CN114595370A - Model training and sorting method and device, electronic equipment and storage medium - Google Patents

Abstract

Embodiments of the present disclosure provide a model training and sorting method, apparatus, electronic device, and storage medium. The model training method includes: acquiring sample data, where the sample data includes sample search information and sample keywords corresponding to the sample objects; in a preset to-be-trained model, feature fusion is performed on the sample keywords and the sample search information to obtain the sample keywords. The comprehensive semantic representation vector is used to obtain the sample sorting parameters of the sample object based on the comprehensive semantic representation vector; the trained model determined based on the sample sorting parameters is used as the sorting model. In the embodiment of the present disclosure, the keyword information of the object is integrated into the ranking model. These keywords are excavated according to the unstructured characteristics of the object, which can better cover and describe the user's intention. The ranking model can capture the semantics of the keywords themselves and the semantic correlation between keywords and user search information, thereby improving the accuracy of the ranking model.

Description

Model training, sorting method, device, electronic device and storage medium

technical field

The present disclosure relates to the technical field of data processing, and in particular, to a model training and sorting method, apparatus, electronic device and storage medium.

Background technique

With the rapid development of Internet technology, users need to find the information they need in the sea of information, just like looking for a needle in a haystack, and search engine technology just solves this problem. A search engine is a retrieval technology that uses specific strategies to retrieve specified information and feed it back to users according to user needs and certain algorithms.

During the search process, multiple objects are first recalled based on the search information input by the user, and then the recalled objects are sorted. In the sorting process, a sorting model is usually used to analyze the recalled objects to obtain sorting parameters, and the sorting parameters are used to sort the recalled objects.

In the prior art, the ranking model is usually modeled by a specific network structure, and is mainly used to analyze the correlation between the search information input by the user and the characteristics of the object itself. However, because the characteristics of the object itself cannot accurately reflect the important characteristics of the object, the analysis accuracy of the ranking model in this method is poor, and the model effect is poor.

SUMMARY OF THE INVENTION

In view of the above problems, the embodiments of the present disclosure propose a model training and sorting method, apparatus, electronic device, and storage medium, so as to improve the accuracy of the sorting model.

According to a first aspect of the embodiments of the present disclosure, a model training method is provided, including: obtaining sample data; the sample data includes sample search information and sample keywords corresponding to sample objects; in a preset model to be trained , perform feature fusion on the sample keywords and the sample search information to obtain a comprehensive semantic representation vector of the sample keywords, and obtain the sample sorting parameters of the sample objects based on the comprehensive semantic representation vector; The sample ranking parameter determines the trained model as the ranking model.

Optionally, performing feature fusion on the sample keywords and the sample search information includes: for each sample keyword, performing feature fusion on the current sample keyword and other sample keywords to obtain a fusion of the current sample keywords. Semantic representation vector; for each sample keyword, feature fusion of the current sample keyword and the sample search information to obtain the correlation between the current sample keyword and the sample search information; based on the fusion semantic representation of each sample keyword vector and relevance, and calculate the comprehensive semantic representation vector of the sample keywords.

Optionally, the feature fusion of the sample keywords and the sample search information includes: querying the feature vector of each sample keyword and the sample search information from a preset correspondence between words and feature vectors. Feature vector; for each sample keyword, perform feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords to obtain the fusion semantic representation vector of the current sample keyword; The fused semantic representation vector of the sample keyword and the feature vector of the sample search information perform feature fusion to obtain the correlation between the current sample keyword and the sample search information; based on the fusion semantic representation vector of each sample keyword and the correlation degree, and the feature vector of the sample search information, and calculate the comprehensive semantic representation vector of the sample keyword.

Optionally, feature fusion is performed on the feature vector of the current sample keyword and the feature vector of other sample keywords, including: performing feature fusion on the feature vector of the current sample keyword and the feature vector of other sample keywords through a self-attention mechanism. , to obtain the fusion semantic representation vector of the current sample keywords.

Optionally, the sample data further includes the confidence level of the sample keyword; the feature fusion of the feature vector of the current sample keyword and the feature vectors of other sample keywords includes: using a self-attention mechanism to analyze the current sample keywords. The feature vector of the word is fused with the feature vectors of other sample keywords to obtain the preliminary fusion semantic representation vector of the current sample keyword; the sample keywords are sorted in descending order according to the confidence, and the key of each sample is obtained based on the sorting result. The position embedding vector of the word; the initial fusion semantic representation vector of the current sample keyword and the position embedding vector are added to obtain the fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further includes the confidence level of the sample keyword; performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information, including: calculating the Confidence and the relevance weight of the search information; based on the fusion semantic representation vector of the current sample keyword, the feature vector of the sample search information, and the relevance weight of the current sample keyword, calculate the intermediate parameter of the current sample keyword ; Calculate the correlation between the current sample keyword and the sample search information based on the intermediate parameter and the preset temperature parameter.

Optionally, based on the fusion semantic representation vector and correlation of each sample keyword, and the feature vector of the sample search information, calculating the comprehensive semantic representation vector of the sample keyword, including: calculating the fusion semantics of each sample keyword. The sum of the product of the representation vector and the correlation degree is obtained to obtain the preliminary semantic representation vector of the sample keyword; the preliminary semantic representation vector and the feature vector of the sample search information are added to obtain the preliminary synthesis of the sample keyword Semantic representation vector; standardize the preliminary comprehensive semantic representation vector to obtain the comprehensive semantic representation vector of the sample keyword.

Optionally, after calculating the sum of the products of the fusion semantic representation vector of each sample keyword and the correlation degree to obtain the preliminary semantic representation vector of the sample keyword, the method further includes: randomly discarding the preliminary semantic representation vector. ; adding the preliminary semantic representation vector to the feature vector of the sample search information, comprising: adding the preliminary semantic representation vector after the random discarding process to the feature vector of the sample search information.

Optionally, the sample data further includes sample description information corresponding to the sample object; obtaining the sample sorting parameters of the sample object based on the comprehensive semantic representation vector, including: Feature fusion is performed on the description information and the sample search information to obtain sample sorting parameters of the sample object.

Optionally, performing feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information includes: performing a deep analysis on the comprehensive semantic representation vector, the sample description information, and the sample search information. feature fusion to obtain the first sample sorting parameter of the sample object; perform shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain the second sample of the sample object Sorting parameter; based on the first sample sorting parameter and the first sample sorting parameter, calculate the sample sorting parameter of the sample object.

Optionally, deep feature fusion is performed on the comprehensive semantic representation vector, the sample description information and the sample search information, including: acquiring the feature vector of the sample description information and the feature vector of the sample search information; based on The integrated semantic representation vector, the feature vector of the sample description information, and the feature vector of the sample search information generate a splicing feature vector; use a preset deep fusion network to perform feature fusion processing on the splicing feature vector to obtain the The first sample sorting parameter.

Optionally, performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information, including: acquiring a feature vector of the sample description information and a feature vector of the sample search information; A preset shallow fusion network is used to perform feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information, and the comprehensive semantic representation vector to obtain the second sample ranking parameter.

According to a second aspect of the embodiments of the present disclosure, there is provided a sorting method, comprising: acquiring search information and keywords corresponding to objects to be sorted; inputting the search information and the keywords into a pre-trained sorting model to obtain The sorting parameters of the objects to be sorted output by the sorting model; the sorting model is obtained by the model training method described in any one of the above; the objects to be sorted are sorted based on the sorting parameters.

According to a third aspect of the embodiments of the present disclosure, a model training apparatus is provided, including: a first acquisition module for acquiring sample data; the sample data includes sample search information and sample keywords corresponding to sample objects; training The module is used to perform feature fusion on the sample keywords and the sample search information in the preset model to be trained, to obtain a comprehensive semantic representation vector of the sample keywords, and obtain the comprehensive semantic representation vector based on the comprehensive semantic representation vector. The sample sorting parameter of the sample object; the determining module is configured to determine the model that has been trained based on the sample sorting parameter as the sorting model.

Optionally, the training module includes: a first fusion unit, configured to perform feature fusion on the current sample keyword and other sample keywords for each sample keyword, to obtain a fusion semantic representation vector of the current sample keyword; The second fusion unit is used to perform feature fusion on the current sample keyword and the sample search information for each sample keyword to obtain the correlation between the current sample keyword and the sample search information; the first calculation unit is used for Based on the fusion semantic representation vector and relevance of each sample keyword, the comprehensive semantic representation vector of the sample keyword is calculated.

