, and determine the multidimensional feature vector of each first word.

[0166] Step 1002: Determine the multidimensional feature vector of each first word in the skill tag sample based on the multidimensional feature vector of each first word segment.

[0167] Step 1003: Input the multidimensional feature vectors of each first word segment in the skill tag sample into the first semantic representation layer to generate the semantic representation vector of the skill tag sample.

[0168] Step 1004: Determine the multidimensional feature vector of each first word in the skill tag sample context based on the multidimensional feature vector of each first word and the preset window size; wherein, the skill tag sample context contains skill tag samples.

[0169] Since the semantic representation of skill tag samples is usually related to the context of the skill tag samples, in order to improve the training effect of the model, we can obtain the multi-dimensional feature vectors of each first word in the context of the skill tag samples, so as to obtain more comprehensive features of the skill tag samples.

[0170] Step 1005: Input the multidimensional feature vectors of each first word segment in the skill tag sample context into the second semantic representation layer to generate the semantic representation vector of the skill tag sample context.

[0171] Step 1006: Perform feature concatenation on the semantic representation vector of the skill tag sample and the semantic representation vector of the skill tag sample context to obtain the concatenated feature vector.

[0172] It is understandable that the context of a skill tag usually has a certain influence on its semantic representation. Therefore, in order to improve the training effect of the model, the features of the semantic representation vector of the skill tag sample and the semantic representation vector of the context of the skill tag sample can be concatenated.

[0173] In the embodiments of this disclosure, the feature concatenation of the semantic representation vector of the skill tag sample and the semantic representation vector of the skill tag sample context refers to the concatenation of features in each dimension. For example, if the dimension of the semantic representation vector of the skill tag sample is 128 and the dimension of the semantic representation vector of the skill tag sample context is 128, then the dimension of the feature vector after feature concatenation is 258.

[0174] It should be noted that since the semantic representation vectors of skill tag samples and the semantic representation vectors of skill tag sample contexts may have relatively high dimensions, the semantic representation vectors of skill tag samples and the semantic representation vectors of skill tag sample contexts can be transformed by fully connected layers first, and then the feature vectors obtained after feature transformation can be concatenated.

[0175] Step 1007: Input the concatenated feature vector into a multi-layer fully connected layer to perform feature transformation on the concatenated feature vector and obtain the transformed feature vector.

[0176] Step 1008: Based on the classification layer, perform classification calculations on the transformed feature vectors to obtain the classification probability of the skill tag samples.

[0177] Step 1009: Train the candidate phrase classification model based on the classification probability of the skill tag samples.

[0178] First, it should be noted that, regarding Figure 5 During the training process of the candidate phrase classification model shown, Figure 8 The implementation of step 805 can include only steps 1003-1009, that is, there is no need to perform word segmentation and multi-dimensional feature vector acquisition again.

[0179] It should also be noted that, in Figure 5 During the training of the candidate phrase classification model shown, steps 1001-1008 above can also be used. Figure 8In step 803, this disclosure will not elaborate further.

[0180] According to the training method of the candidate phrase classification model in the present disclosure, a second semantic representation layer is added to the model structure to generate a semantic representation vector of the candidate phrase context by utilizing the multidimensional feature vectors of each word segment in the candidate phrase context. This makes the features of the obtained candidate phrases more comprehensive, thereby improving the training effect of the candidate phrase model and also improving the transferability of the candidate phrase model.

[0181] To further improve the training performance of the model, this disclosure proposes yet another embodiment.

[0182] Figure 11 This is a flowchart illustrating a training method for another candidate phrase classification model provided in this embodiment. In this embodiment, the candidate phrase classification model includes a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers, and its structure is similar to... Figure 5 The model structure shown is consistent. Based on the above embodiments, Figure 8 Step 802 in the process involves training the candidate phrase classification model based on the first sentence sample and its corresponding skill tag sample, obtaining the trained candidate phrase classification model as follows: Figure 11 As shown:

[0183] Step 1101: Perform word segmentation on the first sentence sample to obtain multiple first words, and determine the multidimensional feature vector of each first word.

[0184] Step 1102: Determine the multidimensional feature vector of each first word in the skill tag sample based on the multidimensional feature vector of each first word segment.

