Data processing methods, apparatus, computer equipment and storage media

Abstract

This invention discloses a data processing method, apparatus, computer device, and storage medium. The method includes: acquiring search behavior data and search documents; the search behavior data includes search behavior terms and corresponding category tags, wherein the probability of using search documents for category search is less than the probability of using search behavior terms for category search; extracting the document topics of the search documents and generating search reference data based on the search documents and corresponding document topics; the category tags associated with the search documents are used as category tags of the document topics; using the search reference data and search behavior data as sample data, and using the sample data to train a target search model to obtain a trained target search model; the trained target search model is used to predict the correspondence probability between input search terms and different category tags, which can improve the accuracy of model training.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] With the continuous development of computer technology, the current training methods for category prediction models generally adopt category mapping, i.e., constructing mapping relationships, or statistical analysis to construct and train model samples. However, research has shown that the current model training methods have the problem of low prediction accuracy. Therefore, how to improve the accuracy in category prediction scenarios has become a current research hotspot. Summary of the Invention

[0003] This invention provides a data processing method, apparatus, computer equipment, and storage medium that can improve the accuracy of model training.

[0004] On one hand, embodiments of the present invention provide a data processing method, including:

[0005] Acquire search behavior data and search documents; the search behavior data includes search behavior terms and category tags corresponding to the search behavior terms, wherein the probability of using the search documents to perform category searches is less than the probability of using the search behavior terms to perform category searches;

[0006] Extract the document topic of the search document, and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic;

[0007] The search reference data and the search behavior data are used as sample data, and the target search model is trained using the sample data to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

[0008] In another aspect, embodiments of the present invention provide a data processing apparatus, including:

[0009] An acquisition unit is used to acquire search behavior data and search documents; the search behavior data includes search behavior terms and category tags corresponding to the search behavior terms, wherein the probability of using the search document to perform a category search is less than the probability of using the search behavior terms to perform a category search;

[0010] The processing unit is used to extract the document topic of the search document and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic;

[0011] The processing unit is further configured to use the search reference data and the search behavior data as sample data, and use the sample data to train the target search model to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

[0012] In another aspect, embodiments of the present invention provide a computer device, including a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, wherein the memory is used to store a computer program supporting the computer device in executing the above-described method, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the following steps:

[0013] Acquire search behavior data and search documents; the search behavior data includes search behavior terms and category tags corresponding to the search behavior terms, wherein the probability of using the search documents to perform category searches is less than the probability of using the search behavior terms to perform category searches;

[0014] Extract the document topic of the search document, and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic;

[0015] The search reference data and the search behavior data are used as sample data, and the target search model is trained using the sample data to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

[0016] In another aspect, embodiments of the present invention provide a computer-readable storage medium storing program instructions, which, when executed by a processor, are used to perform the data processing method as described in the first aspect.

[0017] In this embodiment, when training a target search model, if the computer device cannot obtain a sufficient amount of search behavior data, it can acquire the search behavior data and search documents. Based on topic extraction from the search documents, it can construct search reference data. Then, both the search behavior data and the search reference data can be used as training samples for the target search model, effectively expanding the sample data and improving the accuracy of model training. This achieves effective utilization of useful information during model training, enhancing the flexibility and effectiveness of expanding model sample data. Using search object behavior data and topic-extracted text from search documents as training data for the text classification model maximizes the reflection of the search object's search tendency on frequently occurring head search terms, and effectively predicts the search category for mid- and long-tail search terms. This significantly improves the overall accuracy of intent recognition, allowing the computer device to better understand the search object's search intent, avoiding the appearance of irrelevant category documents through text matching, and improving the search experience. This effectively enhances the computer device's search prediction capabilities. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a data processing system provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic flowchart of a data processing method provided in an embodiment of the present invention;

[0021] Figure 3a This is a schematic diagram illustrating the training process of a target search model provided in an embodiment of the present invention;

[0022] Figure 3b This is a schematic diagram illustrating an example of item attribute extraction provided by an embodiment of the present invention;

[0023] Figure 3c This is a schematic diagram illustrating the prediction and evaluation time consumption provided in an embodiment of the present invention;

[0024] Figure 3d This is a schematic diagram illustrating the comparison of prediction results based on a target search model, provided by an embodiment of the present invention.

[0025] Figure 3eThis is a schematic diagram illustrating feedback based on search input provided in an embodiment of the present invention;

[0026] Figure 3f This is a schematic diagram illustrating a search term expansion method provided in an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram illustrating a search feedback based on a search request, provided by an embodiment of the present invention.

[0028] Figure 5 This is a schematic block diagram of a data processing device provided in an embodiment of the present invention;

[0029] Figure 6 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0030] This application proposes a data processing method that, when training a target search model, a computer device not only constructs corresponding sample data based on search behavior data generated from the search behavior of the search object, but also constructs sample data from search documents with a low probability of being used during the search. This expands the sample data during model training. Since the expansion of the training samples is done using search documents with a low probability of being used during the search, and since these search documents contain a lot of descriptive information about the resource to be fed back, the target search model can learn the semantic information of the resource to be fed back in the search scenario from the search documents. Based on the semantic learning of the resource in the search scenario, the target search model's ability to understand the semantics of the resource to be fed back in the search scenario can be effectively improved. Simultaneously, by training the target search model with search behavior data generated from the search behavior of the search object, the target search model trained by the computer device can analyze and understand the behavioral tendencies of the search object in a search scenario. This allows the target search model to combine the user's search behavior tendencies with the semantic understanding of the resource to be provided when providing search feedback, thereby improving the effectiveness of the target search model's feedback and further increasing the search conversion rate of the corresponding resources using the target search model. In one embodiment, the target search model is a text classification model, which is used to predict the search target (i.e., the category tag corresponding to the input search term) of the search object, and the probability that the corresponding search target is confirmed by the search object. It can be understood that model-based search and model training belong to Artificial Intelligence (AI) technology. Artificial intelligence is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain optimal results. In other words, artificial intelligence (AI) is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce new intelligent machines that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0031] In one embodiment, when a computer device trains a target search model, if the computer device acquires a large amount of search behavior data—that is, if the amount of search behavior data acquired by the computer device meets the training requirements for the target search model—the computer device can directly use the corresponding amount of search behavior data to train the target search model, thereby obtaining a trained target search model. Specifically, after acquiring the search behavior data, the computer device can compare the amount of data corresponding to the acquired search behavior data with a predetermined amount of data (where the predetermined amount of data is sufficient to train the template search model). If the amount of data corresponding to the acquired search behavior data is greater than or equal to the predetermined amount of data, it is determined that the acquired search behavior data meets the training requirements for the target search behavior; otherwise, it is considered that the acquired search behavior data does not meet the training requirements for the target search model.

