Label generation method and device, computer equipment and storage medium

Abstract

The invention provides a label generation method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring review information of a target object; performing information extraction on the evaluation information to obtain feature information in the evaluation information; and generating a review label of the target object according to the feature information. According to the label generation method provided by the embodiment of the invention, after the evaluation information of the target object is obtained, the evaluation information is subjected to information extraction processing, so that feature information of a plurality of simple semantics can be extracted from the evaluation information containing complex semantics; in the process of generating the evaluation label of the target object by using the feature information, the introduction of noise or the loss of semantic information of the evaluation text can be effectively avoided, so that the precision of the finally generated evaluation label is improved.

Description

Technical Field

[0001] The present application relates to the technical field of text processing, and in particular to a label generation method, apparatus, computer equipment and storage medium. Background Art

[0002] With the development of social media, many people like to record their feelings and comments after watching each film or TV work on major websites. Usually, these comments contain intuitive comments on the quality of the work itself, which can be used as a reference for others. Therefore, the labels of film and television works extracted from these comments will also affect the subsequent recommendation effect of video works.

[0003] Currently, the existing technology for extracting labels from numerous comments usually clusters the comment texts based on the similarity between the entire comment texts. However, the above solution is prone to introduce noise into the final result and also loses some comment semantics, thus affecting the accuracy of the final generated labels. Summary of the invention

[0004] Based on this, it is necessary to provide a label generation method, device, computer equipment and storage medium to address the above technical problems, so as to solve the problem that the existing label generation method is easy to introduce noise and lose some comment semantics, thus resulting in insufficient accuracy of the final generated labels.

[0005] In a first aspect, the present application provides a label generation method, comprising:

[0006] Obtain review information of the target object;

[0007] Extracting information from the review information to obtain feature information in the review information;

[0008] A comment label of the target object is generated according to the feature information.

[0009] As a feasible embodiment of the present application, generating a comment label of the target object according to the feature information includes:

[0010] Performing clustering processing on the feature information to obtain a feature information set of the target object;

[0011] Filtering a target feature information set from the feature information sets according to the amount of feature information in each of the feature information sets;

[0012] A comment label of the target object is generated according to the target feature information set.

[0013] As a feasible embodiment of the present application, clustering the feature information to obtain a feature information set of the target object includes:

[0014] Processing the feature information based on the first language model to obtain feature codes corresponding to each feature information;

[0015] Normalizing each of the feature codes to obtain a normalized feature code corresponding to each of the feature information;

[0016] According to the similarity between the normalized feature codes, each feature information is clustered to obtain a feature information set of the target object.

[0017] As a feasible embodiment of the present application, generating a comment label of the target object according to the target feature information set includes:

[0018] For each of the target feature information sets, input feature information in the target feature information set into a second language model to generate a description text corresponding to the target feature information set;

[0019] The description texts corresponding to each target feature information set are fused to obtain a comment label of the target object.

[0020] As a feasible embodiment of the present application, the step of inputting the feature information in the target feature information set into the second language model to generate a description text corresponding to the target feature information set includes:

[0021] Obtaining the correlation probability corresponding to each feature information in the target feature information set; the correlation probability corresponding to the feature information is obtained by processing the feature information through a pre-trained correlation detection model;

[0022] Based on the magnitude relationship of the correlation probabilities corresponding to the respective feature information, the feature information is sequentially input into the second language model to generate a description text corresponding to the target feature information set.

[0023] As a feasible embodiment of the present application, before the step of generating a comment label of the target object according to the feature information, the method further includes:

[0024] Processing the feature information according to a correlation detection model to obtain a correlation probability of the feature information corresponding to the target object;

[0025] Filtering target feature information from the feature information according to the magnitude relationship of the correlation probabilities;

[0026] The step of generating a comment label for the target object according to the feature information includes:

[0027] A comment label of the target object is generated according to the target feature information.

[0028] As a feasible embodiment of the present application, extracting information from the review information to obtain feature information in the review information includes:

[0029] According to the part-of-speech information of each word in the review information, the review information is segmented to obtain target words in the review information and the association relationship between the target words;

[0030] According to the association relationship between the target words, the target words are concatenated and combined to obtain the characteristic information of the review information.

[0031] As a feasible embodiment of the present application, after the step of generating a comment label of the target object according to the feature information, the method further includes:

[0032] Obtain search information of target users;

[0033] Determining a matching degree between the target object and the search information according to the search information and the comment label;

[0034] The target object is recommended to the target user according to the matching degree.

[0035] In a second aspect, the present application provides a label generation device, comprising:

[0036] An acquisition module is used to obtain the review information of the target object;

[0037] An extraction module, used to extract information from the review information to obtain feature information in the review information;

[0038] A generating module is used to generate a comment label of the target object according to the feature information.