Optionally, the training module includes: a query unit for querying the feature vector of each sample keyword and the feature vector of the sample search information from the preset correspondence between words and feature vectors; a third fusion unit , which is used for feature fusion of the feature vector of the current sample keyword and the feature vectors of other sample keywords for each sample keyword to obtain the fusion semantic representation vector of the current sample keyword; the fourth fusion unit is used for each sample keyword. a sample keyword, performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information to obtain the correlation between the current sample keyword and the sample search information; the second computing unit, It is used to calculate the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation degree of each sample keyword and the feature vector of the sample search information.

Optionally, the third fusion unit is specifically configured to perform feature fusion between the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data also includes the confidence level of the sample keywords; the third fusion unit is specifically used to analyze the feature vector of the current sample keyword and the feature vector of other sample keywords through a self-attention mechanism. Perform feature fusion to obtain a preliminary fusion semantic representation vector of the current sample keywords; sort the sample keywords in descending order according to the confidence, and obtain the position embedding vector of each sample keyword based on the sorting result; The initial fusion semantic representation vector and the position embedding vector are added to obtain the fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further includes the confidence of the sample keywords; the fourth fusion unit is specifically configured to calculate the confidence of the current sample keywords and the correlation weight of the search information; based on the current sample The fusion semantic representation vector of the keyword, the feature vector of the sample search information, and the relevance weight of the current sample keyword, calculate the intermediate parameter of the current sample keyword; based on the intermediate parameter and the preset temperature parameter, calculate the current The correlation between the sample keywords and the sample search information.

Optionally, the second calculation unit is specifically configured to calculate the sum of the product of the fusion semantic representation vector and the correlation degree of each sample keyword, and obtain the preliminary semantic representation vector of the sample keyword; The vector is added with the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword; the preliminary comprehensive semantic representation vector is standardized to obtain a comprehensive semantic representation vector of the sample keyword.

Optionally, the second computing unit is further configured to calculate the sum of the product of the fusion semantic representation vector of each sample keyword and the correlation degree, and obtain the preliminary semantic representation vector of the sample keyword. The semantic representation vector is randomly discarded; the second computing unit is specifically configured to add the preliminary semantic representation vector after the random discard process and the feature vector of the sample search information.

Optionally, the sample data further includes sample description information corresponding to the sample object; the training module is specifically configured to perform feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information. , obtain the sample sorting parameters of the sample object.

Optionally, the training module includes: a fifth fusion unit, configured to perform deep feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information, to obtain the first image of the sample object. this sorting parameter; a sixth fusion unit, configured to perform shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain the second sample sorting parameter of the sample object; the third A calculation unit, configured to calculate the sample sorting parameter of the sample object based on the first sample sorting parameter and the first sample sorting parameter.

Optionally, the fifth fusion unit is specifically configured to obtain the feature vector of the sample description information and the feature vector of the sample search information; based on the comprehensive semantic representation vector, the feature vector of the sample description information and A splicing feature vector is generated from the feature vector of the sample search information; the feature fusion process is performed on the splicing feature vector by using a preset deep fusion network to obtain the first sample sorting parameter.

Optionally, the sixth fusion unit is specifically configured to obtain the feature vector of the sample description information and the feature vector of the sample search information; use a preset shallow fusion network to compose the feature vector of the sample description information. and performing feature fusion processing on the feature vector of the sample search information and the comprehensive semantic representation vector to obtain the second sample sorting parameter.

According to a fourth aspect of the embodiments of the present disclosure, there is provided a sorting apparatus, comprising: a second acquisition module for acquiring search information and keywords corresponding to objects to be sorted; a prediction module for combining the search information with keywords The keywords are input into a pre-trained sorting model, and the sorting parameters of the objects to be sorted output by the sorting model are obtained; the sorting model is obtained by the model training method described in any one of the above; the sorting module is used based on The sorting parameter sorts the objects to be sorted.

According to a fifth aspect of embodiments of the present disclosure, there is provided an electronic device, comprising: one or more processors; and one or more computer-readable storage media having instructions stored thereon; When executed by the one or more processors, the processors are caused to execute the model training method described in any one of the above, or execute the sorting method described in any one of the above.

According to a sixth aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, causes the processor to execute the above-mentioned method. The model training method described above, or, perform the ranking method described in any of the above.

Embodiments of the present disclosure provide a model training and sorting method, apparatus, electronic device, and storage medium. During the model training process, sample data is obtained, and the sample data includes sample search information and sample keywords corresponding to the sample objects; in the preset model to be trained, the sample keywords and the sample search information are characterized. Fusion to obtain a comprehensive semantic representation vector of the sample keywords, and obtaining sample sorting parameters of the sample object based on the comprehensive semantic representation vector; and determining a trained model based on the sample sorting parameters as a sorting model. It can be seen that, in the embodiment of the present disclosure, the keyword information of the object is integrated into the ranking model, and these keywords are mined for the unstructured features of the objects. Compared with the unstructured features of the objects, the keywords can be To better refine and summarize the subject of the object, it can cover a wide range of users’ intentions. Therefore, the ranking model incorporating keyword information can capture the semantics of the keywords themselves and the semantic correlation between keywords and user search information, thereby improving the ranking model. accuracy.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present disclosure more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments of the present disclosure. Obviously, the accompanying drawings in the following description are only the implementation of the present disclosure. For some of the drawings in the example, for those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

FIG. 1 is a flowchart of steps of a model training method according to an embodiment of the present disclosure.

FIG. 2 is a schematic diagram of a keyword according to an embodiment of the present disclosure.

FIG. 3 is a flowchart of steps of a sorting method according to an embodiment of the present disclosure.

FIG. 4 is a flowchart of steps of another model training method according to an embodiment of the present disclosure.

FIG. 5 is a flowchart of steps of another sorting method according to an embodiment of the present disclosure.

FIG. 6 is a schematic diagram of an overall processing process according to an embodiment of the present disclosure.

FIG. 7 is a schematic diagram of a semantic encoding network according to an embodiment of the present disclosure.

FIG. 8 is a schematic diagram of a feature interaction network according to an embodiment of the present disclosure.

FIG. 9 is a structural block diagram of a model training apparatus according to an embodiment of the present disclosure.

FIG. 10 is a structural block diagram of a sorting apparatus according to an embodiment of the present disclosure.

Detailed ways

The technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than the whole of the present disclosure. Example. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present disclosure.

The ranking model of the embodiment of the present disclosure can be applied to the ranking link in the search scenario, the ranking link in the recommendation scenario, and so on. By integrating the keyword information of the object into the ranking model, compared with the ranking model that only analyzes the unstructured features of the object itself, the ranking model incorporating the keyword information can capture the semantics of the keywords themselves, as well as the keywords and users. Search for semantic correlations between information, thereby improving the accuracy of ranking models.

Referring to FIG. 1 , a flowchart of steps of a model training method according to an embodiment of the present disclosure is shown.

As shown in Figure 1, the model training method can include the following steps:

Step 101, obtain sample data.

In this embodiment, the sample data may include sample search information and sample keywords corresponding to the sample objects, and the sample data may also include actual sorting parameters of the sample objects.

Sample objects may include, but are not limited to, content objects and the like. Content objects may include, but are not limited to: documents, web pages, pictures, videos, audios, and so on.

The sample search information may include, but is not limited to, search information input by a user, and the like. The form of search information may include, but is not limited to: text form, voice form, picture form, and so on.

Sample keywords refer to words or phrases mined from the relevant information of the sample objects that can better refine and summarize the subject matter of the sample objects, and each sample object may have at least one keyword. Referring to FIG. 2 , a schematic diagram of a keyword according to an embodiment of the present disclosure is shown. Figure 2 shows 4 objects, among which, the keywords of the first object include "Old Shanghai Style Street" and "a good place to go shopping", and the keywords of the second object include "Shanghai Bookstore" and "Nice Tourism" Strategy”, the keywords of the third object include “Shanghai Celebrity Street” and “Old Western-style House”, and the keywords of the fourth object include “Shanghai Travel Strategy” and “must punch attractions”.