[0032] If the computer device acquires insufficient search behavior data when training the target search model, it can be considered that the acquired search behavior data does not meet the training requirements for the target search model. In this case, the computer device can expand the training samples of the target search model based on the data processing method proposed in this application, through the acquisition and topic extraction of search documents. This enables effective training of the target search model and allows for full utilization of effective information during model training. Furthermore, the expansion of training samples based on the processing of search documents enhances the flexibility of expanding the training samples of the target search model. In addition, since the target application corresponding to the target search model cannot generate a large amount of search behavior data in the initial stage, the computer device can determine the training samples of the target search model in the early stages of the application by acquiring search behavior data and processing search documents. Later in the application, if the computer device can acquire more search behavior data, it can directly use the acquired search behavior data to train the target search model if it determines that model training is required at this point.

[0033] In one embodiment, artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing (NLP), and machine learning / deep learning. The data processing methods mentioned in this application primarily involve NLP and machine learning (ML) technologies within AI. NLP is a crucial area within computer science and AI. It studies various theories and methods that enable effective communication between humans and computers using natural language. NLP is a science integrating linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use daily—and thus has a close connection to linguistic research. NLP technologies typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs. Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0034] The search behavior data used when training the target search model includes search terms frequently used by the search object in the search scenario (i.e., those with high frequency of use), and the corresponding category tags. The search term is the search word entered by the search object in the search scenario, which could be something like "toy," "apple," or "medicine cabinet." The corresponding category tag is the resource type of one or more feedback resources selected and confirmed by the search object based on the corresponding search word. In other words, based on the search term and its corresponding category tag, the model effectively reflects the type of feedback resources expected by the search object when entering search terms in the search scenario, thus effectively reflecting the search object's search behavior tendencies. To enrich the training samples for the model, the search documents used to expand the training samples refer to the information sources of the resources provided to the search object, such as the resource name or the content of the feedback webpage.

[0035] Based on the acquisition of search documents, the computer device can extract document topics from the search documents and construct sample data based on these topics. In one embodiment, there are multiple sample data sets. When the computer device uses these sample data to train the target search model, it iteratively inputs the sample data into the target search model. Since the sample data referenced when training the target search model includes sample search terms (including the aforementioned search behavior terms and search reference data) and corresponding category tags, training the target search model based on these sample data means training with these category tags as the learning objective of the target search model. This can improve the training accuracy of the target search model, and by determining the learning objective of the target search model, the training speed of the target search model can be improved. In one embodiment, the low-probability search documents used by the computer device when expanding sample data refer to search documents that are used less frequently but are still in use in certain situations. Based on these search documents, corresponding long-tail or mid-tail keywords can be identified. These long-tail or mid-tail keywords are highly related to the core keyword and are used to further explain and define the core keyword. For example, if the keyword is mung bean, then the corresponding long-tail (or mid-tail) keywords would be methods for using mung beans. Generally speaking, 80% of the search terms entered by the search object in a search scenario are long-tail keywords, and because long-tail keywords generally contain more meanings and connotations...

[0036] In one embodiment, the data processing method can be applied to, for example... Figure 1In the data processing system shown, the system includes a terminal device 10 and a server 11. The terminal device 10 is a device controlled and executed by the search object, while the server 11 is a device that responds to the search scenario of the terminal device 10. Specifically, it can be used to train the target search model required for executing the search, call the trained target search model to respond to the search request, and display the corresponding search results. In one embodiment, the server 11 and the terminal device 10 can be as follows: Figure 1 The two independent physical devices shown can also be two different processing modules integrated into one physical device; this application does not limit the specific implementation of these devices. The terminal device 10 includes, but is not limited to, smartphones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc., and this application does not impose any restrictions on these. The server 11 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, but is not limited to these. Furthermore, this application does not limit the number of servers. By expanding the corresponding sample data during the training of the target search model, computer devices can not only quickly predict the click probability (i.e. the probability of selection and confirmation) of different categories corresponding to the search terms of the target object through the trained neural network model, and give priority to displaying content under the categories that the target object wants, but also better understand the search needs of the target object based on the text semantics of the search document, thereby further improving the click conversion rate of the search.

[0037] Please see Figure 2 This is a schematic flowchart of a data processing method proposed in an embodiment of this application, such as... Figure 2 As shown, the method may include:

[0038] S201, Obtain search behavior data and search documents; the search behavior data includes search behavior terms and corresponding category tags, wherein the probability of using search documents for category search is less than the probability of using search behavior terms for category search.

[0039] Both search behavior data and search documents are used to construct information sources corresponding to the training samples of the target search model. This target search model is a semantic search model based on intent recognition, and its processing results are used to obtain category prediction results for the input text. Intent recognition involves understanding the core search needs of the search object based on the keywords entered by the user and returning the most suitable results. Semantic search means that the work of the search engine (i.e., a search engine equipped with a target search model) no longer focuses solely on the literal meaning of the search query, but rather sees through the surface to accurately capture the true intent behind the search query and uses this intent to perform the search, thus returning the most relevant search results to the user. Furthermore, category prediction is used to predict the category to which the search text belongs; for example, if the search object enters "water," the system should return either "drinking water" or "skincare water." In one embodiment, the processing of the target search model mainly includes offline processing and online processing. In the offline processing, the main purpose is to build model training samples and estimate the category probability corresponding to the text based on the text classification model (i.e., the target search model), and finally obtain the trained text classification model. In the online processing, the search content input by the search object is used to call the offline model to obtain the corresponding category probability. The search results under different categories are weighted and sorted according to the probability, and finally the search results are fed back to the search object.