[0039] In a third aspect, the present application further provides a computer device, the computer device comprising:

[0040] one or more processors;

[0041] Memory; and

[0042] One or more applications, wherein the one or more applications are stored in the memory and are configured to be executed by the processor to execute any of the label generation methods provided above.

[0043] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is loaded by a processor to execute any of the label generation methods described above.

[0044] The label generation method provided in the embodiment of the present application can extract some simple semantic feature information from the review information containing complex semantics by performing information extraction processing on the review information after acquiring the review information of the target object. Therefore, compared with processing the entire review text with complex semantics, in the process of generating the review label of the target object by using these feature information, it can effectively avoid introducing noise or losing the semantic information of the review text, thereby improving the accuracy of the review label finally generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 A schematic diagram of the steps of a label generation method provided in an embodiment of the present application;

[0047] Figure 2 A schematic diagram of a process flow for generating a comment label based on feature information provided in an embodiment of the present application;

[0048] Figure 3 A schematic diagram of a process flow for clustering feature information provided in an embodiment of the present application;

[0049] Figure 4 A schematic diagram of a process flow for generating a comment label of a target object according to a target feature information set provided in an embodiment of the present application;

[0050] Figure 5 A schematic diagram of a flow chart of steps for sorting propositions in a feature information set based on correlation probability provided in an embodiment of the present application;

[0051] Figure 6 A schematic diagram of a process flow for performing correlation detection on feature information provided in an embodiment of the present application;

[0052] Figure 7 A schematic diagram of a process flow for extracting feature information from review information provided in an embodiment of the present application;

[0053] Figure 8A schematic diagram of a process flow for recommending a target object provided in an embodiment of the present application;

[0054] Fig. 9 A complete implementation flow diagram of a label generation method provided in an embodiment of the present application;

[0055] Fig.10 A schematic diagram of the structure of a label generation device provided in an embodiment of the present application;

[0056] Fig.11 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0058] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "plurality" means two or more, unless otherwise clearly and specifically defined.

[0059] In the description of the present application, the word "for example" is used to mean "used as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any technician in the field to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes will not be elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed in the present application.

[0060] To facilitate understanding of the label generation method provided by the embodiment of the present application, the application scenario of the label generation method provided by the embodiment of the present application is first described. The label generation method provided by the embodiment of the present application is generally used to generate the label of the film and television work according to the review text of the film and television work shared by the user on the Internet, so as to facilitate the subsequent recommendation of the film and television work. Specifically, in most cases, the review text is usually composed of long sentences, containing several information units, such as "This is a classic of inspirational movies, you can revisit it when you are confused", and "This is a very inspirational film, in a certain sense, it is a subjective writing of American history". The two review texts, wherein the semantics of the two information units "This is a classic of inspirational movies" and "This is a very inspirational film" in the two review texts overlap, while the semantics of the remaining information units are not the same. This results in that in the existing label generation method for processing the entire review text provided by the user, if the two review texts are regarded as review texts with different semantics, noise will be introduced, and if the two review texts are regarded as review texts with the same semantics, other semantic information in the review texts will be lost, resulting in the final generated label result being not accurate enough.

[0061] In order to solve the above problems, the embodiments of the present application provide a label generation method, device, computer device and storage medium for generating a label for a label of a target object by extracting the review information containing complex semantics of the target object to obtain feature information containing simple semantic information in the review information, wherein the label generation method is usually installed in the label generation device in the form of a computer program, and the label generation device is provided in the computer device in the form of a processor, and the label generation device in the computer device executes the label generation method provided in the embodiments of the present application by running the computer program corresponding to the label generation method. Specifically, Figure 1 As shown, Figure 1 A schematic flow chart of a label generation method provided in an embodiment of the present application, specifically including steps S110 to S130:

[0062] S110, obtaining comment information of the target object.

[0063] In the embodiments of the present application, combined with the above-mentioned description of the application scenario, it can be seen that the target object here usually refers to works of art such as movies, articles and novels. For the convenience of description, the following descriptions are all based on the target objects of movies and novels. Specifically, compared with the review information in other fields, for example, compared with the review information in the field of commodity and product recommendation, the review information of these works of art, in addition to the intuitive review of the quality of the work itself, often also contains other richer emotional information, such as recording personal life or mood at that time, etc. Therefore, the review information of these works of art usually exists in the form of complex long sentences.

[0064] Furthermore, the review information in the embodiment of the present application may be collected from different website pages based on network technologies such as crawlers, and the embodiment of the present application does not specifically limit the method of obtaining the review information of the target object.

[0065] S120: extracting information from the review information to obtain feature information in the review information.

