Intention recognition method and device

Abstract

This application provides an intent recognition method and apparatus. The method includes: determining the target title of at least one target search result from the titles of at least one search result; performing a first feature extraction process on the search term to obtain search term features; performing a second feature extraction process on the target title of the at least one target search result to obtain target title features; performing entity recognition processing on one or more of the search term and target title to obtain entity features; fusing the search term features, target title features, and entity features to obtain fused features; predicting the search intent corresponding to the search term based on the fused features to determine the target intent category corresponding to the search term; and determining the target content corresponding to the target intent category. Predicting search intent based on the fused features of search term features, target title features, and entity features improves the accuracy of intent recognition.

Description

[0001] This application is a divisional application of Chinese Patent Application No. 202110951767.2, filed with the Chinese Patent Office on August 18, 2021, entitled "Intent Recognition Method and Apparatus", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of neural network technology, and in particular to an intention recognition method and apparatus. Background Technology

[0003] In search scenarios, users enter search terms (queries) into search engines. Accurately identifying the user's search intent can improve search accuracy.

[0004] Therefore, how to improve the accuracy of intent recognition is the problem that this application needs to solve. Summary of the Invention

[0005] This application provides an intent recognition method and apparatus to improve the accuracy of intent recognition.

[0006] Firstly, an intent recognition method is provided, the method comprising:

[0007] In response to a user's triggering action on the title of at least one search result in a first search results page, a target title for at least one target search result is determined from the titles of the at least one search result; wherein the first search results page is used to display the title of at least one search result obtained based on a first search term;

[0008] Perform a first feature extraction process on the first search term to obtain the search term features of the first search term;

[0009] The target title of the at least one target search result is subjected to a second feature extraction process to obtain the target title feature;

[0010] Entity recognition processing is performed on one or more of the first search term and the target title to obtain entity features;

[0011] The search term features, the target title features, and the entity features are fused together to obtain fused features;

[0012] Based on the fusion features, the search intent corresponding to the first search term is predicted, and the target intent category corresponding to the first search term is determined;

[0013] Determine the target content corresponding to the target intent category, the target content being displayed in the search results page associated with the first search term.

[0014] In one possible implementation, determining the target title of at least one target search result from the titles of at least one search result in response to a user's triggering action on the title of at least one search result in a first search result page includes:

[0015] In response to a user's triggering action on the title of at least one search result in the first search results page, the click-through rate of the title of the at least one search result is obtained, wherein the click-through rate of the title of each search result in the at least one search results is: the ratio of the number of clicks on the title of each search result by multiple users within a set time period to the total number of clicks on the title of the at least one search result by the multiple users, wherein the multiple users are users who input the first search term and information associated with the first search term;

[0016] The click-through rates of the titles of the at least one search result are sorted in descending order.

[0017] Based on the ranking of the click-through rate of the titles of the at least one search result, the titles of at least one search result whose click-through rate is less than or equal to a set value are selected as the target titles of at least one target search result.

[0018] In another possible implementation, the second feature extraction process on the target title of the at least one target search result to obtain target title features includes:

[0019] Feature extraction is performed on the target title of the at least one target search result to obtain the text feature vector of the target title of the at least one target search result;

[0020] Using the click-through rate of the target title of the at least one target search result as a weight, the text feature vectors of the target titles of the at least one target search result are weighted and summed to obtain the target title features.

[0021] In another possible implementation, the entity recognition processing of one or more of the first search term and the target title to obtain entity features includes:

[0022] Identify at least one entity from one or more of the first search term and the target title to obtain an entity set;

[0023] Determine at least one intent category to which each entity in the entity set belongs;

[0024] Obtain the prior probability of each entity to each intent category. The prior probability is historical statistical information on the probability of an entity to its intent category. The prior probability of each entity to each intent category is the ratio of the first occurrence count of each entity belonging to each intent category to the total number of times each entity belonging to at least one intent category appears in one or more of the first search term and the target title. Wherein, the first occurrence count of each entity belonging to each intent category is the number of times each entity belonging to each intent category appears in one or more of the first search term and the target title.