[0040] In one embodiment, since the training samples used by the computer device to train the target search model are text, the process of training the target search model can also be considered a text classification task based on text search samples. A text classification task refers to assigning text to one or more categories. In this embodiment, a text classification model can be obtained by training a large number of text samples and corresponding categories. Therefore, search category prediction can essentially be converted into a text classification task. In the category prediction scenario, the training samples for training the target search model generally include the click category corresponding to the user's search terms and the category corresponding to the search document. The specific training process for this target search model can be as follows: Figure 3a As shown. Among them, text classification is a common task in natural language processing. With the development of deep learning, target search models based on text classification can predict the classification results of text relatively accurately.

[0041] In one embodiment, the search behavior data and search documents acquired by the computer device can be obtained from corresponding text data, such as text data generated from the search behavior of the search object, or text data extracted from the speech data generated from the search behavior of the search object. In one embodiment, when acquiring search behavior data and search documents, the computer device can perform text processing on the corpus data generated by the search object in the search scenario to obtain the corresponding search behavior data and search documents. Since Chinese text classification generally uses word-level granularity to train the corresponding model, word segmentation is required first. The word segmentation tool used to perform this processing can include, but is not limited to, jieba (a word segmentation tool) and FastText (a word segmentation tool). For vertical fields such as e-commerce, item terms can be added, and for the medical field, disease and drug terms can be added. Optionally, when the search data source is relatively rich, unsupervised new word discovery can be performed on a large amount of search corpus text, such as new word discovery based on cohesion and freedom, to further supplement new words or entity words in professional fields. FastText is an open-source tool for word vector calculation and text classification. In text classification tasks, FastText (a shallow network) achieves accuracy comparable to deep networks, but its training time is orders of magnitude faster. In its implementation, when acquiring search behavior data, the computer device first acquires a corpus of text, which includes historical search text and search results based on that historical text. The search results include at least one search feedback resource obtained from the historical search text, and the search execution status of that resource.

[0042] After acquiring the corpus text, the computer device can perform word segmentation and filtering on the historical search text contained in the corpus text to obtain the search words corresponding to the historical search text. Based on the search feedback resources included in the search results and the corresponding search execution status, the search target of the historical search text is determined. The filtering process is used to further remove stop words from the text corpus. Stop words include, but are not limited to, punctuation marks, modal particles, adverbs, prepositions, and conjunctions. After the computer device determines the search target of the historical search text, the search words can be used as search behavior words, and the search target can be used as a category tag associated with the search behavior words. Thus, search behavior data can be further generated based on the search behavior words associated with category tags. In one embodiment, the corpus text may also contain resource description text. When the corpus text contains resource description text, based on the above processing of the resource description text by the computer device, the search document corresponding to the resource description text can be obtained, thereby achieving simultaneous acquisition of search behavior data and search documents, effectively improving the speed of the computer device in acquiring raw sample data.

[0043] In one embodiment, the computer device acquires multiple sets of search behavior data, and each set of search behavior data is obtained by associating a corresponding search term with a corresponding category tag. The search terms entered by the search object generally follow an 80 / 20 distribution, meaning 20% ​​of the search terms are frequently searched (i.e., the number of searches meets a preset limit), and 80% are less frequently searched terms. For texts with a high number of searches, the click behavior of the search object after the search term largely reflects the relevance between the search term and the category; that is, the more categories clicked, the more likely they are to be related to the intent of the search term. Therefore, the computer device can construct the corresponding search behavior data by combining the search terms in the search scenarios already performed by the search object (i.e., historical search terms) with the categories of resources confirmed based on those historical search terms. The distribution of the search terms entered by the search object and the corresponding confirmed categories is shown in Table 1.

[0044] Table 1

[0045] Search terms Percentage of Quantities Category apple 20.2251％ fruit apple 17.1006％ Bluetooth headphones apple 11.062％ Tablet PC apple 7.2274％ Tablet accessories apple 6.8224％ thin and light laptops apple 6.086％ cell phone Book 97.2174％ books Book 2.7826％ Household Daily Necessities

[0046] In one embodiment, the category shown in Table 1 can be a first-level category as shown in Table 1, or it can be a third-level category. For search terms with search behavior data of the search object, the search terms and click categories of the search object are used as training data for the target search model, so that the target search model can learn the probability distribution of the corresponding category by the search terms and their frequency of occurrence in different categories, as shown in Table 2.

[0047] Table 2

[0048] Search terms First-level category / Second-level category / Third-level category apple Mobile communication / mobile phone / mobile phone apple Mobile Communications / Mobile Phone Accessories / Mobile Phone Screen Protectors apple Fresh produce / fruits / apples Apple's new Mobile communication / mobile phone / mobile phone water Food and Beverages / Beverage Preparation / Drinking Water water Beauty and skincare / facial skincare / skincare toner

[0049] Because some search terms in the total number of search terms may have poor quality, such as typos, garbled text, or excessively long text, the computer device can filter or replace these search terms before incorporating them into the model for training. For example, for typos, a text-based language model or a deep learning-based error correction model can be used to correct and replace the search terms. Garbled text or excessively long text is directly filtered out and not used as training samples. Specifically, the computer device can obtain all search behavior terms and their distribution. The distribution indicates the frequency with which different search behavior terms are used in resource searches. After obtaining all search behavior terms and their distribution, the device can select search behavior terms that meet a preset frequency requirement based on the frequency indicated by the distribution. The selected search behavior terms and their associated category tags are then used as search behavior data.

[0050] S202, extract the document topic of the search document, and generate search reference data based on the search document and the corresponding document topic; the class tags associated with the search document are used as class tags of the document topic.

[0051] Computer devices can construct sample data not only based on search behavior data but also on search documents. While search documents lack corresponding large amounts of search behavior data, making model prediction difficult, the text content of numerous search documents inherently contains information. Therefore, this large amount of document text can be used as training data. For example, in item search engines, a large number of item titles and corresponding item categories are used as training data to obtain a corresponding prediction model. However, since the length of search terms is generally much shorter than the length of document content, training directly based on search document content results in a high error rate. The core reason is that models trained on long texts of document content perform poorly in predicting short texts of search terms. In one embodiment, because long search documents contain a large amount of information that does not conform to the information distribution pattern of search terms, and contain a lot of invalid text information that easily causes noise pollution to the model, data augmentation operations can be performed on search documents to remove invalid information. By introducing a text topic extraction model, core keywords can be extracted and then used for training. For example, in an item search engine, the item title "1st generation mobile phone (A1234) 300GB black supporting China Mobile, China Unicom and China Telecom 5G (5th generation mobile communication technology) dual SIM dual standby mobile phone" is quite long. However, the search target in an item search will generally not show "supports China Mobile, China Unicom and China Telecom". Instead, the search input will be the brand, item, model, size, color, etc. In a general text search engine, it may only search for the brand or model.