[0066] In the embodiment of the present application, information extraction is a method for extracting structured information from text data, which can identify entities in text data and the association relationship between entities. By extracting information from the comment information, structured feature information in the comment information can be obtained. Specifically, the feature information here usually exists in the form of a sentence-level proposition of a single sentence, which usually consists of two parts: "subject" and "predicate". Among them, the predicate is a statement or description of the subject, which can answer questions such as "how" and "what" of the subject. That is, in general, the proposition usually exists in the form of "how is XX" and "what is XX". Therefore, by extracting information from the comment information, the entity in the comment information and the descriptive information related to the entity can be extracted, thereby generating corresponding feature information. For example, taking the aforementioned "This is a very inspiring film, and in a certain sense, it is a subjective writing of American history" as an example, by performing information extraction on the review information, two feature information in the form of sentence-level propositions can be extracted from the review text: "This (film) is a very inspiring film" and "This (film) is a subjective writing of American history". Specifically, the implementation scheme of extracting several different feature information from a review information based on information extraction can be referred to in the following Figure 7 and its explanatory contents.

[0067] S130: Generate a comment label for the target object according to the feature information.

[0068] In the embodiment of the present application, after extracting information from the aforementioned review information containing complex semantics to obtain feature information containing only the semantics of a single sentence in the review information, by further performing semantic recognition processing on the feature information, the feature information of the single sentence semantics can be used to generate a corresponding review label for the target object. For example, taking the feature information "This (film and television work) is a very inspirational film" as an example, by performing semantic recognition on the feature information, the target object, that is, the film and television work, can be labeled with a review label of "inspirational". Alternatively, for the feature information "This (film and television work) is a subjective writing of American history", by performing semantic recognition processing on the feature information, the film and television work can be labeled with a review label such as "America", "History", "Documentary", etc.

[0069] Of course, the above scheme only takes a single feature information as an example to illustrate how to generate a comment label for a target object. In fact, considering the different subjective emotions of different users, there will be differences in their views on the same target object. Therefore, if a comment label for the target object is generated for each feature information, there may be many conflicting comment labels. Therefore, as a feasible implementation scheme of this application, a scheme is proposed to comprehensively generate comment labels for the target object based on cluster analysis of many feature information. The specific implementation scheme can be found in the subsequent Figure 2 and its explanatory contents.

[0070] Furthermore, combined with the above-mentioned descriptions, it can be seen that the review information of these works of art, in addition to the intuitive review of the quality of the work itself, often also contains other richer emotional information, such as recording personal life details or mood at that time, etc. These recorded life details are usually also extracted with corresponding feature information. For example, taking the following comment as an example, "This movie reminds me of my college days, when I borrowed discs from the library video room with my friends and watched movies on an 18-inch small TV with headphones." At this time, based on information extraction, it can be obtained that "This movie will make There are two propositions: "I think of my college years" and "My friend and I borrowed a disc in the library video room, and then watched a movie on an 18-inch TV with headphones". It can be seen that the second proposition is closely related to the reviewer's own situation, but has a lower correlation with the film or TV work. Therefore, before processing the feature information to obtain the review label, the proposition can also be tested for relevance to determine whether the obtained feature information is closely related to the target object, and the feature information closely related to the target object can be used to generate a more accurate review label. The specific implementation scheme can be referred to in the subsequent Figure 6 and its explanatory contents.

[0071] In addition, after generating the review tags of the film and television works, usually, based on the review tags, it can also be used to implement search recommendations for users. For specific implementation solutions, please refer to the subsequent Figure 8 and its explanatory contents.

[0072] The label generation method provided in the embodiment of the present application can extract some simple semantic feature information from the review information containing complex semantics by performing information extraction processing on the review information after acquiring the review information of the target object. Therefore, compared with processing the entire review text with complex semantics, in the process of generating the review label of the target object by using these feature information, it can effectively avoid introducing noise or losing the semantic information of the review text, thereby improving the accuracy of the review label finally generated.

[0073] like Figure 2 As shown, Figure 2 A schematic flow chart of steps for generating a comment label based on feature information provided in an embodiment of the present application, specifically, including steps S210 to S230:

[0074] S210: clustering the feature information to obtain a feature information set of the target object.

[0075] Compared to the review text with complex semantics, since the feature information in the embodiments of the present application usually only contains the semantic information of a single sentence, it is convenient to complete the clustering of the feature information based on the semantic information of each feature information. For example, taking the two feature information provided above, "This (film work) is a classic of inspirational movies" and "This (film work) is a very inspirational movie", the semantics represented by these two feature information are that this film work is very inspirational. Therefore, these two feature information can be merged into the same feature information set, and by clustering different feature information extracted from different review information, feature information with the same semantics can be clustered into the same feature information set. Specifically, the clustering of feature information can extract the feature coding of each feature information based on the language model, and then complete the clustering of feature information based on the similarity between the feature coding. For the specific implementation scheme, please refer to the subsequent Figure 3 and its explanatory contents.