The actual sorting parameter refers to the actual parameter related to the sorting of the sample objects. Actual parameters can include, but are not limited to: click-through rate, conversion rate, exposure, etc.

Step 102: In the preset model to be trained, feature fusion is performed on the sample keywords and the sample search information to obtain a comprehensive semantic representation vector of the sample keywords, and the comprehensive semantic representation vector is obtained based on the comprehensive semantic representation vector. Sample ordering parameters for the sample object.

Input the sample search information and sample keywords corresponding to the sample object into the preset to-be-trained model. In the model to be trained, a comprehensive semantic representation vector of the sample keywords is obtained by feature fusion of the sample keywords and the sample search information. The comprehensive semantic representation vector of sample keywords can capture the user's search intent and improve the relevance of the search. Therefore, the sample sorting parameters obtained based on the comprehensive semantic representation vector of sample keywords can more accurately reflect the sorting characteristics of sample objects. .

Step 103 , determining the trained model based on the sample sorting parameter as the sorting model.

The model to be trained obtains the sample sorting parameters of the sample objects based on the comprehensive semantic representation vector, and outputs the sample sorting parameters of the sample objects. Whether the training is complete can be determined based on the sample sorting parameters and the actual sorting parameters of the sample object.

Optionally, the loss function can be calculated based on the sample ranking parameters and the actual ranking parameters of the sample object. The loss function is used to measure the degree of inconsistency between the predicted value of the model and the true value. If the loss function is small, it indicates that the model is very close to the real distribution of the data, and the model has good performance; if the loss function is large, it indicates that the model is far different from the real distribution of the data, and the model has poor performance. The task of training a model is to use an optimization method to find the loss function that minimizes the corresponding model parameters. Therefore, when the loss function satisfies a preset condition (for example, the loss function is smaller than a certain threshold), it can be determined that the training is completed. Optionally, the loss function may include, but is not limited to, a cross-entropy loss function, an exponential loss function, a Dice loss function, a cross-union loss function, and the like.

In response to determining that the training is not completed, the parameters of the model to be trained can be updated, and the training can be continued until the training is completed. In response to determining that the training is complete, the trained model is used as the ranking model.

Based on the sorting model trained by the model training method shown in FIG. 1 above, the objects to be sorted can be sorted.

Referring to FIG. 3 , a flowchart of steps of a sorting method according to an embodiment of the present disclosure is shown.

As shown in Figure 3, the sorting method may include the following steps:

Step 301: Obtain search information and keywords corresponding to the objects to be sorted.

Objects to be sorted refer to objects that need to be sorted. For example, multiple objects recalled in the recall link can be used as objects to be sorted.

The search information corresponding to the objects to be sorted may include, but is not limited to, search information input by the user, and the like. The form of search information may include, but is not limited to: text form, voice form, picture form, and so on. The search information corresponding to the multiple objects to be sorted may be the same, for example, the search information corresponding to the multiple objects to be sorted recalled based on the same search information is the same.

The keywords corresponding to the objects to be sorted refer to words or phrases mined from the relevant information of the objects to be sorted, which can better refine and summarize the subject of the objects to be sorted, and each object to be sorted may have at least one keyword. For the specific process of mining keywords, relevant processing may be performed according to actual experience, and this embodiment will not discuss in detail here.

Step 302: Input the search information and the keywords into a pre-trained sorting model, and obtain the sorting parameters of the objects to be sorted output by the sorting model.

The search information and keywords corresponding to the objects to be sorted are input into the pre-trained sorting model. In the sorting model, the search information and keywords corresponding to the objects to be sorted are feature-fused to obtain the comprehensive semantic representation vector of the keywords. The comprehensive semantic representation vector obtains the sorting parameters of the objects to be sorted, and the sorting model outputs the sorting parameters of the objects to be sorted. The sorting parameters of the objects to be sorted may include but are not limited to: click-through rate, conversion rate, exposure, and so on.

Step 303: Sort the objects to be sorted based on the sorting parameters.

Based on the sorting parameters of the objects to be sorted, the objects to be sorted can be sorted according to preset rules. For example, mathematical operations (such as weighted calculation, average calculation, etc.) are performed based on the sorting parameters of the objects to be sorted to obtain comprehensive sorting parameters of the objects to be sorted, and the objects to be sorted are sorted based on the comprehensive sorting parameters. For example, the larger the comprehensive sorting parameter, the higher the sorting order, and so on.

In the embodiment of the present disclosure, the keyword information of the sample objects is integrated into the ranking model. These keywords are mined for the unstructured features of the objects. Compared with the unstructured features of the objects, the keywords can better Refining and summarizing the subject of the object can cover and describe a wide range of users’ intentions. Therefore, the ranking model incorporating keyword information can capture the semantics of the keywords themselves and the semantic correlation between keywords and user search information, thereby improving the ranking model. accuracy.

Referring to FIG. 4 , a flowchart of steps of another model training method according to an embodiment of the present disclosure is shown.

As shown in Figure 4, the model training method can include the following steps:

Step 401, obtain sample data.

In this embodiment, the sample data may include sample search information, sample keywords, and sample description information corresponding to the sample objects, and the sample data may also include actual sorting parameters of the sample objects.

For sample objects, sample search information, sample keywords and actual sorting parameters, reference may be made to the relevant description of the above step 101 .

The sample description information may include, but is not limited to: the own characteristics of the sample object (such as document characteristics of documents, picture characteristics of pictures, etc.), context characteristics, user characteristics, and the like. Among them, self-features may include but are not limited to: title, logo, author, and so on. Contextual features may include, but are not limited to: time, location, and the like. User characteristics may include, but are not limited to, the user's age, gender, preferences, and the like.

Step 402, in the preset to-be-trained model, perform feature fusion on the sample keyword and the sample search information to obtain a comprehensive semantic representation vector of the sample keyword, and compare the comprehensive semantic representation vector and the sample search information. The sample description information and the sample search information are feature-fused to obtain sample sorting parameters of the sample object.

Input the sample search information, sample keywords and sample description information corresponding to the sample object into the preset to-be-trained model. In the model to be trained, a comprehensive semantic representation vector of the sample keywords is obtained by feature fusion of the sample keywords and the sample search information. The comprehensive semantic representation vector of sample keywords can capture the user's search intent and improve search relevance. Through feature fusion of the comprehensive semantic representation vector of the sample keywords, the sample description information and the sample search information, the sample sorting parameters of the sample object can be obtained, which can conduct sufficient feature interaction and improve the search relevance and result quality.

Step 403: Determine the trained model based on the sample sorting parameter as the sorting model.

For the specific process of step 403, reference may be made to the relevant description of the above-mentioned step 103.

Based on the sorting model trained by the model training method shown in FIG. 4, the objects to be sorted can be sorted.

Referring to FIG. 5 , a flowchart of steps of another sorting method according to an embodiment of the present disclosure is shown.

As shown in Figure 5, the sorting method may include the following steps:

Step 501: Obtain search information, keywords and description information corresponding to the objects to be sorted.

For the objects to be sorted, search information and keywords, refer to the relevant description of the above step 301 .

The description information of the objects to be sorted may include, but is not limited to: the own characteristics of the objects to be sorted (such as document characteristics of documents, picture characteristics of pictures, etc.), context characteristics, user characteristics, and so on.

Step 502: Input the search information, the keyword and the description information into a pre-trained sorting model, and obtain sorting parameters of the objects to be sorted output by the sorting model.

Input the search information, keywords and description information corresponding to the objects to be sorted into the pre-trained sorting model. In this sorting model, the keywords corresponding to the objects to be sorted and the sample search information are feature-fused to obtain the comprehensive semantic representation vector of the keywords, and the comprehensive semantic representation vector, description information and search information of the keywords are feature-fused to obtain the objects to be sorted. The sorting parameters of the sorting model output the sorting parameters of the objects to be sorted.

Step 503: Sort the objects to be sorted based on the sorting parameters.

For the specific process of step 503, reference may be made to the relevant description of the above-mentioned step 303.

In the embodiment of the present disclosure, the keyword of the object is introduced, and a set of general ranking framework is proposed, which integrates the interaction relationship between the semantic information of the keyword, the search information and the object description information. This framework can improve the expressive ability of keyword semantic vector, and further fully interact with important features in the scene, so as to improve the relevance of search and the quality of results.