[0185] Step 1103: Input the multidimensional feature vectors of each first word segment in the skill tag sample into the first semantic representation layer to generate the semantic representation vector of the skill tag sample.

[0186] Step 1104: Determine the multidimensional feature vector of each first word in the skill tag sample context based on the multidimensional feature vector of each first word and the preset window size; wherein, the skill tag sample context contains skill tag samples.

[0187] Step 1105: Input the multidimensional feature vectors of each first word segment in the skill tag sample context into the second semantic representation layer to generate the semantic representation vector of the skill tag sample context.

[0188] Step 1106: Generate other feature representation vectors for skill tag samples based on the preset candidate phrase feature engineering.

[0189] In some embodiments of this disclosure, candidate phrase feature engineering aims to extract additional features from the raw data to improve the accuracy of model training. As an example, candidate phrase feature engineering may include features such as the number of characters in the candidate phrase, the position of the candidate phrase in the sentence to be extracted, and whether the candidate phrase consists of English words.

[0190] As one implementation method, the information of each feature in the candidate phrase feature engineering corresponding to the skill tag sample can be determined according to the candidate phrase feature engineering, and then other feature representation vectors of the skill tag sample can be generated based on the expression of feature information.

[0191] Step 1107: Perform feature concatenation on the semantic representation vector of the skill tag sample, the semantic representation vector of the skill tag sample context, and other feature representation vectors of the skill tag sample to obtain the concatenated feature vector.

[0192] It should be noted that, since the semantic representation vectors of skill tag samples, the semantic representation vectors of skill tag sample contexts, and other representation vectors of skill tag samples may have relatively high dimensionality, the semantic representation vectors of skill tag samples, the semantic representation vectors of skill tag sample contexts, and other representation vectors of skill tag samples can be transformed separately through fully connected layers, and then the feature vectors obtained after feature transformation can be concatenated. As an example, the structure of the candidate phrase classification model in this embodiment can also be as follows. Figure 10 As shown, the outputs of the first and second semantic representation layers can be transformed through fully connected layers. Other feature representation vectors can also be transformed through fully connected layers first, and then the transformed features are concatenated and input into multiple fully connected layers.

[0193] Step 1108: Input the concatenated feature vector into a multi-layer fully connected layer to perform feature transformation on the concatenated feature vector and obtain the transformed feature vector.

[0194] Step 1109: Based on the classification layer, perform classification calculations on the transformed feature vectors to obtain the classification probability of the skill tag samples.

[0195] Step 1110: Train the candidate phrase classification model based on the classification probability of the skill tag samples.

[0196] First, it should be noted that during the training process of the candidate phrase classification model in this embodiment of the disclosure... Figure 8 The implementation of step 805 can include only steps 1103-1110, that is, there is no need to perform word segmentation and multi-dimensional feature vector acquisition again.

[0197] It should also be noted that, during the training of the candidate phrase classification model according to the embodiments of this disclosure, the above steps 1101-1109 can also be used... Figure 8 In step 803, this disclosure will not elaborate further.

[0198] According to the training method of the candidate phrase classification model in the present disclosure, other feature representation vectors of skill tag samples can be generated through preset candidate phrase feature engineering, making the obtained spliced ​​features richer, thereby further improving the training effect of the model and making the accuracy of the trained candidate phrase classification model higher.

[0199] To implement the above embodiments, this disclosure proposes a skill tag extraction device.

[0200] Figure 12 This is a structural block diagram of a skill tag extraction device provided in an embodiment of this disclosure. Figure 12 As shown, the device may include:

[0201] The first determining module 1201 is used to perform word segmentation on the sentence to be extracted, obtain multiple words, and determine the multi-dimensional feature vector of each word.

[0202] The first extraction module 1202 is used to extract candidate phrases from the statement to be extracted;

[0203] The second determining module 1203 is used to determine the multidimensional feature vector of each word in the candidate phrase based on the multidimensional feature vector of each word segment;

[0204] The first generation module 1204 is used to generate the semantic representation vector of the candidate phrase based on the multidimensional feature vector of each word segment in the candidate phrase;

[0205] The second extraction module 1205 is used to extract skill tags from the sentences to be extracted based on the semantic representation vector of the candidate phrases.