[0076] S220: Filter out a target feature information set from the feature information sets according to the amount of feature information in each of the feature information sets.

[0077] In an embodiment of the present application, after obtaining the feature information set of the target object based on the aforementioned clustering processing of the feature information, since the feature information contained in each feature information set is feature information with the same or similar semantics, the number of feature information in the cluster can, to a certain extent, reflect the degree of recognition of different users' cognition of the target object. Under normal circumstances, therefore, the target feature information set can be filtered out from the feature information set according to the number of feature information in each feature information set.

[0078] Specifically, as a common feasible implementation scheme, the generated feature information sets can be sorted according to the size of the feature information in the cluster, and the feature information sets with a smaller number, that is, the ones ranked lower, are discarded, and only the feature information sets with a larger number, that is, the k feature information sets ranked higher are retained as the target feature information sets. Of course, as another feasible implementation scheme, it is also possible to use the feature information set with the number of feature information in the cluster greater than the preset number threshold N as the target feature information set based on the size relationship between the number of feature information in the cluster and the preset number threshold N. Among them, the parameters used in the screening process, such as the number k of the target feature information set, or the number threshold N, can be based on actual needs, or can be pre-set based on the actual size of the review text.

[0079] S230: Generate a comment label for the target object according to the target feature information set.

[0080] In an embodiment of the present application, for each feature information set obtained by clustering, since the semantics of the feature information contained in each feature information set are relatively similar, a comment label for the target object can be obtained by processing the feature information in the feature information set, and by processing different feature information sets, multiple comment labels for the target can be generated.

[0081] Among them, the comment label of the target object generated based on the feature information in the feature information set can also be obtained based on the language model. The specific implementation scheme can be referred to in the subsequent Figure 4 and its explanatory contents.

[0082] like Figure 3 As shown, Figure 3 A schematic flow chart of a step of clustering feature information provided in an embodiment of the present application, specifically, comprising steps S310 to S330:

[0083] S310: Process the feature information based on a first language model to obtain feature codes corresponding to each piece of feature information.

[0084] Normally, for semantic processing of feature information, it is usually necessary to convert the feature information in text form into feature coding, that is, it is necessary to complete the coding of the feature information. Specifically, this step can be implemented in advance through a language model in the field of text processing. For example, as a common feasible implementation scheme, a preset pre-trained language model can be used to complete the processing of feature information and obtain the feature coding corresponding to the feature information. Specifically, the pre-trained language model here can be ELMo (Embedding from Language Models, a pre-trained language model for generating word vectors), GPT (Generative Pre-Trained Transformer, generative pre-trained language model), BERT (Bidirectional Encoder Representations from Transformers, bidirectional encoding pre-trained language model), etc. Of course, based on experimental effect verification, the pre-trained language model here can adopt the ELMo language model.

[0085] S320, normalizing each of the feature codes to obtain a normalized feature code corresponding to each of the feature information.

[0086] In the embodiment of the present application, on the basis of the feature codes corresponding to the feature information obtained above, by further performing standard normalization processing on each feature code, it can be ensured that the obtained normalized feature code is more in line with the standardization and stability. Specifically, the calculation formula for normalizing the feature code is as follows:

[0087]

[0088] Among them, h i That is, the feature code of the i-th feature information, ‖h i ‖ is the modulus of the feature code, and It is the normalized feature code obtained after normalizing the feature code.

[0089] S330: Clustering each of the feature information according to the similarity between the normalized feature codes to obtain a feature information set of the target object.

[0090] In an embodiment of the present application, after the normalized feature codes corresponding to each feature information are obtained, the similarity between the normalized feature codes can describe the similarity between the semantics corresponding to each sentence proposition. Specifically, the similarity between the normalized feature codes can be calculated based on the calculation formula of cosine similarity. The calculation formula of cosine similarity is not described in the embodiment of the present application. On this basis, based on the similarity between the normalized feature codes, the feature information is clustered to obtain a set of feature information of the target object. Specifically, as a feasible embodiment of the present application, the feature information can be clustered by using K-means (k-means clustering algorithm).

[0091] like Figure 4 As shown, Figure 4 A schematic flow chart of steps for generating a comment label of a target object according to a target feature information set provided in an embodiment of the present application, specifically, comprising steps S410 to S420:

[0092] S410: For each target feature information set, input feature information in the target feature information set into a second language model to generate a description text corresponding to the target feature information set.