In the following, content search (specifically, document search) in a search scenario is used as an example for description. However, in practical applications, it is not limited to this, and the sorting process of any object in any scene can be processed.

In the search scenario, the sorting method of fusion keywords refers to integrating the keyword information of the object into the sorting model, so that the sorting model can capture the semantics of the keywords themselves and the semantic correlation between keywords and user search information, thereby Improve search relevance and result quality.

Content search (specifically, document search) is an important module of comment search, which undertakes the important mission of ecological construction of comment content. Currently, the content understanding process produces high-coverage keywords (words or phrases), which are mined from document-related text, compared to unstructured documents. Text information, keywords can better refine and summarize the topic of the document, and each document will be marked with several keyword tags. How to implement the keywords of these documents into the sorting scenario of content search, mine the correlation between keywords and user search information (also known as search terms), capture users' search intentions and influence the distribution of search data, is the basic The problem to be solved by the disclosed embodiments.

In the following, the introduction will be based on the use of keywords in ranking and the correlation modeling method in the ranking model.

In the search scenario, according to the way the keyword tags are used in sorting, they can be divided into the following two categories:

①Explicit feature modeling

a. Single-dimensional features of keywords: that is, use the features of the keywords themselves as additional input features of the model, such as TF-IDF (Term Frequency-Inverse Document Frequency, term frequency-inverse text frequency) statistical features or pre-training vectors, etc. .

②Implicit semantic modeling

a. Single-dimensional implicit semantic representation of keywords: take keywords as a whole and map them to unique identifiers, and initialize word vectors randomly; or treat keywords as sequences of words, initialize word vectors randomly, and use a bag-of-words model or sequence The model is used to aggregate the semantic representation vector of the keyword, and finally the semantic representation vector of the keyword is learned end-to-end through the model.

b. Based on the pre-trained model: The pre-trained model is used as a feature extractor for keyword labels, and then combined with the ranking model to perform fine-tuning in the ranking task.

In search scenarios, methods for modeling search relevance in the ranking phase can be divided into the following two categories:

① Mining correlation features

Mining explicit intersection features between the document's text and user search information, such as features such as the edit distance between the document's title or body and search information.

②Modeling correlation in model structure

In search scenarios, relevance ranking methods can model the relevance between document text and user search information through a specific network structure. E.g:

a. Introduce the double-tower structure into the sorting model, that is, add two network sub-tower structures to the sorting model, the text and search information of the document are respectively obtained through the double-tower to obtain the representation vector, and then the feature interaction is performed through the dot product, and finally the text and The representation or dot product result of the search information can be used as the input of the ranking model together with the representation vector of ranking other features for deep fusion.

b. Use a sequence representation model, such as LSTM (Long Short-Term Memory) or Transformers (transformers), to stitch together the text and search information of the document to capture the correlation between the text and search information.

c. Introduce a pre-training model, introduce the BERT (Bidirectional Encoder Representations from Transformers) encoder into the sorting model, encode the text or search information of the document, and fine-tune in the downstream task of sorting .

However, if the above method is simply adopted, there will be the following problems:

1. The above-mentioned sorting method of fused keywords cannot effectively capture the deep semantic information of keywords. For example: when the keyword itself is used as the input feature, if the method used is to mine the literally explicit features such as the edit distance between the keyword itself and the search information, then when the keyword and the search information are literally inconsistent, but semantically There are synonyms, similarities, and hyponyms, etc., and such explicit relationship features cannot capture the semantic correlation between keywords and search information. When modeling the latent semantics of keywords, if random initialization and end-to-end learning are adopted, it may be difficult for the ranking model to learn the latent semantic representation of the keywords themselves due to the sparse user behavior in the training samples, namely: The representation ability of the learned representation vector is poor.

2. The keywords themselves have noise. If the keywords are integrated alone, the noise of the keywords themselves will be ignored, resulting in poor robustness of the model. Usually, for a document, each keyword has a confidence score, which measures how confident it represents the document. Some keywords have low confidence, and direct introduction will bring noise to the model. Even if some keywords have a high degree of confidence, under a certain search information, the keyword itself and the search information may be irrelevant. Directly introducing the keyword is also important for the model to capture the semantic correlation between the search information and the document. noise.

3. The above ranking correlation modeling methods are mainly based on the correlation between the unstructured text of the document and the search information, lack of structured keywords that can model the document, and discover the relationship between the user's search intent and the document topic search ranking framework. The keywords of the document are the extraction and generalization of the document topic, which can cover and describe the user's extensive search intent in the search scenario. If fully utilized in ranking models, it is crucial to improve the relevance of search results.

In order to better integrate keywords into the ranking model, thereby improving the relevance and quality of search, the following two optimizations are carried out in this embodiment for the above method:

1. This embodiment proposes a deep semantic coding network that integrates keywords and user search information, which can effectively mine the semantic information of keywords, automatically filter noise information, and improve the accuracy and robustness of semantic representation.

2. This embodiment introduces keywords, and proposes a set of general sorting processes and frameworks based on pre-training semantic vectors, fusing keyword semantic information, modeling tags, and interaction between important features of search scenarios. The framework improves the expression ability of keyword semantic vectors through pre-training vectors, and further conducts sufficient feature interaction with important context, document, user and other features in the search scene to improve the relevance of search and the quality of results.

Referring to FIG. 6 , a schematic diagram of an overall processing procedure of an embodiment of the present disclosure is shown. Figure 6 proposes a general ranking framework based on keywords and pre-trained vectors, which can effectively integrate keywords into the ranking model, extract the interaction between semantic vectors and modeling features. As shown in Figure 6, the overall processing process may include two parts: the preprocessing flow of keywords and pretrained vectors, and the sorting model. The following are introduced separately.

1. The preprocessing process of keywords and pre-training vectors.

First, mine keywords and train pre-trained models from massive corpora; then, build keyword dictionaries based on online exposure log data and mined keywords; then, use pre-trained models to reason and extract keywords Pretrained vector. The specific process is as follows:

(1) Training of the pre-trained model. The pre-trained model is obtained by training based on massive corpus. The pre-training model can include the BERT model, etc. The BERT model is a vertical domain pre-training model based on the Transformers structure, and performs well in tasks such as single sentence assignment, sequence labeling, and inter-sentence relationships.

(2) Keyword mining. Using unsupervised methods, such as dependency syntax analysis, TF-IDF, text clustering and other methods, the massive corpus is mined for keywords, and a large number of keywords are obtained. The mined keywords can well measure the theme of the content. After the keyword is obtained by mining, the confidence level of the keyword can also be obtained.

(3) Construction of keyword dictionary. Based on online content search exposure log data for statistics, low-frequency filtering, etc., from the exposed content-related keywords and user search information, filter low-frequency words, count and build a keyword dictionary (including keywords and search information).

(4) Generation of keyword pre-training vectors. Using the above-mentioned pre-training model, the pre-training vector (that is, the feature vector) of each word in the keyword dictionary is extracted, and the keyword pre-training vector table (that is, the correspondence between the word and the feature vector) is obtained. The pre-training vector contains rich semantic information, which can well describe the semantic correlation between different keywords or search information. This vector is the average pooling result of the token sequence representation vectors in the word. Compared with the [CLS]token vector, the vector obtained by average pooling has more complete semantic features.

Second, the sorting model.

Ranking models can include semantic encoding networks and feature interaction networks. On the one hand, the semantic coding network is used to extract the comprehensive semantic representation vector of keywords; on the other hand, the feature interaction network is used for feature fusion, and finally the predicted value of the model is output for offline training and online reasoning. The specific process details are as follows:

(1) Perform keyword semantic encoding based on the keyword pre-training vector table, and obtain the comprehensive semantic representation vector of the keyword.

Through the semantic coding network, the feature fusion of keywords and search information is carried out, and the comprehensive semantic representation vector of keywords is obtained.

Referring to FIG. 7 , a schematic diagram of a semantic coding network according to an embodiment of the present disclosure is shown. As shown in Figure 7, the semantic coding network may include an Input Layer, an Embedding Layer, a Self-attention Layer, a Confidence-Aware Aggregator and an output layer. Layer (Output Layer).