[0206] In some embodiments of this disclosure, the first extraction module 1202 is specifically used for:

[0207] From the multidimensional feature vector of each word segment, obtain the part-of-speech tag of each word segment;

[0208] Based on the preset candidate phrase templates and the part-of-speech tags of each word, candidate phrases are extracted from the sentences to be extracted.

[0209] In some embodiments of this disclosure, the second determining module 1203 is specifically used for:

[0210] Identify at least one target word contained in the candidate phrase;

[0211] Based on the multidimensional feature vector of each word segment, determine the multidimensional feature vector of each target word segment.

[0212] In some embodiments of this disclosure, the second extraction module 1205 is specifically used for:

[0213] Based on the semantic representation vector of the candidate phrase, obtain the classification probability of the candidate phrase;

[0214] When the classification probability is greater than a preset threshold, the candidate phrase is identified as a skill tag.

[0215] In other embodiments of this disclosure, the device may further include:

[0216] The third determining module 1206 is used to determine the multidimensional feature vector of each word in the candidate phrase context based on the multidimensional feature vector of each word and the preset window size; wherein, the candidate phrase context contains the candidate phrase;

[0217] The second generation module 1207 is used to generate a semantic representation vector of the candidate phrase context based on the multidimensional feature vectors of each word segment in the candidate phrase context.

[0218] The second extraction module 1205 is also used for:

[0219] Skill tags are extracted from the sentences to be extracted based on the semantic representation vectors of the candidate phrases and the semantic representation vectors of the candidate phrase contexts.

[0220] In some further embodiments of this disclosure, the device may further include:

[0221] The third generation module 1208 is used to generate other feature representation vectors of candidate phrases based on the preset candidate phrase feature engineering.

[0222] The second extraction module 1205 is also used for:

[0223] Skill tags are extracted from the sentences to be extracted based on the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase.

[0224] As one implementation method, the device can be based on a preset candidate phrase classification model; the candidate phrase classification model includes a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers; wherein, the first generation module 1205 is specifically used for:

[0225] The multidimensional feature vectors of each word segment in the candidate phrase are input into the first semantic representation layer to generate the semantic representation vector of the candidate phrase.

[0226] Specifically, the second generation module 1207 is used for:

[0227] The multidimensional feature vectors of each word segment in the candidate phrase context are input into the second semantic representation layer to generate the semantic representation vector of the candidate phrase context;

[0228] Specifically, the second extraction module 1205 is used for:

[0229] The semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase are concatenated to obtain the concatenated feature vector.

[0230] The concatenated feature vector is input into a multi-layer fully connected layer to perform feature transformation on the concatenated feature vector and obtain the transformed feature vector.

[0231] The transformed feature vector is input into the classification layer to perform classification calculations based on the transformed feature vector, thereby obtaining the classification probability of the candidate phrase.

[0232] When the classification probability is greater than a preset threshold, the candidate phrase is identified as a skill tag.

[0233] In some further embodiments of this disclosure, the device may further include:

[0234] The fourth determination module 1209 is used to determine the category to which the skill tags belong after extracting skill tags from the statements to be extracted.

[0235] The skill tag extraction device according to embodiments of this disclosure extracts candidate phrases from the sentence to be extracted, generates semantic representation vectors of the candidate phrases based on the multidimensional feature vectors of each word segment in the candidate phrases, and extracts skill tags from the sentence to be extracted based on the semantic representation vectors of the candidate phrases, thereby achieving automatic skill tag extraction. Utilizing the multidimensional feature vectors of each word segment in the candidate phrases for skill tag extraction can effectively improve the accuracy of skill tag extraction and provide necessary conditions for enterprises to achieve intelligent talent management.

[0236] To implement the above embodiments, this disclosure provides a training apparatus for a candidate phrase classification model.

[0237] Figure 13 This is a structural block diagram of a training device for a candidate phrase classification model provided in an embodiment of this disclosure. Figure 13 As shown, the device may include:

[0238] The acquisition module 1301 is used to acquire the labeled training set and the unlabeled dataset; wherein, the labeled training set includes the first sentence sample and its corresponding skill label sample, and the unlabeled dataset includes the second sentence sample and its corresponding candidate phrase sample.