[0093] In order to realize the fusion of feature information in the feature information set into a complete descriptive text as a comment label for the target object, that is, to realize the fusion of text, the embodiment of the present application also provides an implementation scheme for text fusion based on a pre-trained language model. Specifically, since the pre-trained language model here needs to complete the encoding and decoding of the text to generate the corresponding text. Therefore, the pre-trained language model here is usually required to have a language generation function. For example, as a feasible embodiment of the present application, the second pre-trained language model can adopt BART (Bidirectional and Auto-Regressive Transformers, a text encoding and decoding model with contextual information and autoregressive characteristics). Compared with other language models, BART has stronger semantic analysis capabilities. Even if the input proposition itself is not grammatical or not fluent, BART can still tend to generate a complete and coherent description text.

[0094] Furthermore, considering that BART has the ability to analyze contextual information, that is, in the process of semantic analysis, the order of the input feature information will be considered. Therefore, as a feasible embodiment of the present application, when the feature information in the target feature information set is input into the second pre-trained language model, the feature information can be sorted based on the probability of correlation between the proposition and the target object detected by the aforementioned correlation detection model, and the single proposition can be segmented with special marks. The specific implementation scheme can be referred to in the subsequent Figure 5 and its explanatory contents.

[0095] S420, fusing the description texts corresponding to the target feature information sets to obtain a comment label for the target object.

[0096] After completing the processing of the target feature information set and obtaining the corresponding description text for each target feature information set, the final multiple comment labels of the target object can be obtained by further fusing the description texts corresponding to these target feature information sets.

[0097] like Figure 5 As shown, Figure 5 A schematic flow chart of a step of sorting propositions in a feature information set based on correlation probability provided in an embodiment of the present application, specifically, comprising steps S510 to S520:

[0098] S510: Obtain the correlation probability corresponding to each feature information in the target feature information set.

[0099] In the embodiment of the present application, the correlation probability corresponding to each feature information refers to the correlation probability between the feature information and the target object. Generally, the correlation probability corresponding to the feature information can be obtained by processing the feature information with a pre-trained correlation detection model. It should be noted that the correlation detection model here is the correlation detection model mentioned in the aforementioned step S130. For the implementation scheme of the correlation detection model, please refer to the subsequent Figure 6 and its explanatory contents.

[0100] S520: Based on the magnitude relationship of the correlation probabilities corresponding to the respective feature information, the feature information is sequentially input into a second language model to generate a description text corresponding to the target feature information set.

[0101] After obtaining the correlation probability describing the degree of association between the feature information and the target object through the aforementioned correlation detection model, for each feature information set, the feature information in the feature information set is sorted according to the size relationship of the correlation probability corresponding to each feature information, and then sequentially input into the second pre-trained language model, so that BART's ability to analyze contextual information can be better utilized to output more accurate description text.

[0102] like Figure 6 As shown, Figure 6 A schematic flow chart of a step of performing correlation detection on feature information provided in an embodiment of the present application, specifically, comprising steps S610 to S630:

[0103] S610: Process the feature information according to a correlation detection model to obtain a correlation probability of the feature information corresponding to the target object.

[0104] In the embodiment of the present application, the correlation detection model can be pre-trained by using a binary classification method based on a supervised training method. Specifically, the steps of training the correlation detection model mainly include the following steps:

[0105] (1) Pre-establish a training set, where the training set contains true propositions related to the characteristics of film and television works, such as the aforementioned "This movie reminds me of my college years", and false propositions unrelated to the characteristics of film and television works, such as "My friend and I borrowed a disc from the library video room, and then watched the movie on an 18-inch small TV with headphones." In general, true propositions can be manually labeled as 1, and false propositions can be manually labeled as 0, that is, 1 and 0 are used as the true labels of true propositions and false propositions respectively;

[0106] (2) Building an initial correlation detection model. For example, as a feasible embodiment of the present application, the initial correlation detection model here selects a combination of BERT (Bidirectional Encoder Representation from Transformers, a pre-trained model based on bidirectional encoding) and a fully connected layer;

[0107] (3) Train the initial correlation detection model. Specifically, train a training proposition x in the training set. i =[CLS]s1s2…s l Input to BERT in the initial relevance detection model, where [CLS] is a special marker at the beginning of the proposition, s1~s l are the word vectors of the l words in the proposition, and the contextual semantic representation of the proposition can be obtained H = BERT (x i), then the component H0 corresponding to the [CLS] tag position in the context semantic representation H is input into the fully connected layer in the initial relevance detection model to obtain the probability P = sigmoid(H0W+b) of the proposition corresponding to the true or false proposition, and the probability P∈[0,1], where W and b are the weight parameter and bias parameter to be trained in the fully connected layer, respectively, and sigmoid is the activation function. Then, the cross entropy between the probability P and the true label of this sample (i.e., 0 and 1 in the above) is calculated to obtain the loss value of the model, and the parameters in the model are updated and optimized based on back propagation to obtain an updated relevance detection model, until the loss value obtained by outputting the training proposition to the updated model is less than the preset loss value threshold for several consecutive rounds, and the current relevance detection model is the relevance detection model pre-trained in this proposal.