Input layer: The input of the input layer is keywords and search information. As shown in FIG. 7 , the keywords include keyword ₁ , keyword ₂ , . . . , keyword _n , and the search information is query.

关键词keyword₂的特征向量为

…，关键词keyword_n的特征向量为

搜索信息query的特征向量为e_q。Embedding layer: The embedding layer is used to query the feature vector of each keyword and the feature vector of search information based on the correspondence between words and feature vectors. As shown in Figure 7, the feature vector of the keyword keyword ₁ is

The feature vector of the keyword keyword ₂ is

..., the feature vector of the keyword keyword _n is

The feature vector of the search information query is e _q .

来初始化表征向量参数矩阵E，该参数矩阵是可学习的，会在排序模型训练时进行微调。Based on the keyword pre-training vector table (that is, the correspondence between words and feature vectors), the parameters of the embedding layer of the semantic coding network are initialized, and all related features share the embedding layer. That is, the parameter space is shared by keyword features and user search information features, which can significantly reduce the complexity and prevent overfitting. Compared with random initialization, initialization using pre-trained feature vectors can ensure that the model can effectively capture and focus on text-level relevance features from the beginning, thereby affecting the ranking learning process. Specifically, using pretrained feature vectors

to initialize the representation vector parameter matrix E, which is learnable and fine-tuned when the ranking model is trained.

Self-attention layer: used to discover the relationship between keywords.

There may be multiple keywords in a document, and there is a synergistic relationship between different keywords, such as "basketball", "training", it is difficult to fully describe the theme of this document by looking at any keyword alone. Only a few keywords can characterize the subject of the document, and can infer what kind of search information is the most appropriate to display the results of the document to the user. Clearly, when a user searches for "basketball training", this document is highly relevant and meets the user's needs. Therefore, the relationship between keywords can be mined to help the model better capture the relationship between each keyword and user search information.

Taking offline training as an example, for each sample keyword, the self-attention layer performs feature fusion of the current sample keyword and other sample keywords, and obtains the fusion semantic representation vector of the current sample keyword.

Optionally, for each sample keyword, the self-attention layer performs feature fusion on the feature vector of the current sample keyword and the feature vector of other sample keywords through the self-attention mechanism to obtain the fusion semantic representation vector of the current sample keyword. .

The relationship between keywords is mined through the self-attention mechanism, and the fused keyword semantic representation vector is obtained, so that each keyword can integrate the semantic information of other keywords.

The formula is expressed as formula 1:

表示样本关键词的特征向量，

表示样本关键词的融合语义表征向量。该种方式对应于图7所示的情况。In formula 1, Self-Attention() represents the calculation process of the self-attention mechanism,

the feature vector representing the sample keywords,

A fused semantic representation vector representing sample keywords. This approach corresponds to the situation shown in FIG. 7 .

Optionally, the self-attention layer performs feature fusion on the feature vector of the current sample keyword and the feature vector of other sample keywords through the self-attention mechanism to obtain the preliminary fusion semantic representation vector of the current sample keyword; Confidence is sorted in descending order of the sample keywords, and the position embedding vector of each sample keyword is obtained based on the sorting result; the preliminary fusion semantic representation vector of the current sample keyword and the position embedding vector are added to obtain the fusion of the current sample keywords. Semantic representation vector.

接着，通过自注意力机制来挖掘关键词与关键词之间关系，得到融合后的关键词语义表征序列，使得每个关键词都能够融入其它关键词的语义信息。For a certain document i, the confidence level of each keyword k is different. Firstly, the keywords are arranged in descending order according to the confidence level to form the keyword sequence {k ₁ ,k ₂ ,...,k _n }, and then the initial characterization sequence of keywords is formed by looking up the table, namely:

Then, the relationship between keywords and keywords is mined through the self-attention mechanism, and the fused keyword semantic representation sequence is obtained, so that each keyword can be integrated into the semantic information of other keywords.

The formulas are expressed as formula 2 and formula 3:

表示样本关键词的特征向量，

表示样本关键词的初步融合语义表征向量。公式三中，p_i表示置信度排名第i的样本关键词k_i的位置嵌入向量，p_i能够从全局层面捕捉不同排名的样本关键词的置信度偏置，

表示置信度排名第i的样本关键词k_i的初步融合语义表征向量，

表示置信度排名第i的样本关键词k_i的融合语义表征向量。In formula 2, Self-Attention() represents the calculation process of the self-attention mechanism,

the feature vector representing the sample keywords,

Preliminary fused semantic representation vector representing sample keywords. In formula 3, pi represents the position embedding vector of the _i - _th sample keyword ki with confidence ranking, and _pi can capture the confidence bias of sample keywords with different rankings from the global level,

represents the initial fusion semantic representation vector of the _i -th sample keyword ki with confidence ranking,

Represents the fused semantic representation vector of the _i -th sample keyword ki with confidence ranking.

Confidence-aware aggregation layer: It is used to automatically select effective keywords related to user search information, and automatically filter noise information.

The user's search information may be related to some keywords, but not related to some keywords, so it is necessary to select the most relevant keywords to the search information.

Taking offline training as an example, the confidence-aware aggregation layer performs feature fusion on the current sample keyword and the sample search information for each sample keyword, and obtains the correlation between the current sample keyword and the sample search information. The fusion semantic representation vector and relevance of keywords are used to calculate the preliminary comprehensive semantic representation vector of sample keywords.

Optionally, for each sample keyword, the confidence-aware aggregation layer performs feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information, to obtain the relationship between the current sample keyword and the sample search information. Relevance: Based on the fusion semantic representation vector and relevancy of each sample keyword, and the feature vector of the sample search information, the preliminary comprehensive semantic representation vector of the sample keyword is calculated.

Optionally, the process of performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information to obtain the correlation between the current sample keyword and the sample search information may include: calculating the current sample The confidence of the keyword and the relevance weight of the search information; the current sample keyword is calculated based on the fusion semantic representation vector of the current sample keyword, the feature vector of the sample search information, and the relevance weight of the current sample keyword The intermediate parameter; based on the intermediate parameter and the preset temperature parameter, calculate the correlation between the current sample keyword and the sample search information.

As shown in Figure 7, in the confidence-aware convergence layer, through the Attention Network, the attention mechanism is used to calculate the correlation between confidence-aware, user search information and keyword fusion semantic representation vectors weights and adaptively select useful information from the fused semantic representation vector of keywords.

The formula of the confidence degree of the ith sample keyword k _i and the relevance weight of the search information is expressed as formula 4:

表示第i个样本关键词k_i的置信度，

表示第i个样本关键词k_i的相关度权重。In formula 4, softmax() represents the calculation process of the softmax function,

Represents the confidence of the _i -th sample keyword ki,

Represents the relevance weight of the _ith sample keyword ki.

则第i个样本关键词k_i的中间参数的公式表示如公式五：If corresponding to the above formula 1, the fusion semantic representation vector of the _i -th sample keyword ki is

Then the formula of the intermediate parameter of the ith sample keyword k _i is expressed as formula 5:

表示第i个样本关键词k_i的融合语义表征向量，e_q表示样本搜索信息的特征向量，s(q，k_i)表示第i个样本关键词k_i的中间参数。In formula 5, MLP represents a multi-layer feedforward neural network, || represents a splicing operation,

represents the fused semantic representation vector of the _ith sample keyword ki, e _q represents the feature vector of the sample search information, and s(q, _ki ) represents the intermediate parameter of the _ith sample keyword ki.

则第i个样本关键词k_i的中间参数的公式表示如公式六：If corresponding to the above formulas 2 and 3, the fusion semantic representation vector of the i-th sample keyword k _i is

Then the formula of the intermediate parameter of the i-th sample keyword k _i is expressed as formula 6:

表示第i个样本关键词k_i的融合语义表征向量，其他参数参照上述公式五的相关描述。In formula six,

Represents the fusion semantic representation vector of the _i -th sample keyword ki, and other parameters refer to the relevant description of the above formula 5.