[0239] Training module 1302 is used to train the candidate phrase classification model based on the first sentence sample and its corresponding skill tag sample, so as to obtain the trained candidate phrase classification model.

[0240] The prediction module 1303 is used to predict the classification probability of candidate phrase samples in the unlabeled dataset based on the trained candidate phrase classification model.

[0241] Update module 1304 is used to update the labeled training set and the unlabeled dataset based on the classification probability;

[0242] Among them, the training module 1302 is also used to train the trained candidate phrase classification model based on the updated labeled training set.

[0243] In one implementation, the update module 1304 is specifically used for:

[0244] Obtain target candidate phrase samples whose classification probability is greater than a probability threshold from the candidate phrase samples;

[0245] Add the target candidate phrase sample and its corresponding second sentence sample to the labeled training set; wherein, the target candidate phrase sample serves as the skill label sample of its corresponding second sentence sample;

[0246] Remove the target candidate phrase samples and their corresponding second sentence samples from the unlabeled dataset.

[0247] In some embodiments of this disclosure, the candidate phrase classification model includes: a first semantic representation layer, a classification layer, and multiple fully connected layers; the training module 1302 is specifically used for:

[0248] The first sentence sample is segmented into words to obtain multiple first words, and the multidimensional feature vector of each first word is determined.

[0249] Based on the multidimensional feature vector of each first word segment, determine the multidimensional feature vector of each first word segment in the skill tag sample;

[0250] The multidimensional feature vectors of each first word segment in the skill tag sample are input into the first semantic representation layer to generate the semantic representation vector of the skill tag sample;

[0251] Based on multiple fully connected layers, feature transformation is performed on the semantic representation vector of skill tag samples to obtain the transformed feature vector;

[0252] Based on the classification layer, the transformed feature vector is classified to obtain the classification probability of the skill tag sample.

[0253] The candidate phrase classification model is trained based on the classification probability of the skill tag samples.

[0254] Specifically, training module 1302 is used for:

[0255] Obtain the true classification values ​​of skill tag samples;

[0256] The loss value is obtained based on the true classification value and classification probability of the skill tag sample.

[0257] The candidate phrase classification model is trained based on the loss value.

[0258] In other embodiments of this disclosure, the candidate phrase classification model further includes a second semantic representation layer; the training module 1302 is also used for:

[0259] Based on the multidimensional feature vector of each first word segment and the preset window size, determine the multidimensional feature vector of each first word segment in the skill tag sample context; wherein, the skill tag sample context contains skill tag samples;

[0260] The multidimensional feature vectors of each first word in the skill tag sample context are input into the second semantic representation layer to generate the semantic representation vector of the skill tag sample context;

[0261] The semantic representation vector of the skill tag sample and the semantic representation vector of the skill tag sample context are concatenated to obtain the concatenated feature vector;

[0262] The concatenated feature vectors are input into a multi-layer fully connected layer to perform feature transformation on the concatenated feature vectors, thereby obtaining the transformed feature vectors.

[0263] In some further embodiments of this disclosure, the training module 1302 is also used for:

[0264] Based on the preset candidate phrase feature engineering, generate other feature representation vectors for skill tag samples;

[0265] The semantic representation vector of the skill tag sample, the semantic representation vector of the skill tag sample context, and other feature representation vectors of the skill tag sample are concatenated to obtain the concatenated feature vector.

[0266] The training apparatus for the candidate phrase classification model according to embodiments of this disclosure trains the candidate phrase classification model based on a labeled training set, and continuously updates the labeled training set and the unlabeled dataset based on the prediction results of the classification probabilities of candidate phrase samples in the unlabeled dataset by the trained candidate phrase classification model, thereby iteratively training the candidate phrase classification model. In other words, model training can be achieved using only a small amount of labeled data, which not only improves the training effect of the model through self-training but also reduces the cost of training data.

[0267] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0268] Figure 14 A schematic block diagram of an example electronic device 1400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0269] like Figure 14 As shown, device 1400 includes a computing unit 1401, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1402 or a computer program loaded from storage unit 1408 into random access memory (RAM) 1403. The RAM 1403 may also store various programs and data required for the operation of device 1400. The computing unit 1401, ROM 1402, and RAM 1403 are interconnected via bus 1404. Input / output (I / O) interface 1405 is also connected to bus 1404.