[0052] In one embodiment, when a computer device retrieves a search document, it can acquire the resource description text of the resource to be fed back (such as the item to be fed back), and then perform data augmentation processing on the resource description text to obtain the augmented text corresponding to the resource to be fed back, and use the obtained augmented text as the search document; or, the computer device can acquire the resource description text of the resource to be fed back, and extract the domain keywords of the corresponding resource domain from the resource description text based on the resource domain corresponding to the resource to be fed back, and use the extracted domain keywords as the search document. When the computer device performs data augmentation processing on the resource description text, it can first perform word segmentation processing on the resource description text to obtain the corresponding descriptive words, and then obtain synonyms based on the descriptive words, and randomly combine the synonyms of the obtained descriptive words with the original descriptive words to obtain the new resource description text as the augmented text, so as to realize the data augmentation processing of the resource description text. In one embodiment, when the computer device extracts domain keywords of the corresponding resource domain from the resource description text based on the resource domain of the resource to be fed back, the computer device may first obtain the attribute information and / or functional information of the resource to be fed back, thereby determining the resource domain of the resource to be fed back based on the obtained attribute information and / or functional information. After the computer device determines the resource domain of the resource to be fed back, it may obtain the text describing the attributes and / or functions of the resource to be fed back from the resource description text of the resource to be fed back, and use the obtained text describing the attributes and / or functions as the domain keywords of the resource description text in the corresponding resource domain.

[0053] In one embodiment, since search engines in different domains may use different methods for topic extraction, general text search engines may employ TF-IDF (term frequency–inverse document frequency, a commonly used weighting technique for information retrieval and data mining), TextRank (a keyword extraction algorithm), LDA (a topic model), deep learning, etc., to extract text keywords. For example, TF-IDF can extract important words in the text through the global document's term frequency (TF) and inverse document frequency (IDF), ultimately retaining only important words with TF-IDF greater than a certain threshold as training samples. TextRank, LDA, and deep learning calculate text importance using different algorithms, ultimately extracting the core topic of the search document as the training document content. For vertical domains, domain-specific terms can be obtained. For example, in the e-commerce domain, keywords can be extracted and then trained using item attribute words. Item attributes can be identified through a dictionary of terms such as brand, item, model, size, and color. The dictionary can be derived from the product attribute content uploaded by the search object, specifically as follows: Figure 3b As shown in the diagram. In the specific implementation, since the search document includes one or more domain keywords extracted from the resource description text within the corresponding resource domain, when the computer device extracts the domain keywords from the resource description text based on the resource domain corresponding to the resource to be fed back, it can obtain one or more general attributes that need to be referenced when describing the resource within the corresponding resource domain. It can also extract the description text of the resource to be fed back under any of these general attributes from the resource description text and associate each general attribute with its corresponding description text as domain keywords. The search reference data obtained by concatenating the attribute words after topic extraction is shown in Table 3.

[0054] Table 3

[0055]

[0056]

[0057] After obtaining the search reference data and search behavior data, the computer device can train the target search model based on the search reference data and search behavior data, and then proceed to step S203.

[0058] S203, use the search reference data and search behavior data as sample data, and use the sample data to train the target search model to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

[0059] Expanding the training samples of the model by searching documents is feasible when the computer device has limited search behavior data or the target application corresponding to the target search model cannot obtain a sufficient amount of search behavior data. When the computer device has access to a large amount of search behavior data, it can directly use the acquired search behavior data for model training. Alternatively, the computer device can further expand the samples based on the search documents and then train the target search model using the expanded training samples. Traditional methods for selecting a specific target search model include KNN (K-Nearest Neighbor), Naive Bayes classification, and Support Vector Machines. However, with the development of deep learning, deep learning-based text classification models such as RNN (a type of recurrent neural network) and BERT (a language processing model) can achieve higher accuracy. However, deep learning models are generally more complex and time-consuming when used online. Since search involves multiple processes, the time requirement for category prediction cannot be too high, for example, within 10ms. Therefore, to ensure the performance of the search interface in returning the search results, a fast text classification tool based on a shallow neural network can be used. In one embodiment, to improve the model's training accuracy while maintaining its training accuracy, the computer device can further remove the influence of obscure categories by setting minCountLabel in the category prediction scenario, and use softmax as the loss function to estimate the probabilities of multiple categories, which can further improve the accuracy by 5%. Finally, the average time of the Fasttext-based model is controlled within 1ms, ensuring that the online search interface completes category prediction within milliseconds. The time consumption indicator is as follows: Figure 3c As shown. In one embodiment, minCountLabel is used to reduce the number of training samples that do not occur frequently (i.e., less than the target value), thereby effectively reducing the number of training samples while ensuring the model's training accuracy, and thus effectively improving the model's training efficiency.

[0060] After training the target search model, the computer device can test its capabilities. In one embodiment, taking an item search engine (i.e., integrating the target search model into the search engine) as an example, the computer device can split a large product dataset into two parts: 90% as the training set and 10% as the test set. The final trained model achieves a top-1 recall of 94.8% and a top-3 recall of 98.9% on the test set, as shown in Table 4 below.

[0061] Table 4

[0062] index Recall rate TOP1 accuracy 94.8％ TOP3 accuracy 98.9％

[0063] In one embodiment, TOP1 hit refers to the probability that the first-ranked feedback resource among the feedback resources returned by the target search model is selected by the search object, while TOP3 refers to the probability that one of the top three feedback resources returned by the target search model is selected by the search object. Therefore, as shown in Table 4 above, after training the target search model based on the scheme of this application embodiment, the resulting trained search model in the resource recommendation scenario can be illustrated as follows: Figure 3d As shown, it can be seen that the accuracy of search prediction based on the trained target search model is quite good.