[0025] Entity features corresponding to the entity set are generated based on the prior probability of each intent category to which each entity belongs.

[0026] In another possible implementation, generating entity features corresponding to the entity set based on the prior probability of each intent category to which each entity belongs includes:

[0027] The frequency of each entity in the entity set is obtained, wherein the frequency of each entity is the ratio of the number of times each entity appears in one or more of the first search term and the target title to the total number of times the entity set appears in one or more of the first search term and the target title.

[0028] Using the frequency of occurrence of each entity in the entity set as the weight, the prior probabilities of each intent category to which each entity in the entity set belongs are weighted and summed to obtain the entity features corresponding to the entity set.

[0029] In yet another possible implementation, the method further includes:

[0030] The prior probability of each entity in the entity set to each intent category is reduced in dimensionality, and the resulting prior probability dimension of each entity in the entity set to each intent category is consistent with the number of the target intent categories.

[0031] In another possible implementation, the fusion processing of the search term features, the target title features, and the entity features to obtain fused features includes:

[0032] The vectors of the search term features, the target title features, and the entity features are concatenated to obtain the fused features.

[0033] In another possible implementation, predicting the search intent corresponding to the first search term based on the fusion features and determining the target intent category corresponding to the first search term includes:

[0034] The fusion features are then subjected to dimensionality reduction processing, and the dimensionality of the resulting fusion features is consistent with the number of target intent categories.

[0035] Based on the fusion features obtained after the processing, the classification probability corresponding to at least one intent category is obtained;

[0036] The classification probabilities corresponding to the at least one intent category are normalized.

[0037] Among the classification probabilities corresponding to at least one intent category obtained after normalization, the intent category with the highest classification probability is determined as the target intent category.

[0038] Secondly, an intent recognition device is provided, the device comprising:

[0039] A first determining unit is configured to, in response to a user's triggering operation on the title of at least one search result in a first search result page, determine the target title of at least one target search result from the titles of the at least one search result; wherein the first search result page is configured to display the title of at least one search result obtained based on a first search term;

[0040] The first feature extraction unit is used to perform a first feature extraction process on the first search term to obtain the search term features of the first search term;

[0041] The second feature extraction unit is used to perform second feature extraction processing on the target title of the at least one target search result to obtain target title features;

[0042] The third feature extraction unit is used to perform entity recognition processing on one or more of the first search term and the target title to obtain entity features;

[0043] The feature fusion unit is used to fuse the search term features, the target title features, and the entity features to obtain fused features;

[0044] The prediction unit is used to predict the search intent corresponding to the first search term based on the fusion features, and determine the target intent category corresponding to the first search term;

[0045] The second determining unit is used to determine the target content corresponding to the target intent category, and the target content is used to be displayed in the search results page associated with the first search term.

[0046] In one possible implementation, the first determining unit includes:

[0047] The first acquisition unit is configured to, in response to a user's triggering operation on the title of at least one search result in the first search results page, acquire the click-through rate of the title of the at least one search result, wherein the click-through rate of the title of each search result in the at least one search result is: the ratio of the number of clicks on the title of each search result by multiple users within a set time period to the total number of clicks on the title of the at least one search result by the multiple users, wherein the multiple users are users who input the first search term and information associated with the first search term;

[0048] A sorting unit is used to sort the click-through rates of the titles of the at least one search result in descending order.

[0049] A filtering unit is configured to filter the titles of at least one search result whose click-through rate (CTR) is less than or equal to a set value, based on the ranking result of the CTR of the titles of the at least one search result, and use them as the target titles of at least one target search result.

[0050] In yet another possible implementation, the second feature extraction unit includes:

[0051] The second acquisition unit is used to extract features from the target titles of the at least one target search result to obtain the text feature vector of the target title of the at least one target search result.

[0052] The first weighted summation unit is used to perform a weighted summation of the text feature vectors of the target titles of the at least one target search result, using the click-through rate of the target title of the at least one target search result as a weight, to obtain the target title features.