[0108] Specifically, by processing the feature information based on the correlation detection model obtained through the training, the correlation probability of the feature information corresponding to the target object can be obtained, wherein the closer the correlation probability is to 1, the higher the degree of correlation between the feature information and the target object. Conversely, if the correlation probability is closer to 0, the lower the degree of correlation between the feature information and the target object, the more likely the feature information is an emotional description proposition expressed by the user related to the personal situation.

[0109] S620: Filter out target feature information from the feature information according to the magnitude relationship of the correlation probabilities.

[0110] In an embodiment of the present application, after obtaining the correlation probability describing the degree of association between the proposition and the target object based on the aforementioned correlation detection model, several feature information with higher correlation probabilities can be screened out from the feature information as target feature information based on the size relationship of these correlation probabilities.

[0111] In addition, in addition to the above-mentioned feature information can be screened based on the size relationship of the relevance probability, the relevance probability of the feature information corresponding to the target object can also be further used to determine the sorting order of the feature information in the feature information set in the process of finally generating a description text based on the feature information set, so as to input it into the pre-trained language model. Figure 5 Provided implementation solutions.

[0112] S630: Generate a comment label for the target object according to the target feature information.

[0113] Unlike the aforementioned step S130 in which the comment label of the target object is generated based on the feature information, the embodiment of the present application generates the comment label of the target object based on the target feature information that is more relevant to the target object selected from the feature information. Of course, the specific scheme for generating the comment label can refer to the aforementioned Figure 2 to Figure 5 The solutions provided differ only in the feature information they process.

[0114] like Figure 7 As shown, Figure 7 A schematic flow chart of steps for extracting feature information from review information provided in an embodiment of the present application, specifically, including steps S710 to S720:

[0115] S710: Perform word segmentation processing on the review information according to part-of-speech information of each word in the review information to obtain target words in the review information and associations between the target words.

[0116] In the embodiments of the present application, information extraction generally refers to extracting entity information in the review information and the association relationship between entities. Therefore, the review information can be segmented based on the part-of-speech information of each word in the review information, so as to obtain target words with specific part-of-speech information in the review information, for example, nouns that can be used to refer to the subject in the proposition, or adjectives that can be used to describe the subject in the proposition, etc., as well as the association relationship between these words.

[0117] S720: According to the association relationship between the target words, the target words are merged to obtain feature information of the review information.

[0118] In the embodiment of the present application, after completing the word segmentation processing of the review information and extracting the target words therein, the target words are further fused based on the association relationship between the target words to obtain the feature information of the review information. Specifically, the fusion here usually involves splicing and combining the target words with association relationships. It should be noted that since the BART model with strong semantic analysis capabilities is used in the subsequent process of generating review tags, the feature information obtained by splicing and combining the target words here may not strictly require that it conforms to grammatical habits.

[0119] like Figure 8 As shown, Figure 8 A schematic diagram of a process flow for recommending a target object provided in an embodiment of the present application, specifically, comprising steps S810 to S830:

[0120] S810, obtaining search information of the target user.

[0121] In the embodiment of the present application, after obtaining the comment label of the target object based on the comment information of the target object, the comment label will be stored in association with the target object for subsequent retrieval recommendation. For example, the comment label can be stored in association with the target object in the retrieval system. At this time, the retrieval system can obtain the retrieval information of the target user for subsequent matching recommendation.

[0122] S820: Determine a matching degree between the target object and the search information according to the search information and the comment tag.

[0123] In the embodiment of the present application, the degree of match between the target object and the search information can be determined by calculating the similarity between the search information and the comment tag, or determining whether the comment tag of the target object contains the keyword in the search information input by the user. Specifically, if the similarity between the search information and the comment tag is high, for example, higher than a preset threshold, or the comment tag of the target object contains the keyword in the search information input by the user, then it can be determined that the degree of match between the target object and the search information is high, otherwise it can be determined that the degree of match between the target object and the search information is low.

[0124] S830: Recommend the target object to the target user according to the matching degree.

[0125] In an embodiment of the present application, further, when the degree of match between the target object and the search information is high, it can be considered that the target object is the content that the target user expects to retrieve, and the target object can be recommended to the target user. Conversely, when the degree of match between the target object and the search information is low, it can be considered that the target object is not the content that the target user expects to retrieve, and the target object can be filtered.