[0064] After training the target search model, the computer device can perform input prediction using this trained model. This means it can execute an online search prediction service by calling the target search model; essentially, it provides a category prediction interface to the search engine by loading the trained model file. When a searcher enters a search term, the category prediction service, through model estimation, can obtain the probability values ​​(range 0-1) of the search term corresponding to different categories. For example, if the search term is "car," the computer device can obtain the probabilities of different categories from the target search model as follows: Figure 3e As shown, further, when retrieving search results, the ranking weight of documents in the corresponding categories is increased based on the predicted probability of different categories, displaying content in the category the search object wants. In specific implementation, when the computer device calls the trained target search model to predict the category target tag, it can first obtain the search request sent by the target object, and the search cache information generated after the search prediction processing of the trained target search model; wherein, the search request contains the target search term, the search cache information contains one or more reference search terms that have been processed by the trained target search model, and one or more predicted category target tags obtained by the corresponding reference search terms based on the search prediction processing of the trained target search model, and the corresponding prediction probability; then, the target search term can be matched with one or more reference search terms contained in the search cache information to obtain the matching result, and the target feedback resource corresponding to the target search term can be determined based on the matching result.

[0065] Since the predicted results of the target search model after training are stable for search terms, to improve search performance and reduce the time spent calling the category prediction service, the search category prediction results can be cached locally or on a cache server. The cache key is the search term for the search object, the cache content is the probability under different categories, and the caching time is the interval between model training. When the search term category prediction cache is hit, no category prediction request is initiated, and the cached result is used directly, which can effectively improve search performance. Optionally, since some search terms have multiple synonyms or hyponyms, for example, the search object searches for "books" or "books" less often, but searches for "books" more often. Since "books" have more behavioral data that can better reflect the search object's search intent, the search term can be expanded with synonyms based on word vectors and hyponyms based on knowledge graphs. The expanded words can be used for prediction to improve the category prediction accuracy of uncommon words. A diagram illustrating the expansion of hyponyms based on search terms can be shown below. Figure 3f As shown, hyponyms and hypernyms refer to words that correspond to each other in a hyponym relationship. The hyponym relationship refers to the semantic relationship between a hyponym of a subtype and a hypernym of a parent type. In a semantic field, a hyponym is contained within a hypernym.

[0066] Below, in conjunction with Figure 4 This paper describes the online prediction process based on a target search model. After receiving a search request from a target, the computer device matches the target search term in the search request with one or more reference search terms in the search cache information to obtain a matching result. Based on the matching result, the computer device determines the target feedback resource corresponding to the target search term. Specifically, when determining the target feedback resource based on the matching result, if the matching result indicates that there is a reference search term in the one or more reference search terms in the search cache information that matches the target search term, the computer device retrieves one or more predicted target tags and their corresponding prediction probabilities from the search cache information. Each predicted target tag corresponds to one feedback resource. Further, based on the retrieved predicted target tags and their corresponding prediction probabilities, the feedback resources corresponding to each predicted target tag are sorted in descending order. Based on the descending order, the feedback resources that meet the preset sequence number are selected as the target feedback resources corresponding to the target search term.

[0067] After determining one or more predicted target tags and their corresponding predicted probabilities for a reference search term, the computer device sorts the resources to be fed back for each predicted target tag in descending order based on its predicted probability. Specifically, if the computer device determines predicted target tags 1, 2, and 3 with predicted probabilities of 0.3, 0.5, and 0.2 respectively, then during the descending sort, it can obtain resource a corresponding to target tag 1, resource b corresponding to target tag 2, and resource c corresponding to target tag 3. The resulting sequence, combining these predicted probabilities, could be: resource b, resource a, resource c. After sorting the resources to be fed back in descending order based on the predicted probabilities, the computer device can select the resources that meet the preset sequence number as the target feedback resources. In the aforementioned sequence of resources b, a, and c, resource b has the sequence number 1, resource a has the sequence number 2, and resource c has the sequence number 3. The preset sequence number can be 2. In this case, the computer device can use the resources preceding 2 and the resources corresponding to 2 as target feedback resources. Therefore, in the aforementioned sequence of resources b, a, and c, resources b and a are used as target feedback resources.

[0068] In one embodiment, when the computer device determines the target feedback resource corresponding to the target search term based on the matching result, it can also first call the trained target search model to perform search prediction processing on the target search term if there is no reference search term that successfully matches the target search term among one or more reference search terms included in the search cache information indicated by the matching result, thereby obtaining one or more predicted target tags corresponding to the target search term and their corresponding prediction probabilities; one predicted target tag corresponds to one feedback resource; furthermore, based on the prediction probabilities of each predicted target tag corresponding to the target search term, the target feedback resource corresponding to the target search term can be selected from the corresponding feedback resources of the predicted target tags corresponding to the target search term.

[0069] When a computer device fails to perform search prediction processing on a target search term using a trained target search model, it can perform semantic expansion processing based on the meaning of the target search term to obtain one or more expanded search terms. Semantic expansion processing includes one or more of the following: synonym expansion processing, hypernym expansion processing, and hyponym expansion processing. Furthermore, the computer device can use any of these expanded search terms as a new target search term and call the trained target search model to perform search prediction processing on the new target search term until the search prediction processing is successful, thereby effectively improving the search success rate of the computer device.

[0070] In this embodiment, when training a target search model, if the computer device cannot obtain a sufficient amount of search behavior data, it can acquire the search behavior data and search documents. Based on topic extraction from the search documents, it can construct search reference data. Then, both the search behavior data and the search reference data can be used as training samples for the target search model, effectively expanding the sample data and improving the accuracy of model training. This achieves effective utilization of useful information during model training, enhancing the flexibility and effectiveness of expanding model sample data. Using search object behavior data and topic-extracted text from search documents as training data for the text classification model maximizes the reflection of the search object's search tendency on frequently occurring head search terms, and effectively predicts the search category for mid- and long-tail search terms. This significantly improves the overall accuracy of intent recognition, allowing the computer device to better understand the search object's search intent, avoiding the appearance of irrelevant category documents through text matching, and improving the search experience. This effectively enhances the computer device's search prediction capabilities.