The formula of the correlation between the i-th sample keyword _ki and the sample search information is expressed as formula 7:

a _A (q, k _i )=softmax(s(q, k _i )/τ) Formula 7

In formula 7, softmax() represents the calculation process of the softmax function, a _A (q, _ki ) represents the correlation between the _ith sample keyword ki and the sample search information, and τ represents the temperature parameter. The temperature parameter can increase the importance difference of different keywords, that is, keywords that can reflect the content theme and are more related to the user's search information are more likely to obtain greater weights, thereby contributing to the comprehensive semantic representation vector. The larger the value is, the noise keywords that have nothing to do with the search information have a small contribution, which can significantly reduce the interference of the noise keywords. In a word, combining the temperature parameter and the confidence-aware attention mechanism can effectively distinguish the importance of keywords at different confidence levels and improve the robustness of the model.

Optionally, based on the fusion semantic representation vector and correlation of each sample keyword, and the feature vector of the sample search information, the process of calculating the preliminary comprehensive semantic representation vector of the sample keyword may include: calculating the key of each sample. The sum of the product of the fusion semantic representation vector of the word and the correlation degree, to obtain the preliminary semantic representation vector of the sample keyword; adding the preliminary semantic representation vector and the feature vector of the sample search information to obtain the sample key Preliminary synthetic semantic representation vector for words.

则样本关键词的初步语义表征向量的公式表示如公式八：If corresponding to the above formula 1, the fusion semantic representation vector of the _i -th sample keyword ki is

Then the formula of the preliminary semantic representation vector of the sample keyword is expressed as formula 8:

则样本关键词的初步语义表征向量的公式表示如公式九：If corresponding to the above formulas 2 and 3, the fusion semantic representation vector of the i-th sample keyword k _i is

Then the formula of the preliminary semantic representation vector of the sample keyword is expressed as formula 9:

In Formula 8 and Formula 9, h _K represents the preliminary semantic representation vector of the sample keyword, and Σ represents the summation.

The formula of the preliminary comprehensive semantic representation vector of the sample keywords is expressed as formula ten:

表示样本关键词的初步综合语义表征向量。In formula ten, e _q represents the feature vector of the sample search information,

Preliminary synthetic semantic representation vector representing sample keywords.

Optionally, because the keyword may be inaccurate, forcibly introducing the keyword into the ranking model may bring noise to the model. In order to ensure robustness, the Dropout mechanism can be used to randomly discard the initial Part of the neuron in the semantic representation vector.

Therefore, after calculating the sum of the products of the fusion semantic representation vector of each sample keyword and the correlation degree to obtain the preliminary semantic representation vector of the sample keyword, the method further includes: randomly discarding the preliminary semantic representation vector.

The formula of the preliminary semantic representation vector after random discarding is expressed as formula 11:

h′ _K = dropout(h _K ) formula 11

In formula 11, dropout( ) represents the processing process of the drop mechanism, h′ _K represents the preliminary semantic representation vector after random drop processing, and h _K represents the preliminary semantic representation vector of the sample keyword.

Correspondingly, the preliminary semantic representation vector after the random discarding process is added to the feature vector of the sample search information to obtain the preliminary comprehensive semantic representation vector of the sample keywords.

The formula of the preliminary comprehensive semantic representation vector of the sample keywords is expressed as formula 12:

表示样本关键词的初步综合语义表征向量。In formula 12, e _q represents the feature vector of the sample search information,

Preliminary synthetic semantic representation vector representing sample keywords.

Output layer: used to output the comprehensive semantic representation vector of keywords based on the preliminary comprehensive semantic representation vector of keywords.

Taking offline training as an example, the initial comprehensive semantic representation vector of sample keywords is normalized at the output layer, and the comprehensive semantic representation vector of sample keywords is obtained.

The formula of the comprehensive semantic representation vector of the sample keywords is expressed as formula 13:

表示样本关键词的综合语义表征向量。In Formula 13, LayerNorm() represents the Layer Normalization process,

A comprehensive semantic representation vector representing sample keywords.

(2) Perform deep feature fusion and shallow feature fusion based on the comprehensive semantic representation vector of keywords to obtain the predicted value.

The comprehensive semantic representation vector of keywords is input into the feature interaction network for feature fusion. Through the feature interaction network, the sorting parameters of objects are obtained based on the comprehensive semantic representation vector of keywords. Specifically, feature fusion is performed on the comprehensive semantic representation vector, description information and search information of keywords to obtain the ranking parameters of objects.

Referring to FIG. 8 , a schematic diagram of a feature interaction network according to an embodiment of the present disclosure is shown. Figure 8 takes the multi-target model as an example, which mainly includes a deep fusion network (MMOE (Multi-gate Mixture-of-Experts) Network) and a shallow fusion network (FM (Factor Machine, factorization machine) Network) ). The underlying input features include the description information of the object, the search information (query) and the comprehensive semantic representation vector of the keywords obtained through the aforementioned semantic encoding network (Keyword Encoder). Wherein, as shown in FIG. 8 , the description information of the object may include, but is not limited to, a document feature (Item Feature), a context feature (Context Feature), and the like.

Taking offline training as an example, optionally, the process of performing feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information may include: performing the feature fusion on the comprehensive semantic representation vector, the sample description performing deep feature fusion with the sample search information to obtain the first sample sorting parameter of the sample object; performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information, Obtain the second sample sorting parameter of the sample object; and calculate the sample sorting parameter of the sample object based on the first sample sorting parameter and the first sample sorting parameter.

Optionally, deep feature fusion is performed on the comprehensive semantic representation vector, the sample description information and the sample search information, including: acquiring the feature vector of the sample description information and the feature vector of the sample search information; based on The integrated semantic representation vector, the feature vector of the sample description information, and the feature vector of the sample search information generate a splicing feature vector; a preset deep fusion network is used to perform feature fusion processing on the splicing feature vector to obtain the first Sample sorting parameters.

As shown in Figure 8, a feature vector is extracted for each feature of the sample description information by means of embedding matrix or sequence modeling, and the feature vector of the sample search information is obtained. The feature vectors of the same type are concatenated (Concat), as shown in Figure 8, the feature vectors of the document features are concatenated, and the feature vectors of the context features and the feature vectors of the sample search information are concatenated. The spliced vector is then spliced with the comprehensive semantic representation vector of the sample keywords to generate a spliced feature vector. The splicing feature vector is initially fused through the fusion layer (Fusion Network) as the input of the deep fusion network (MMoE Network). , output the task-specific predicted value, that is, the first sample ranking parameter.

Optionally, performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information, including: acquiring a feature vector of the sample description information and a feature vector of the sample search information; A preset shallow fusion network is used to perform feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information, and the comprehensive semantic representation vector to obtain the second sample ranking parameter.

As shown in Figure 8, a feature vector is extracted for each feature of the sample description information by means of embedding matrix or sequence modeling, and the feature vector of the sample search information is obtained. The feature vector of the sample description information, the feature vector of the sample search information, and the comprehensive semantic representation vector of the sample keyword are used as the input of the shallow fusion network (FM Network). The shallow fusion network combines the comprehensive semantic representation vector of the keyword and some memory The feature vector of the characteristic sparse feature performs sufficient feature interaction, and outputs the predicted value shared by the task, that is, the second sample ranking parameter.

Optionally, the first sample sorting parameter and the first sample sorting parameter are correspondingly added to obtain the sample sorting parameters of each task corresponding to the sample object. Sample ranking parameters may include, but are not limited to, click-through rate, conversion rate, exposure, user interaction behavior parameters, and the like.

In the embodiments of the present disclosure, a confidence-aware keyword semantic coding network is proposed, which can effectively extract semantic information in keywords, automatically filter noise information, and improve the robustness of semantic representation; at the same time, a fusion key The framework that integrates the semantic representation vector of words into the ranking model can fully interact with the important context, document, user and other features in the search scene with the help of the expressive ability of pre-trained vectors and keywords, thereby improving the relevance of search results. and quality.

By introducing keyword information into the ranking model, initializing parameters based on the pre-training vector, and then fine-tuning in the ranking model, the ranking model can effectively capture semantic relevance information and improve the relevance of ranking. Through the attention mechanism based on confidence perception, the importance of different keywords can be better distinguished, noise can be reduced, and the robustness of semantic representation can be further ensured through temperature parameters and dropout mechanism.