[0126] To facilitate understanding of the complete implementation process of a label generation method provided in an embodiment of the present application, the following will be combined with the aforementioned Figures 1 to 8 The embodiment provided provides a complete implementation scheme of a tag generation method. Specifically, Fig. 9 As shown, Fig. 9 A complete implementation flow diagram of a label generation method provided in an embodiment of the present application is provided. Combined with the flow diagram, a complete implementation scheme of the label generation method includes the following steps:

[0127] (1) Proposition extraction: The collected film and television review information is subjected to information extraction processing to obtain feature information in the review information. Specifically, the film and television review information is usually composed of a mixture of multiple propositions. By using the part-of-speech information of each word in the review information, target words of specific parts of speech in the review information are extracted, such as nouns referring to the subject of the proposition or adjectives describing the subject of the proposition, as well as the associations between these target words. Then, based on the associations between the target words, these target words are rearranged and combined, so that some feature information can be extracted from the review information;

[0128] (2) Relevance detection: By using the pre-trained relevance detection model to process each feature information, we can obtain the probability value describing whether the proposition is relevant to the film and television work itself, that is, the relevance probability. When the relevance probability is lower than a certain threshold, the proposition can be classified as irrelevant. After filtering and cleaning irrelevant propositions, we can extract target feature information that is more relevant to the film and television work from the feature information.

[0129] (3) Cluster analysis: The obtained n target feature information is input into the pre-trained language model ELMo model to obtain the feature encoding h1, h2, …h of each target feature information. n , and then use the standard normalization formula to complete the normalization of each feature encoding, and then get the normalized vector of each target feature information Then use the normalized vector The similarity between them is used as the distance value, and the target feature information is clustered based on the K-mean clustering algorithm to obtain several clusters;

[0130] (4) Cluster sorting and fusion: After obtaining several clusters, first sort the clusters according to the amount of target feature information in each cluster, retain a number of clusters with high rankings, or retain clusters with target feature information in the clusters greater than a preset threshold as target clusters for subsequent label generation, and then for each target cluster, arrange the target feature information in the cluster according to the correlation probability corresponding to each target feature information in the cluster, and then sequentially input them into the pre-trained language model BART model to obtain a descriptive text. After performing the above operation on each target cluster, several descriptive texts can be obtained. By summarizing these texts, several comment labels of the film and television work can be obtained;

[0131] (5) Content recommendation: After the review tag is associated with the film or TV work and stored, for example, in a retrieval system, the subsequent retrieval system can determine the matching degree of the film or TV work relative to the retrieval information input by the user and the review tag of the film or TV work, and then determine whether to recommend the film or TV work to the user based on the matching degree.

[0132] In order to better implement the label generation method provided in the embodiment of the present application, on the basis of the label generation method provided in the embodiment of the present application, the embodiment of the present application also provides a label generation device, such as Fig.10 As shown, the label generating device 1000 includes:

[0133] An acquisition module 1010 is used to acquire the review information of the target object;

[0134] An extraction module 1020 is used to extract information from the review information to obtain feature information in the review information;

[0135] The generating module 1030 is used to generate a comment label of the target object according to the feature information.

[0136] In some embodiments of the present application, the generation module 1030 is also used to perform clustering processing on the feature information to obtain a feature information set of the target object; based on the number of feature information in each feature information set, filter out a target feature information set from the feature information set; and generate a comment label for the target object based on the target feature information set.

[0137] In some embodiments of the present application, the generation module 1030 is further used to process the feature information based on the first language model to obtain feature codes corresponding to each piece of feature information; normalize each piece of feature code to obtain a normalized feature code corresponding to each piece of feature information; cluster each piece of feature information according to the similarity between the normalized feature codes to obtain a feature information set of the target object.

[0138] In some embodiments of the present application, the generation module 1030 is also used to input the feature information in each target feature information set into the second language model to generate a description text corresponding to the target feature information set; and fuse the description texts corresponding to each target feature information set to obtain a comment label for the target object.

[0139] In some embodiments of the present application, the generation module 1030 is also used to obtain the correlation probability corresponding to each feature information in the target feature information set; the correlation probability corresponding to the feature information is obtained by processing the feature information with a pre-trained correlation detection model; based on the size relationship of the correlation probabilities corresponding to each feature information, the feature information is sequentially input into the second language model to generate a description text corresponding to the target feature information set.

[0140] In some embodiments of the present application, the generation module 1030 is also used to process the feature information according to a correlation detection model to obtain a correlation probability of the feature information corresponding to the target object; based on the size relationship of the correlation probabilities, filter out target feature information from the feature information; and generate a comment label for the target object based on the target feature information.

[0141] In some embodiments of the present application, the extraction module 1020 is also used to perform word segmentation on the comment information according to the part-of-speech information of each word in the comment information, to obtain target words in the comment information and the association relationship between the target words; based on the association relationship between the target words, the target words are merged to obtain feature information of the comment information.

[0142] In some embodiments of the present application, the generation module 1030 is also used to obtain search information of the target user; determine the matching degree between the target object and the search information based on the search information and the comment label; and recommend the target object to the target user based on the matching degree.