[0071] Based on the description of the above data processing method embodiments, this invention also proposes a data processing apparatus, which can be a computer program (including program code) running on the aforementioned computer device. This data processing apparatus can be used to perform tasks such as... Figure 2 For the data processing method described above, please refer to [link / reference]. Figure 5 The data processing device includes an acquisition unit 501 and a processing unit 502.

[0072] The acquisition unit 501 is used to acquire search behavior data and search documents; the search behavior data includes search behavior terms and category tags corresponding to the search behavior terms, wherein the probability of using the search document to perform a category search is less than the probability of using the search behavior terms to perform a category search.

[0073] The processing unit 502 is used to extract the document topic of the search document and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic;

[0074] The processing unit 502 is further configured to use the search reference data and the search behavior data as sample data, and use the sample data to train the target search model to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

[0075] In one embodiment, the acquisition unit 501 is specifically used for:

[0076] Obtain a corpus of text, which includes historical search text and search results obtained by data search based on the historical search text; the search results include at least one search feedback resource obtained by data search based on the historical search text, and the search execution status of the corresponding search feedback resource;

[0077] The historical search text contained in the corpus is segmented and filtered to obtain the search words corresponding to the historical search text. Based on the search feedback resources and corresponding search execution status included in the search results, the search target of the historical search text is determined.

[0078] The search terms are used as search behavior terms, and the search target is used as a category tag associated with the search behavior terms;

[0079] Search behavior data is generated based on search behavior terms associated with related target tags.

[0080] In one embodiment, the number of search behavior data is multiple, and one search behavior data is obtained by associating a corresponding search behavior term with a corresponding category tag; the acquisition unit 501 is also used to acquire the full set of search behavior terms and the distribution of the full set of search behavior terms, the distribution being used to indicate the frequency with which different search behavior terms are used in resource searches.

[0081] The processing unit 502 is further configured to select search behavior words that meet a preset frequency level from the full set of search behavior words based on the frequency level of each search behavior word indicated by the distribution when performing resource searches.

[0082] The processing unit 502 is further configured to use the selected search behavior terms and associated category tags as search behavior data.

[0083] In one embodiment, the acquisition unit 501 is specifically used for:

[0084] Obtain the resource description text of the resource to be fed back, perform data augmentation processing on the resource description text to obtain the augmented text corresponding to the resource to be fed back, and use the obtained augmented text as a search document; or,

[0085] Obtain the resource description text of the resource to be fed back, and extract the domain keywords of the corresponding resource domain from the resource description text based on the resource domain corresponding to the resource to be fed back, and use the extracted domain keywords as search documents.

[0086] In one embodiment, the search document includes one or more domain keywords extracted from the resource description text, within the corresponding resource domain; the acquisition unit 501 is specifically used for:

[0087] Based on the resource domain corresponding to the resource to be fed back, obtain one or more general attributes that need to be referenced when describing the resource in the corresponding resource domain;

[0088] Extract the description text of the resource to be fed back under any general attribute from the resource description text, and associate each general attribute and the corresponding description text as domain keywords.

[0089] In one embodiment, the acquisition unit 501 is further configured to acquire a search request sent by the target object, and search cache information generated after the target search model has been called for search prediction processing; wherein, the search request includes target search terms, and the search cache information includes one or more reference search terms that have been called for search prediction processing of the target search model, and one or more predicted target tags obtained by the corresponding reference search terms based on the target search model, and the corresponding prediction probabilities;

[0090] The processing unit 502 is further configured to perform matching processing on the target search term and one or more reference search terms contained in the search cache information to obtain matching results, and determine the target feedback resource corresponding to the target search term based on the matching results.

[0091] In one embodiment, the processing unit 502 is specifically used for:

[0092] When the matching result indicates that there is a reference search term that successfully matches the target search term among one or more reference search terms contained in the search cache information, one or more predicted target tags and their corresponding predicted probabilities corresponding to the successfully matched reference search terms are obtained from the search cache information; wherein, one predicted target tag corresponds to one resource to be fed back.

[0093] Based on one or more predicted target tags and their corresponding predicted probabilities, the resources to be fed back corresponding to each predicted target tag are sorted in descending order.

[0094] Based on the order after descending sorting, the resources to be fed back that meet the preset sequence number are taken as the target feedback resources corresponding to the target search term.

[0095] In one embodiment, the processing unit 502 is specifically used for:

[0096] When the matching result indicates that there is no reference search term that successfully matches the target search term among the one or more reference search terms contained in the search cache information, the trained target search model is invoked to perform search prediction processing on the target search term to obtain one or more predicted target tags corresponding to the target search term and their corresponding prediction probabilities; one predicted target tag corresponds to one resource to be fed back.

[0097] Based on the predicted probabilities of each predicted target tag corresponding to the target search term, the target feedback resource corresponding to the target search term is selected from the corresponding feedback resources of the predicted target tags corresponding to the target search term.

[0098] In one embodiment, when the trained target search model fails to perform search prediction processing on the target search term, the processing unit 502 is further configured to perform semantic expansion processing based on the meaning of the target search term to obtain one or more expanded search terms of the target search term; wherein, the semantic expansion processing includes one or more of the following: synonym expansion processing, hypernym expansion processing, and hyponym expansion processing;

[0099] The processing unit 502 is further configured to take any extended search term as a new target search term and call the trained target search model to perform search prediction processing on the new target search term until the search prediction processing of the trained target search model is successfully performed.

[0100] In this embodiment, when training the target search model, if the acquisition unit 501 cannot acquire a sufficient amount of search behavior data, it can acquire the search behavior data and search documents. The processing unit 502 can then construct search reference data based on topic extraction from the search documents. Both the search behavior data and the search reference data can then be used as training samples for the target search model, effectively expanding the sample data and improving the accuracy of model training. This achieves effective utilization of useful information during model training, enhancing the flexibility and effectiveness of expanding model sample data. Using search object behavior data and text extracted from search documents based on topics as training data for the text classification model maximizes the reflection of the search object's search tendency on frequently occurring head search terms, while also effectively predicting the search category for mid- and long-tail search terms. This significantly improves the overall accuracy of intent recognition, allowing for a better understanding of the search object's search intent and avoiding irrelevant category documents resulting from text matching. This enhances the search experience for the search object and effectively improves search prediction capabilities.