[0053] In yet another possible implementation, the third feature extraction unit includes:

[0054] The identification unit is used to identify at least one entity from one or more of the first search term and the target title to obtain an entity set;

[0055] The third determining unit is used to determine at least one intent category to which each entity in the entity set belongs;

[0056] The third acquisition unit is used to acquire the prior probability of each entity to each intent category. The prior probability is historical statistical information on the probability of an entity to an intent category. The prior probability of each entity to each intent category is the ratio of the first occurrence count of each entity belonging to each intent category to the total number of times each entity belonging to at least one intent category appears in one or more of the first search term and the target title. Wherein, the first occurrence count of each entity belonging to each intent category is the number of times each entity belonging to each intent category appears in one or more of the first search term and the target title.

[0057] The generation unit is used to generate entity features corresponding to the entity set based on the prior probability of each intent category to which each entity belongs.

[0058] In yet another possible implementation, the generation unit includes:

[0059] The fourth acquisition unit is used to acquire the frequency of occurrence of each entity in the entity set, wherein the frequency of occurrence of each entity is the ratio of the number of times each entity appears in one or more of the first search term and the target title to the total number of times the entity set appears in one or more of the first search term and the target title.

[0060] The second weighted summation unit is used to perform weighted summation on the prior probabilities of each intent category to which each entity in the entity set belongs, using the frequency of occurrence of each entity in the entity set as the weight, to obtain the entity features corresponding to the entity set.

[0061] In yet another possible implementation, the device further includes:

[0062] The first dimensionality reduction processing unit is used to perform dimensionality reduction processing on the prior probability of each entity to each intent category in the entity set. The dimension of the prior probability of each entity to each intent category in the entity set after processing is consistent with the number of target intent categories.

[0063] In another possible implementation, the feature fusion unit is used to concatenate the vectors of the search term features, the target title features, and the entity features to obtain the fused features.

[0064] In yet another possible implementation, the second determining unit includes:

[0065] The second dimensionality reduction processing unit is used to perform dimensionality reduction processing on the fused features, and the dimension of the fused features obtained after processing is consistent with the number of target intent categories.

[0066] The fifth acquisition unit is used to obtain the classification probability corresponding to at least one intent category based on the fusion features obtained after the processing.

[0067] A normalization processing unit is used to normalize the classification probabilities corresponding to the at least one intent category;

[0068] The fourth determining unit is used to determine the intent category with the highest classification probability among the classification probabilities corresponding to at least one intent category obtained after normalization processing, and use it as the target intent category.

[0069] Thirdly, an intent recognition device is provided, including an input device and an output device, and further comprising:

[0070] A processor, adapted to implement one or more instructions; and,

[0071] A computer storage medium storing one or more instructions adapted to be loaded by the processor and executed by the method of the first aspect or any implementation thereof.

[0072] Fourthly, a computer storage medium is provided, the computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded by a processor and executed by the method described in the first aspect or any implementation thereof.

[0073] The solution provided in this application has the following advantages:

[0074] The fusion of search term features, target title features, and entity features predicts search intent, improving the accuracy of intent recognition. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a flowchart illustrating an intent recognition method provided in an embodiment of this application;

[0077] Figure 2 This is a flowchart illustrating another intent recognition method provided in an embodiment of this application;

[0078] Figure 3 This is a schematic diagram of an intent recognition model provided in an embodiment of this application;

[0079] Figure 4 This is a schematic diagram of the structure of an intent recognition device provided in an embodiment of this application;

[0080] Figure 5 This is a schematic diagram of another intent recognition device provided in an embodiment of this application. Detailed Implementation

[0081] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0082] As disclosed in this application, the relevant information and data can be stored on a blockchain.

[0083] like Figure 1 The diagram shown is a flowchart illustrating an intent recognition method provided in an embodiment of this application. Exemplarily, the method may include the following steps:

[0084] 101. In response to a user's triggering action on the title of at least one search result in a first search results page, determine the target title of at least one target search result from the titles of at least one search result; wherein the first search results page is used to display the title of at least one search result obtained based on a first search term.