Keywords are very important structured information for content-based communities and platforms. They can be used to describe the topic of content and users' explicit interest preferences, and are widely used in content search and recommendation scenarios. Among them, the search scene can be used to capture the correlation between user search words and keywords; the recommendation scene can construct the keyword sequence of the content that the user has clicked, which can be used to describe the user's long-term and short-term interests and build user portrait tags, and then based on the user portraits Tags and content keywords are used to generalize and accurately capture user interests and improve recommendation performance. This method may be adopted in all content platforms.

Referring to FIG. 9 , a structural block diagram of a model training apparatus according to an embodiment of the present disclosure is shown.

As shown in Figure 9, the model training device may include the following modules:

The first obtaining module 901 is used for obtaining sample data; the sample data includes sample search information and sample keywords corresponding to the sample object;

A training module 902, configured to perform feature fusion on the sample keywords and the sample search information in a preset model to be trained, to obtain a comprehensive semantic representation vector of the sample keywords, based on the comprehensive semantic representation vector obtaining the sample sorting parameters of the sample object;

The determining module 903 is configured to determine the model that has been trained based on the sample sorting parameter as the sorting model.

Optionally, the training module 902 includes: a first fusion unit, configured to perform feature fusion on the current sample keyword and other sample keywords for each sample keyword, to obtain a fusion semantic representation vector of the current sample keyword; The second fusion unit is configured to perform feature fusion on the current sample keyword and the sample search information for each sample keyword, to obtain the correlation between the current sample keyword and the sample search information; the first calculation unit, using Based on the fusion semantic representation vector and relevancy of each sample keyword, the comprehensive semantic representation vector of the sample keyword is calculated.

Optionally, the training module 902 includes: a query unit, configured to query the feature vector of each sample keyword and the feature vector of the sample search information from the preset correspondence between words and feature vectors; the third fusion unit, which is used for feature fusion of the feature vector of the current sample keyword and the feature vectors of other sample keywords for each sample keyword, to obtain the fusion semantic representation vector of the current sample keyword; the fourth fusion unit is used for For each sample keyword, feature fusion is performed on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information to obtain the correlation between the current sample keyword and the sample search information; the second calculation unit , which is used to calculate the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and relevance of each sample keyword and the feature vector of the sample search information.

Optionally, the third fusion unit is specifically configured to perform feature fusion between the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data also includes the confidence level of the sample keywords; the third fusion unit is specifically used to analyze the feature vector of the current sample keyword and the feature vector of other sample keywords through a self-attention mechanism. Perform feature fusion to obtain a preliminary fusion semantic representation vector of the current sample keywords; sort the sample keywords in descending order according to the confidence, and obtain the position embedding vector of each sample keyword based on the sorting result; The initial fusion semantic representation vector and the position embedding vector are added to obtain the fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further includes the confidence of the sample keywords; the fourth fusion unit is specifically configured to calculate the confidence of the current sample keywords and the correlation weight of the search information; based on the current sample The fusion semantic representation vector of the keyword, the feature vector of the sample search information, and the relevance weight of the current sample keyword, calculate the intermediate parameter of the current sample keyword; based on the intermediate parameter and the preset temperature parameter, calculate the current The correlation between the sample keywords and the sample search information.

Optionally, the second calculation unit is specifically configured to calculate the sum of the product of the fusion semantic representation vector and the correlation degree of each sample keyword, and obtain the preliminary semantic representation vector of the sample keyword; The vector is added with the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword; the preliminary comprehensive semantic representation vector is standardized to obtain a comprehensive semantic representation vector of the sample keyword.

Optionally, the second computing unit is further configured to calculate the sum of the product of the fusion semantic representation vector of each sample keyword and the correlation degree, and obtain the preliminary semantic representation vector of the sample keyword. The semantic representation vector is randomly discarded; the second computing unit is specifically configured to add the preliminary semantic representation vector after the random discard process and the feature vector of the sample search information.

Optionally, the sample data further includes sample description information corresponding to the sample object; the training module 902 is specifically configured to characterize the comprehensive semantic representation vector, the sample description information and the sample search information. Fusion to obtain the sample sorting parameters of the sample object.

Optionally, the training module 902 includes: a fifth fusion unit, configured to perform deep feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain the first a sample sorting parameter; a sixth fusion unit, configured to perform shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a second sample sorting parameter of the sample object; the sixth A third calculating unit, configured to calculate the sample sorting parameter of the sample object based on the first sample sorting parameter and the first sample sorting parameter.

Optionally, the fifth fusion unit is specifically configured to obtain the feature vector of the sample description information and the feature vector of the sample search information; based on the comprehensive semantic representation vector, the feature vector of the sample description information and A splicing feature vector is generated from the feature vector of the sample search information; the feature fusion process is performed on the splicing feature vector by using a preset deep fusion network to obtain the first sample sorting parameter.

Optionally, the sixth fusion unit is specifically configured to obtain the feature vector of the sample description information and the feature vector of the sample search information; use a preset shallow fusion network to compose the feature vector of the sample description information. and performing feature fusion processing on the feature vector of the sample search information and the comprehensive semantic representation vector to obtain the second sample sorting parameter.

Referring to FIG. 10 , a structural block diagram of a sorting apparatus according to an embodiment of the present disclosure is shown.

As shown in Figure 10, the sorting device may include the following modules:

The second obtaining module 1001 is used to obtain search information and keywords corresponding to the objects to be sorted;

The prediction module 1002 is configured to input the search information and the keywords into a pre-trained sorting model, and obtain the sorting parameters of the objects to be sorted output by the sorting model; The model training method described above is obtained.

A sorting module 1003, configured to sort the objects to be sorted based on the sorting parameters.

In the embodiment of the present disclosure, the keyword information of the sample objects is integrated into the ranking model. These keywords are mined for the unstructured features of the objects. Compared with the unstructured features of the objects, the keywords can better Refining and summarizing the subject of the object can cover and describe a wide range of users’ intentions. Therefore, the ranking model incorporating keyword information can capture the semantics of the keywords themselves and the semantic correlation between keywords and user search information, thereby improving the ranking model. accuracy.

As for the apparatus embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and reference may be made to the partial description of the method embodiment for related parts.

In an embodiment of the present disclosure, an electronic device is also provided. The electronic device may include one or more processors, and one or more computer-readable storage media having stored thereon instructions, such as application programs. When the instructions are executed by one or more processors, the processor is caused to perform the model training method as in any of the above embodiments, or, perform the ranking method as in any of the above embodiments.

In an embodiment of the present disclosure, there is also provided a non-transitory computer-readable storage medium on which a computer program is stored, the computer program can be executed by a processor of an electronic device, and when the computer program is executed by the processor , causing the processor to execute the model training method described in any of the above embodiments, or to execute the sorting method described in any of the above embodiments.

The above-mentioned processor may be a general-purpose processor, which may include, but is not limited to, a central processing unit (CPU for short), a network processor (NP for short), and a digital signal processor (Digital Signal Processing for short) DSP), Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. .

The above-mentioned computer-readable storage medium may include but is not limited to: read-only memory (Read OnlyMemory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), compact disc read-only memory (Compact Disc ReadOnly Memory, CD-ROM for short), Electronic Erasable Programmable ReadOnly Memory (EEPROM for short), hard disk, floppy disk, flash memory, and so on.

The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used with teaching based on this. The structure required to construct such a system is apparent from the above description. Furthermore, embodiments of the present disclosure are not directed to any particular programming language. It should be understood that various programming languages may be utilized to implement the contents of the embodiments of the present disclosure described herein, and that the above descriptions of specific languages are intended to disclose the best modes of implementation of the embodiments of the present disclosure.

In the description provided herein, numerous specific details are set forth. It will be understood, however, that embodiments of the present disclosure may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Similarly, it is to be understood that in the above descriptions of exemplary embodiments of the present disclosure, various features of the embodiments of the present disclosure are sometimes grouped together into a single Examples, figures, or descriptions thereof. However, this method of disclosure should not be interpreted as reflecting an intention that the claimed embodiments of the disclosure require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the Detailed Description are hereby expressly incorporated into this Detailed Description, with each claim standing on its own as a separate embodiment of embodiments of the present disclosure.