[0270] Multiple components in device 1400 are connected to I / O interface 1405, including: input unit 1406, such as a keyboard, mouse, etc.; output unit 1407, such as various types of displays, speakers, etc.; storage unit 1408, such as a disk, optical disk, etc.; and communication unit 1409, such as a network card, modem, wireless transceiver, etc. Communication unit 1409 allows device 1400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0271] The computing unit 1401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1401 performs the various methods and processes described above, such as methods for extracting skill tags, and / or methods for training candidate phrase classification models. For example, in some embodiments, the methods for extracting skill tags, and / or the methods for training candidate phrase classification models, can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1400 via ROM 1402 and / or communication unit 1409. When the computer program is loaded into RAM 1403 and executed by computing unit 1401, one or more steps of the skill tag extraction method and / or candidate phrase classification model training method described above can be performed. Alternatively, in other embodiments, computing unit 1401 can be configured to perform the skill tag extraction method and / or candidate phrase classification model training method by any other suitable means (e.g., by means of firmware).

[0272] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0273] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0274] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0275] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0276] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0277] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0278] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0279] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for extracting skill tags, comprising: The extracted sentence is segmented to obtain multiple segments, and a multi-dimensional feature vector is determined for each segment. The multi-dimensional feature vector includes semantic features, part-of-speech features, and dependency parsing features. Extract candidate phrases from the statements to be extracted; Based on the multidimensional feature vector of each of the segmented words, determine the multidimensional feature vector of each segmented word in the candidate phrase; Based on the multidimensional feature vectors of each word segment in the candidate phrase, a semantic representation vector of the candidate phrase is generated; Based on the multidimensional feature vector of each segmented word and a preset window size, the multidimensional feature vector of each segmented word in the candidate phrase context is determined; wherein, the candidate phrase context contains the candidate phrase; Based on the multidimensional feature vectors of each word segment in the candidate phrase context, a semantic representation vector of the candidate phrase context is generated; Based on the preset candidate phrase feature engineering, other feature representation vectors of the candidate phrase are generated; Skill tags are extracted from the statement to be extracted based on the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase.

2. The method according to claim 1, wherein, The step of extracting candidate phrases from the statements to be extracted includes: From the multidimensional feature vector of each segmented word, obtain the part-of-speech tag of each segmented word; Based on the preset candidate phrase template and the part-of-speech tag of each word segment, candidate phrases are extracted from the sentence to be extracted.

3. The method according to claim 1, wherein, The step of determining the multidimensional feature vector of each word in the candidate phrase based on the multidimensional feature vector of each word segment includes: Identify at least one target word contained in the candidate phrase; The multidimensional feature vector of each target word is determined based on the multidimensional feature vector of each word segmentation.

4. The method according to claim 1, wherein, The method is implemented based on a preset candidate phrase classification model; the candidate phrase classification model includes a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers; generating the semantic representation vector of the candidate phrase based on the multidimensional feature vectors of each word segment in the candidate phrase includes: The multidimensional feature vectors of each word segment in the candidate phrase are input into the first semantic representation layer to generate the semantic representation vector of the candidate phrase; The step of generating a semantic representation vector of the candidate phrase context based on the multidimensional feature vectors of each word segment in the candidate phrase context includes: The multidimensional feature vectors of each word segment in the candidate phrase context are input into the second semantic representation layer to generate the semantic representation vector of the candidate phrase context; The step of extracting skill tags from the statement to be extracted based on the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase includes: The semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase are concatenated to obtain the concatenated feature vector; The concatenated feature vector is input into the multi-layer fully connected layer to perform feature transformation on the concatenated feature vector and obtain the transformed feature vector. The transformed feature vector is input into the classification layer to perform classification calculation based on the transformed feature vector, thereby obtaining the classification probability of the candidate phrase; In response to the classification probability being greater than a preset threshold, the candidate phrase is determined as a skill tag.

5. The method according to claim 1, after extracting skill tags from the statements to be extracted, further comprising: Determine the category to which the skill tag belongs.