[0101] Please see Figure 6 This is a schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Figure 6 The computer device shown in this embodiment may include: one or more processors 601; one or more input devices 602; one or more output devices 603; and a memory 604. The processors 601, input devices 602, output devices 603, and memory 604 are connected via a bus 605. The memory 604 stores a computer program, which includes program instructions, and the processor 601 executes the program instructions stored in the memory 604.

[0102] The memory 604 may include volatile memory, such as random-access memory (RAM); the memory 604 may also include non-volatile memory, such as flash memory, solid-state drive (SSD), etc.; the memory 604 may also include a combination of the above types of memory.

[0103] The processor 601 may be a central processing unit (CPU). The processor 601 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), etc. The PLD may be a field-programmable gate array (FPGA), generic array logic (GAL), etc. The processor 601 may also be a combination of the above structures.

[0104] In this embodiment of the invention, the memory 604 is used to store a computer program, the computer program including program instructions, and the processor 601 is used to execute the program instructions stored in the memory 604 to implement the above-mentioned... Figure 2 The steps of the corresponding method.

[0105] In one embodiment, the processor 601 is configured to invoke the program instructions to execute:

[0106] Acquire search behavior data and search documents; the search behavior data includes search behavior terms and category tags corresponding to the search behavior terms, wherein the probability of using the search documents to perform category searches is less than the probability of using the search behavior terms to perform category searches;

[0107] Extract the document topic of the search document, and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic;

[0108] The search reference data and the search behavior data are used as sample data, and the target search model is trained using the sample data to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

[0109] In one embodiment, the processor 601 is configured to invoke the program instructions to execute:

[0110] Obtain a corpus of text, which includes historical search text and search results obtained by data search based on the historical search text; the search results include at least one search feedback resource obtained by data search based on the historical search text, and the search execution status of the corresponding search feedback resource;

[0111] The historical search text contained in the corpus is segmented and filtered to obtain the search words corresponding to the historical search text. Based on the search feedback resources and corresponding search execution status included in the search results, the search target of the historical search text is determined.

[0112] The search terms are used as search behavior terms, and the search target is used as a category tag associated with the search behavior terms;

[0113] Search behavior data is generated based on search behavior terms associated with related target tags.

[0114] In one embodiment, the number of search behavior data is multiple, and each search behavior data is obtained by associating a corresponding search behavior term with a corresponding category tag; the processor 601 is configured to call the program instructions to execute:

[0115] Obtain the full set of search terms and their distribution, the distribution being used to indicate the frequency with which different search terms are used in resource searches.

[0116] Based on the frequency of each search behavior term indicated by the distribution, search behavior terms that meet the preset frequency are selected from the full set of search behavior terms.

[0117] The selected search behavior terms and associated category tags are used as search behavior data.

[0118] In one embodiment, the processor 601 is configured to invoke the program instructions to execute:

[0119] Obtain the resource description text of the resource to be fed back, perform data augmentation processing on the resource description text to obtain the augmented text corresponding to the resource to be fed back, and use the obtained augmented text as a search document; or,

[0120] Obtain the resource description text of the resource to be fed back, and extract the domain keywords of the corresponding resource domain from the resource description text based on the resource domain corresponding to the resource to be fed back, and use the extracted domain keywords as search documents.

[0121] In one embodiment, the search document includes one or more domain keywords extracted from resource description text within the corresponding resource domain; the processor 601 is configured to invoke the program instructions to execute:

[0122] Based on the resource domain corresponding to the resource to be fed back, obtain one or more general attributes that need to be referenced when describing the resource in the corresponding resource domain;

[0123] Extract the description text of the resource to be fed back under any general attribute from the resource description text, and associate each general attribute and the corresponding description text as domain keywords.

[0124] In one embodiment, the processor 601 is configured to invoke the program instructions to execute:

[0125] The system obtains a search request sent by the target object and search cache information generated after the target search model has been called for search prediction processing; wherein the search request contains target search terms, and the search cache information contains one or more reference search terms that have been called for search prediction processing of the target search model, and one or more predicted target tags obtained by the corresponding reference search terms based on the target search model, and the corresponding prediction probabilities.

[0126] The target search term is matched with one or more reference search terms contained in the search cache information to obtain matching results, and the target feedback resource corresponding to the target search term is determined based on the matching results.

[0127] In one embodiment, the processor 601 is configured to invoke the program instructions to execute:

[0128] When the matching result indicates that there is a reference search term that successfully matches the target search term among one or more reference search terms contained in the search cache information, one or more predicted target tags and their corresponding predicted probabilities corresponding to the successfully matched reference search terms are obtained from the search cache information; wherein, one predicted target tag corresponds to one resource to be fed back.

[0129] Based on one or more predicted target tags and their corresponding predicted probabilities, the resources to be fed back corresponding to each predicted target tag are sorted in descending order.

[0130] Based on the order after descending sorting, the resources to be fed back that meet the preset sequence number are taken as the target feedback resources corresponding to the target search term.

[0131] In one embodiment, the processor 601 is configured to invoke the program instructions to execute:

[0132] When the matching result indicates that there is no reference search term that successfully matches the target search term among the one or more reference search terms contained in the search cache information, the trained target search model is invoked to perform search prediction processing on the target search term to obtain one or more predicted target tags corresponding to the target search term and their corresponding prediction probabilities; one predicted target tag corresponds to one resource to be fed back.

[0133] Based on the predicted probabilities of each predicted target tag corresponding to the target search term, the target feedback resource corresponding to the target search term is selected from the corresponding feedback resources of the predicted target tags corresponding to the target search term.

[0134] In one embodiment, when the trained target search model fails to perform search prediction on the target search term, the processor 601 is configured to invoke the program instructions to execute:

[0135] Based on the meaning of the target search term, semantic expansion processing is performed to obtain one or more expanded search terms of the target search term; wherein, the semantic expansion processing includes one or more of the following: synonym expansion processing, higher-level semantic expansion processing, and lower-level semantic expansion processing;

[0136] Take any extended search term as a new target search term, and call the trained target search model to perform search prediction processing on the new target search term until the search prediction processing of the trained target search model is successful.