[0143] For the specific definition of the label generation device, please refer to the definition of the label generation method above, which will not be repeated here. Each module in the above-mentioned label generation device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0144] In some embodiments of the present application, the label generating device 1000 may be implemented in the form of a computer program. The computer program may be implemented in a computer program such as Fig.11 The computer device shown in the figure is run. The memory of the computer device can store various program modules constituting the label generating device 1000, for example, Fig.10 The acquisition module 1010, extraction module 1020 and generation module 1030 are shown. The computer program composed of various program modules enables the processor to execute the steps of the label generation method of each embodiment of the present application described in this specification.

[0145] For example, Fig.11 The computer device shown can be Fig.10 The acquisition module 1010 in the label generation device 1000 shown executes step S110. The computer device can execute step S120 through the extraction module 1020. The computer device can execute step S130 through the generation module 1030. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external computer device through a network connection. When the computer program is executed by the processor, a label generation method is implemented.

[0146] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0147] In some embodiments of the present application, a computer device is provided, comprising one or more processors; a memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the following steps:

[0148] Obtain review information of the target object;

[0149] Extracting information from the review information to obtain feature information in the review information;

[0150] A comment label of the target object is generated according to the feature information.

[0151] In some embodiments of the present application, a computer-readable storage medium is provided, which stores a computer program. The computer program is loaded by a processor, so that the processor performs the following steps:

[0152] Obtain review information of the target object;

[0153] Extracting information from the review information to obtain feature information in the review information;

[0154] A comment label of the target object is generated according to the feature information.

[0155] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Any reference to memory, storage, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0156] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The above is a detailed introduction to a label generation method, device, computer equipment and storage medium provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A label generation method, characterized in that: include: Obtain review information of the target object; Extracting information from the review information to obtain feature information in the review information; A comment label of the target object is generated according to the feature information.

2. The label generation method according to claim 1, characterized in that: The step of generating a comment label for the target object according to the feature information includes: Performing clustering processing on the feature information to obtain a feature information set of the target object; Filtering a target feature information set from the feature information sets according to the amount of feature information in each of the feature information sets; A comment label of the target object is generated according to the target feature information set.

3. The label generation method according to claim 2, characterized in that: The clustering process of the feature information to obtain a feature information set of the target object includes: Processing the feature information based on the first language model to obtain feature codes corresponding to each feature information; Normalizing each of the feature codes to obtain a normalized feature code corresponding to each of the feature information; According to the similarity between the normalized feature codes, each feature information is clustered to obtain a feature information set of the target object.

4. The label generation method according to claim 2, characterized in that: Generating a comment label of the target object according to the target feature information set includes: For each of the target feature information sets, input feature information in the target feature information set into a second language model to generate a description text corresponding to the target feature information set; The description texts corresponding to each target feature information set are fused to obtain a comment label of the target object.

5. The label generation method according to claim 4, characterized in that: The step of inputting the feature information in the target feature information set into the second language model to generate a description text corresponding to the target feature information set includes: Obtaining the correlation probability corresponding to each feature information in the target feature information set; the correlation probability corresponding to the feature information is obtained by processing the feature information through a pre-trained correlation detection model; Based on the magnitude relationship of the correlation probabilities corresponding to the respective feature information, the feature information is sequentially input into the second language model to generate a description text corresponding to the target feature information set.

6. The label generation method according to claim 1, characterized in that: Before the step of generating a comment label of the target object according to the feature information, the method further includes: Processing the feature information according to a correlation detection model to obtain a correlation probability of the feature information corresponding to the target object; Filtering target feature information from the feature information according to the magnitude relationship of the correlation probabilities; The step of generating a comment label for the target object according to the feature information includes: A comment label of the target object is generated according to the target feature information.

7. The label generation method according to claim 1, characterized in that: The extracting the review information to obtain feature information in the review information includes: According to the part-of-speech information of each word in the review information, the review information is segmented to obtain target words in the review information and the association relationship between the target words; According to the association relationship between the target words, the target words are fused to obtain feature information of the review information.

8. The label generation method according to any one of claims 1 to 7, characterized in that: After the step of generating a comment label of the target object according to the feature information, the method further includes: Obtain search information of target users; Determining a matching degree between the target object and the search information according to the search information and the comment label; The target object is recommended to the target user according to the matching degree.

9. A label generating device, characterized in that: include: An acquisition module is used to obtain the review information of the target object; An extraction module, used to extract information from the review information to obtain feature information in the review information; A generating module is used to generate a comment label of the target object according to the feature information.

10. A computer device, characterized in that: The computer device comprises: one or more processors; Memory; and One or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the processor to implement the label generation method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program is loaded by a processor to execute the label generation method according to any one of claims 1 to 8.

Images