6. A method for training a candidate phrase classification model, wherein the candidate phrase classification model comprises: The system consists of a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers, including: Obtain an labeled training set and an unlabeled dataset; wherein, the labeled training set includes a first statement sample and its corresponding skill tag sample, and the unlabeled dataset includes a second statement sample and its corresponding candidate phrase sample; The first sentence sample is segmented to obtain multiple first segments, and a multi-dimensional feature vector is determined for each first segment. The multi-dimensional feature vector includes semantic features, part-of-speech features, and dependency parsing features. Based on the multidimensional feature vector of each of the first word segments, determine the multidimensional feature vector of each of the first word segments in the skill tag sample; The multidimensional feature vectors of each first word segment in the skill tag sample are input into the first semantic representation layer to generate the semantic representation vector of the skill tag sample; Based on the multidimensional feature vector of each first word segment and a preset window size, determine the multidimensional feature vector of each first word segment in the skill tag sample context; wherein, the skill tag sample context includes the skill tag sample; The multidimensional feature vectors of each first word segment in the skill tag sample context are input into the second semantic representation layer to generate the semantic representation vector of the skill tag sample context; Based on the preset candidate phrase feature engineering, other feature representation vectors of the skill tag samples are generated; The semantic representation vector of the skill tag sample, the semantic representation vector of the skill tag sample context, and other feature representation vectors of the skill tag sample are concatenated to obtain the concatenated feature vector; The concatenated feature vector is input into the multi-layer fully connected layer to perform feature transformation on the concatenated feature vector and obtain the transformed feature vector. Based on the classification layer, the transformed feature vector is classified to obtain the classification probability of the skill tag sample; The candidate phrase classification model is trained based on the classification probabilities of the skill tag samples. Based on the trained candidate phrase classification model, predict the classification probability of candidate phrase samples in the unlabeled dataset; Based on the classification probabilities, update the labeled training set and the unlabeled dataset; The candidate phrase classification model is trained based on the updated labeled training set.

7. The method according to claim 6, wherein, The step of updating the labeled training set and the unlabeled dataset based on the classification probability includes: Obtain target candidate phrase samples whose classification probability is greater than a probability threshold from the candidate phrase samples; The target candidate phrase sample and its corresponding second sentence sample are added to the labeled training set; wherein, the target candidate phrase sample serves as the skill tag sample of its corresponding second sentence sample; The target candidate phrase sample and its corresponding second sentence sample are deleted from the unlabeled dataset.

8. The method according to claim 6, wherein, The step of training the candidate phrase classification model based on the classification probabilities of the skill tag samples includes: Obtain the true classification value of the skill tag sample; The loss value is obtained based on the true classification value and the classification probability of the skill tag sample. The candidate phrase classification model is trained based on the loss value.

9. A skill tag extraction device, comprising: The first determining module is used to perform word segmentation on the sentence to be extracted, obtain multiple words, and determine the multi-dimensional feature vector of each word, wherein the multi-dimensional feature vector includes semantic features, part-of-speech features, and dependency parsing features. The first extraction module is used to extract candidate phrases from the statement to be extracted; The second determining module is used to determine the multidimensional feature vector of each word in the candidate phrase based on the multidimensional feature vector of each word segmentation. The first generation module is used to generate the semantic representation vector of the candidate phrase based on the multidimensional feature vectors of each word segment in the candidate phrase; The third determining module is used to determine the multidimensional feature vector of each word in the candidate phrase context based on the multidimensional feature vector of each word segment and a preset window size; wherein, the candidate phrase context contains the candidate phrase; The second generation module is used to generate a semantic representation vector of the candidate phrase context based on the multidimensional feature vectors of each word segment in the candidate phrase context. The third generation module is used to generate other feature representation vectors of the candidate phrases based on the preset candidate phrase feature engineering. The second extraction module is used to extract skill tags from the statement to be extracted based on the semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase.

10. The apparatus according to claim 9, wherein, The first extraction module is specifically used for: From the multidimensional feature vector of each segmented word, obtain the part-of-speech tag of each segmented word; Based on the preset candidate phrase template and the part-of-speech tag of each word segment, candidate phrases are extracted from the sentence to be extracted.

11. The apparatus according to claim 9, wherein, The second determining module is specifically used for: Identify at least one target word contained in the candidate phrase; The multidimensional feature vector of each target word is determined based on the multidimensional feature vector of each word segmentation.