[0137] This invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned actions. Figure 2 The method embodiment shown. The computer-readable storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0138] The above-disclosed embodiments are merely partial examples of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that all or part of the processes for implementing the above embodiments, and equivalent variations made in accordance with the claims of the present invention, still fall within the scope of the invention.

Claims

1. A data processing method, characterized in that, include: The system acquires search behavior data and search documents. The search behavior data includes search behavior terms and corresponding category tags, wherein the probability of using the search documents for category search is less than the probability of using the search behavior terms for category search. The methods for acquiring search documents include: acquiring the resource description text of the resource to be fed back, performing data augmentation on the resource description text to obtain augmented text corresponding to the resource to be fed back, and using the obtained augmented text as the search document; or, acquiring the resource description text of the resource to be fed back, extracting domain keywords from the resource description text based on the resource domain corresponding to the resource to be fed back, and using the extracted domain keywords as the search document. Extract the document topic of the search document, and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic; The search reference data and the search behavior data are used as sample data, and the target search model is trained using the sample data to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

2. The method as described in claim 1, characterized in that, The acquisition of search behavior data includes: Obtain a corpus of text, which includes historical search text and search results obtained by data search based on the historical search text; the search results include at least one search feedback resource obtained by data search based on the historical search text, and the search execution status of the corresponding search feedback resource; The historical search text contained in the corpus is segmented and filtered to obtain the search words corresponding to the historical search text. Based on the search feedback resources and corresponding search execution status included in the search results, the search target of the historical search text is determined. The search terms are used as search behavior terms, and the search target is used as a category tag associated with the search behavior terms; Search behavior data is generated based on search behavior terms associated with related target tags.

3. The method as described in claim 1 or 2, characterized in that, The number of search behavior data is multiple, and each search behavior data is obtained by associating a corresponding search behavior term with a corresponding category tag; the method further includes: Obtain the full set of search terms and their distribution, the distribution being used to indicate the frequency with which different search terms are used in resource searches. Based on the frequency of each search behavior term indicated by the distribution, search behavior terms that meet the preset frequency are selected from the full set of search behavior terms. The selected search behavior terms and associated category tags are used as search behavior data.

4. The method as described in claim 1, characterized in that, The search document includes one or more domain keywords extracted from the resource description text, belonging to the corresponding resource domain; the extraction of domain keywords from the resource description text based on the resource domain corresponding to the resource to be fed back includes: Based on the resource domain corresponding to the resource to be fed back, obtain one or more general attributes that need to be referenced when describing the resource in the corresponding resource domain; Extract the description text of the resource to be fed back under any general attribute from the resource description text, and associate each general attribute and the corresponding description text as domain keywords.

5. The method as described in claim 1, characterized in that, The method further includes: The system obtains a search request sent by the target object, and search cache information generated after the target search model has been called for search prediction processing; wherein the search request contains target search terms, and the search cache information contains one or more reference search terms that have been called for search prediction processing of the target search model, and one or more predicted target tags obtained by the corresponding reference search terms based on the target search model, and the corresponding prediction probabilities. The target search term is matched with one or more reference search terms contained in the search cache information to obtain matching results, and the target feedback resource corresponding to the target search term is determined based on the matching results.

6. The method as described in claim 5, characterized in that, The step of determining the target feedback resource corresponding to the target search term based on the matching result includes: When the matching result indicates that there is a reference search term that successfully matches the target search term among one or more reference search terms contained in the search cache information, one or more predicted target tags and their corresponding predicted probabilities corresponding to the successfully matched reference search terms are obtained from the search cache information; wherein, one predicted target tag corresponds to one resource to be fed back. Based on one or more predicted target tags and their corresponding predicted probabilities, the resources to be fed back corresponding to each predicted target tag are sorted in descending order. Based on the order after descending sorting, the resources to be fed back that meet the preset sequence number are taken as the target feedback resources corresponding to the target search term.

7. The method as described in claim 5, characterized in that, The step of determining the target feedback resource corresponding to the target search term based on the matching result includes: When the matching result indicates that there is no reference search term that successfully matches the target search term among the one or more reference search terms contained in the search cache information, the trained target search model is invoked to perform search prediction processing on the target search term to obtain one or more predicted target tags corresponding to the target search term and their corresponding prediction probabilities; one predicted target tag corresponds to one resource to be fed back. Based on the predicted probabilities of each predicted target tag corresponding to the target search term, the target feedback resource corresponding to the target search term is selected from the corresponding feedback resources of the predicted target tags corresponding to the target search term.

8. The method as described in claim 7, characterized in that, When the trained target search model fails to perform search prediction processing on the target search term, the method further includes: Based on the meaning of the target search term, semantic expansion processing is performed to obtain one or more expanded search terms of the target search term; wherein, the semantic expansion processing includes one or more of the following: synonym expansion processing, higher-level semantic expansion processing, and lower-level semantic expansion processing; Take any extended search term as a new target search term, and call the trained target search model to perform search prediction processing on the new target search term until the search prediction processing of the trained target search model is successful.

9. A data processing apparatus, characterized in that, include: The acquisition unit is used to acquire search behavior data and search documents; The search behavior data includes search behavior terms and corresponding category tags, wherein the probability of using the search document for category search is less than the probability of using the search behavior terms for category search. The methods for obtaining the search document include: obtaining the resource description text of the resource to be fed back, performing data augmentation on the resource description text to obtain the augmented text corresponding to the resource to be fed back, and using the obtained augmented text as the search document; or, obtaining the resource description text of the resource to be fed back, extracting domain keywords of the corresponding resource domain from the resource description text based on the resource domain corresponding to the resource to be fed back, and using the extracted domain keywords as the search document. The processing unit is configured to extract the document topic of the search document and generate search reference data based on the search document and the corresponding document topic; the category tags associated with the search document are used as the category tags of the document topic; The processing unit is further configured to use the search reference data and the search behavior data as sample data, and use the sample data to train the target search model to obtain the trained target search model; the trained target search model is used to predict the correspondence probability between the input search term and different types of target tags.

10. A computer device, characterized in that, The system includes a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 8.

12. A computer program product comprising computer instructions stored in a computer-readable storage medium; a processor of a computer device reading the computer instructions from the computer-readable storage medium, the processor executing the computer instructions to cause the computer device to perform the method as described in any one of claims 1 to 8.

Images