Those skilled in the art will understand that the modules in the device in the embodiment can be adaptively changed and arranged in one or more devices different from the embodiment. The modules or units or components in the embodiments may be combined into one module or unit or component, and further they may be divided into multiple sub-modules or sub-units or sub-assemblies. All features disclosed in this specification (including accompanying claims, abstract and drawings) and any method so disclosed may be employed in any combination, unless at least some of such features and/or procedures or elements are mutually exclusive. All processes or units of equipment are combined. Each feature disclosed in this specification (including accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

The various component embodiments of the embodiments of the present disclosure may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all functions of some or all components of the dynamic picture generating apparatus according to the embodiments of the present disclosure. Embodiments of the present disclosure may also be implemented as apparatus or apparatus programs for performing a part or all of the methods described herein. Such a program implementing embodiments of the present disclosure may be stored on a computer-readable medium, or may be in the form of one or more signals. Such signals may be downloaded from Internet sites, or provided on carrier signals, or in any other form.

It should be noted that the above-described embodiments illustrate rather than limit embodiments of the present disclosure, and that alternative embodiments may be devised by those skilled in the art without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Embodiments of the present disclosure may be implemented by means of hardware comprising several distinct elements, as well as by means of suitably programmed computers. In a unit claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, and third, etc. do not denote any order. These words can be interpreted as names.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described systems, devices and units may refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

The above are only specific implementations of the embodiments of the present disclosure, but the protection scope of the embodiments of the present disclosure is not limited thereto, and any person skilled in the art is within the technical scope disclosed by the embodiments of the present disclosure , changes or substitutions can be easily conceived, and all should be included within the protection scope of the embodiments of the present disclosure.

Claims (17)

1. a model training method, is characterized in that, comprises: Obtain sample data; the sample data includes sample search information and sample keywords corresponding to the sample object; In the preset to-be-trained model, feature fusion is performed on the sample keyword and the sample search information to obtain a comprehensive semantic representation vector of the sample keyword, and the sample object is obtained based on the comprehensive semantic representation vector. sample sorting parameters; The trained model is determined based on the sample sorting parameters as the sorting model. 2. The method according to claim 1, wherein the feature fusion is performed on the sample keywords and the sample search information, comprising: For each sample keyword, feature fusion is performed on the current sample keyword and other sample keywords to obtain the fusion semantic representation vector of the current sample keyword; For each sample keyword, feature fusion is performed on the current sample keyword and the sample search information to obtain the correlation between the current sample keyword and the sample search information; Based on the fusion semantic representation vector and relevance of each sample keyword, the comprehensive semantic representation vector of the sample keyword is calculated. 3. The method according to claim 1, wherein the feature fusion is performed on the sample keywords and the sample search information, comprising: From the preset correspondence between words and feature vectors, query the feature vector of each sample keyword and the feature vector of the sample search information; For each sample keyword, feature fusion is performed between the feature vector of the current sample keyword and the feature vectors of other sample keywords to obtain the fusion semantic representation vector of the current sample keyword; For each sample keyword, feature fusion is performed on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information to obtain the correlation between the current sample keyword and the sample search information; Based on the fusion semantic representation vector and relevance of each sample keyword, and the feature vector of the sample search information, the comprehensive semantic representation vector of the sample keyword is calculated. 4. The method according to claim 3, wherein the feature fusion of the feature vector of the current sample keyword and the feature vectors of other sample keywords, comprising: The feature vector of the current sample keyword and the feature vector of other sample keywords are fused through the self-attention mechanism, and the fused semantic representation vector of the current sample keyword is obtained. 5. The method according to claim 3, wherein the sample data further comprises the confidence of the sample keywords; feature fusion is performed on the feature vector of the current sample keyword and the feature vectors of other sample keywords, include: Through the self-attention mechanism, the feature vector of the current sample keyword and the feature vector of other sample keywords are feature fusion, and the preliminary fusion semantic representation vector of the current sample keyword is obtained; Sort the sample keywords in descending order according to the confidence level, and obtain the position embedding vector of each sample keyword based on the sorting result; The initial fusion semantic representation vector of the current sample keyword and the position embedding vector are added to obtain the fusion semantic representation vector of the current sample keyword. 6. The method according to claim 3, wherein the sample data further includes the confidence of the sample keywords; the fusion semantic representation vector of the current sample keywords and the feature vector of the sample search information are performed. Feature fusion, including: Calculate the confidence of the current sample keyword and the relevance weight of the search information; Calculate the intermediate parameter of the current sample keyword based on the fusion semantic representation vector of the current sample keyword, the feature vector of the sample search information, and the relevance weight of the current sample keyword; Based on the intermediate parameter and the preset temperature parameter, the correlation between the current sample keyword and the sample search information is calculated. 7. The method according to claim 3, wherein the comprehensive semantic representation vector of the sample keywords is calculated based on the fusion semantic representation vector and correlation of each sample keyword and the feature vector of the sample search information ,include: Calculate the sum of the product of the fusion semantic representation vector of each sample keyword and the correlation degree to obtain the preliminary semantic representation vector of the sample keyword; adding the preliminary semantic representation vector and the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword; Standardize the preliminary comprehensive semantic representation vector to obtain the comprehensive semantic representation vector of the sample keyword. 8. The method of claim 7, wherein: After calculating the sum of the product of the fusion semantic representation vector of each sample keyword and the correlation degree to obtain the preliminary semantic representation vector of the sample keyword, the method further includes: randomly discarding the preliminary semantic representation vector; Adding the preliminary semantic representation vector and the feature vector of the sample search information includes: adding the preliminary semantic representation vector after random discarding processing to the feature vector of the sample search information. 9 . The method according to claim 1 , wherein the sample data further comprises sample description information corresponding to the sample object; obtaining the sample sorting parameters of the sample object based on the comprehensive semantic representation vector, comprising: 10 . Feature fusion is performed on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain sample sorting parameters of the sample object. 10. The method according to claim 9, wherein feature fusion is performed on the comprehensive semantic representation vector, the sample description information and the sample search information, comprising: performing deep feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain the first sample sorting parameter of the sample object; performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a second sample sorting parameter of the sample object; A sample ranking parameter of the sample object is calculated based on the first sample ranking parameter and the first sample ranking parameter. 11. The method according to claim 10, wherein performing deep feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information, comprising: Obtain the feature vector of the sample description information and the feature vector of the sample search information; Generate a splicing feature vector based on the comprehensive semantic representation vector, the feature vector of the sample description information and the feature vector of the sample search information; A preset deep fusion network is used to perform feature fusion processing on the spliced feature vector to obtain the first sample sorting parameter. 12. The method according to claim 10, wherein performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information, comprising: Obtain the feature vector of the sample description information and the feature vector of the sample search information; A preset shallow fusion network is used to perform feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information, and the comprehensive semantic representation vector to obtain the second sample ranking parameter. 13. A sorting method, comprising: Obtain the search information and keywords corresponding to the objects to be sorted; Inputting the search information and the keywords into a pre-trained sorting model to obtain sorting parameters of the objects to be sorted output by the sorting model; The model training method is obtained; The objects to be sorted are sorted based on the sorting parameters. 14. A model training device, comprising: a first acquisition module, configured to acquire sample data; the sample data includes sample search information and sample keywords corresponding to the sample object; A training module, configured to perform feature fusion on the sample keywords and the sample search information in a preset model to be trained, to obtain a comprehensive semantic representation vector of the sample keywords, and obtain a comprehensive semantic representation vector based on the comprehensive semantic representation vector sample sorting parameters of the sample object; A determination module, configured to determine a model that has been trained based on the sample sorting parameters as a sorting model. 15. A sorting device, comprising: The second acquisition module is used to acquire search information and keywords corresponding to the objects to be sorted; A prediction module, configured to input the search information and the keywords into a pre-trained sorting model, and obtain the sorting parameters of the objects to be sorted output by the sorting model; The model training method described in any one is obtained; A sorting module, configured to sort the objects to be sorted based on the sorting parameters. 16. An electronic device, comprising: one or more processors; and one or more computer-readable storage media having instructions stored thereon; When executed by the one or more processors, the instructions cause the processors to perform the model training method as claimed in any one of claims 1 to 12, or to perform the sorting method as claimed in claim 13 . 17. A computer-readable storage medium, characterized in that a computer program is stored thereon, and when the computer program is executed by a processor, the processor is made to execute the method according to any one of claims 1 to 12. A model training method, or, performing the ranking method as claimed in claim 13 .

Images