[0085] In a search scenario, a user enters their first search term into the search box of a search engine. This first search term can be any text content, including characters, words, sentences, etc. For example, entering "Sito online viewing".

[0086] After entering the first search term, at least one search result title will be displayed on the first search results page. Search results can include content resources such as articles and images. Each search result has a title. This title serves as a summary of the search results and contains key features. Clicking on the title can further display at least one search result on the first search results page, or request the search result corresponding to that title from a third-party server.

[0087] The titles of at least one search result displayed on the first search results page, based on the first search term, correspond to the content the user wants to search for. The user can trigger an action on all or part of the titles of at least one search result. Furthermore, different users entering the first search term and similar search terms may trigger actions on the same or different search result titles. Therefore, in response to a user's trigger action on the titles of at least one search result on the first search results page, a target title for at least one target search result can be determined from the titles of at least one search result. This target title of the target search result can reflect the user's search intent.

[0088] 102. Perform the first feature extraction process on the first search term to obtain the search term features of the first search term.

[0089] The first search term is what the user enters into the search engine, directly reflecting what the user wants to search for and their search intent. Therefore, the first search term can be processed to extract its first feature, resulting in a search term feature. This search term feature is a vector containing information extracted from the first search term.

[0090] 103. Perform second feature extraction processing on the target title of at least one target search result to obtain the target title feature.

[0091] After determining the target title of at least one target search result, the target title not only includes rich textual features but also user-triggered action features, which can further reflect the user's search intent. Therefore, a second feature extraction process can be performed on the target title of at least one target search result to obtain the target title feature. This target title feature is a vector containing information extracted from the target title.

[0092] 104. Perform entity recognition processing on one or more of the first search term and the target title to obtain entity features.

[0093] Besides the primary search term and target title, the entities within these terms are also crucial information reflecting user intent. For example, the "Sito" in "Sito TV series ending" and the "Great Ruler" in "The Great Ruler novel" contain important information about the user's intent. Therefore, entity features were introduced into the model.

[0094] Entities refer to common people, movies, songs, places, organizations, etc. Specifically, the first search term can contain entities, the target title can also contain entities, or both the first search term and the target title can contain entities.

[0095] In the preprocessing stage, entity recognition processing is performed on one or more of the first search term and the target title to identify the entities of one or more of the first search term and the target title.

[0096] The identified entities can then be processed to obtain entity features. These entity features are vectors containing information extracted from the entities.

[0097] 105. The search term features, target title features, and entity features are fused to obtain fused features.

[0098] By fusing the extracted search term features, target title features, and entity features, the user's search intent can be comprehensively reflected from the dimensions of search terms, target titles, and entities.

[0099] 106. Based on the fusion features, predict the search intent corresponding to the first search term and determine the target intent category corresponding to the first search term.

[0100] For the first search term entered by a user, there can be one or more intent categories. These intent categories can be derived based on content, entities, etc. For example, if a user enters the first search term "Jiang Ziya," "Jiang Ziya" can be considered an entity, and based on this entity, it can be divided into two intent categories: querying for movies and querying for people.

[0101] However, based on search term features, target title features, and entity features, an intent category, namely the target intent category, can be uniquely predicted. This fused feature comprehensively, fully, and accurately reflects the user's search intent. Therefore, predicting the search intent corresponding to the first search term based on the fused feature can accurately determine the target intent category corresponding to the first search term.

[0102] For example, fused features can be input into a classifier to obtain the classification result of the search intent. This classification result specifically includes the classification probability corresponding to each intent category. For instance, if a user inputs the first search term "Jiang Ziya," and the intent categories include: querying the movie "Jiang Ziya" and querying the character "Jiang Ziya," where the classification probability for querying the movie "Jiang Ziya" is 90%, and the classification probability for querying the character "Jiang Ziya" is 10%, then the obtained search intent classification result is querying the movie "Jiang Ziya," meaning the target intent category is querying the movie "Jiang Ziya."