12. The apparatus according to claim 9, wherein, The device is implemented based on a preset candidate phrase classification model; the candidate phrase classification model includes a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers; the first generation module is specifically used for: The multidimensional feature vectors of each word segment in the candidate phrase are input into the first semantic representation layer to generate the semantic representation vector of the candidate phrase; Specifically, the second generation module is used for: The multidimensional feature vectors of each word segment in the candidate phrase context are input into the second semantic representation layer to generate the semantic representation vector of the candidate phrase context; Specifically, the second extraction module is used for: The semantic representation vector of the candidate phrase, the semantic representation vector of the candidate phrase context, and other feature representation vectors of the candidate phrase are concatenated to obtain the concatenated feature vector; The concatenated feature vector is input into the multi-layer fully connected layer to perform feature transformation on the concatenated feature vector and obtain the transformed feature vector. The transformed feature vector is input into the classification layer to perform classification calculation based on the transformed feature vector, thereby obtaining the classification probability of the candidate phrase; In response to the classification probability being greater than a preset threshold, the candidate phrase is determined as a skill tag.

13. The apparatus according to claim 9, further comprising: The fourth determining module is used to determine the category to which the skill tags belong after extracting skill tags from the statements to be extracted.

14. A training apparatus for a candidate phrase classification model, the candidate phrase classification model comprising: The system consists of a first semantic representation layer, a second semantic representation layer, a classification layer, and multiple fully connected layers, including: The acquisition module is used to acquire an labeled training set and an unlabeled dataset; wherein, the labeled training set includes a first statement sample and its corresponding skill tag sample, and the unlabeled dataset includes a second statement sample and its corresponding candidate phrase sample; The training module is used to perform word segmentation on the first sentence sample to obtain multiple first words and determine the multidimensional feature vector of each first word; based on the multidimensional feature vector of each first word, determine the multidimensional feature vector of each first word in the skill tag sample, wherein the multidimensional feature vector includes semantic features, part-of-speech features, and dependency parsing features; input the multidimensional feature vector of each first word in the skill tag sample to the first semantic representation layer to generate the semantic representation vector of the skill tag sample; based on the multidimensional feature vector of each first word and a preset window size, determine the multidimensional feature vector of each first word in the context of the skill tag sample; wherein the context of the skill tag sample includes the skill tag sample; and input the multidimensional feature vector of each first word in the context of the skill tag sample... The multidimensional feature vector is input into the second semantic representation layer to generate the semantic representation vector of the skill tag sample context; other feature representation vectors of the skill tag sample are generated according to the preset candidate phrase feature engineering; the semantic representation vector of the skill tag sample, the semantic representation vector of the skill tag sample context, and the other feature representation vectors of the skill tag sample are concatenated to obtain the concatenated feature vector; the concatenated feature vector is input into the multi-layer fully connected layer to perform feature transformation on the concatenated feature vector to obtain the transformed feature vector; based on the classification layer, the transformed feature vector is classified to obtain the classification probability of the skill tag sample; the candidate phrase classification model is trained according to the classification probability of the skill tag sample. The prediction module is used to predict the classification probability of candidate phrase samples in the unlabeled dataset based on the trained candidate phrase classification model. The update module is used to update the labeled training set and the unlabeled dataset based on the classification probabilities; The training module is further configured to train the trained candidate phrase classification model based on the updated labeled training set.

15. The apparatus according to claim 14, wherein, The update module is specifically used for: Obtain target candidate phrase samples whose classification probability is greater than a probability threshold from the candidate phrase samples; The target candidate phrase sample and its corresponding second sentence sample are added to the labeled training set; wherein, the target candidate phrase sample serves as the skill tag sample of its corresponding second sentence sample; The target candidate phrase sample and its corresponding second sentence sample are deleted from the unlabeled dataset.

16. The apparatus according to claim 14, wherein, The training module is specifically used for: Obtain the true classification value of the skill tag sample; The loss value is obtained based on the true classification value and the classification probability of the skill tag sample. The candidate phrase classification model is trained based on the loss value.

17. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-5, and / or, perform the method of any one of claims 6-8.

18. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-5, and / or to perform the method according to any one of claims 6-8.

19. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-5, and / or implements the method according to any one of claims 6-8.

Images