[0103] 107. Determine the target content that corresponds to the target intent category.

[0104] After determining the target intent category, the target content corresponding to the target intent category can be obtained. This target content can then be displayed on the search results page associated with the first search term. For example, the target content can be displayed on the first search results page after the user enters the first search term; alternatively, the target content can be displayed on the second search results page after the user, based on the determined target intent category, enters a similar search term (a modified search term) to the first search term.

[0105] By combining the above steps, an accurate intent recognition model can be trained.

[0106] The search process generally includes three stages: search term understanding, results page recall, and results page ranking. Among these, intent recognition is a crucial step in search term understanding. Taking a specific search as an example, if a user searches for "Sito online viewing," the intent recognition model trained in this embodiment will determine that the target intent category corresponding to the first search term is "watching TV series." The subsequent results page recall stage receives the target intent category and recalls video resources related to Sito TV series, such as "Sito Aladdin Big Card" and "Sito Short Videos." Finally, the results page ranking stage sorts the recalled resources, placing more important resources at the top of the results page. For example, the "Sito TV series" big card will appear at the top of the results page, allowing users to directly click on it to watch the TV series.

[0107] An intent recognition method provided in the embodiments of this application predicts search intent based on the fusion features of search term features, target title features, and entity features, thereby improving the accuracy of intent recognition.

[0108] Another embodiment provides an intent recognition method based on rule templates. This method manually compiles corresponding rule templates based on representative search term data for each intent category, and then matches the user's input search terms against these templates to determine the target intent category (e.g., querying the weather, querying for movies). This approach has high accuracy, but requires manual data collection and rule formulation. Finding data corresponding to the intent category from a large amount of data is very labor-intensive, and template generation relies on prior human knowledge. Furthermore, rule template-based solutions can only cover popular search term types within each intent category, resulting in low recall. In contrast, the intent recognition method in this embodiment does not require manual data collection and rule formulation, can accurately determine the target intent category, and has a high recall rate.

[0109] like Figure 2 The diagram shown is a flowchart of another intent recognition method provided in an embodiment of this application. Exemplarily, the method may include the following steps:

[0110] 201. In response to a user's trigger action on the title of at least one search result in the first search results page, obtain the click-through rate of the title of at least one search result.

[0111] In a search scenario, a user enters their first search term into the search box of a search engine. This first search term can be any text content, including characters, words, sentences, etc. For example, entering "Sito online viewing".

[0112] After entering the first search term, at least one search result title will be displayed on the first search results page. Search results can include content resources such as articles and images. Each search result has a title. This title serves as a summary of the search results and contains key features. Clicking on the title can further display at least one search result on the first search results page, or request the search result corresponding to that title from a third-party server.

[0113] The first search results page displays the titles of at least one search result obtained based on the first search term, corresponding to the content the user wants to search for. After the titles of at least one search result are entered on the first search results page, the user can click on the title that interests them.

[0114] Users clicking on titles they are interested in reflects their search intent. Specifically, the click-through rate (CTR) of the title of each search result within at least one search result can be obtained. The CTR of the title of each search result within at least one search result is the ratio of the number of clicks made by multiple users on the title of each search result within a set time period to the total number of clicks made by multiple users on the title of at least one search result. The multiple users are those who entered a first search term and information associated with that first search term.

[0115] For example, it is possible to collect or store the number of clicks made by all users on at least one search result title for each title within a set time period (e.g., 7 days) after all users enter a first search term and related information on the same search engine. For example, the first search term can be denoted as q, and the click-through rate of each title in the corresponding at least one search result title can be denoted as a set. Among them, t i This represents the i-th title. The click-through rate (CTR) for the i-th title is calculated as shown in Formula 1 below:

[0116]

[0117] in, This represents the total number of clicks made by all users on the i-th title within a specified time period; This represents the total number of clicks made by all users on the title of at least one search result.

[0118] For example, after entering the first search term, the first search results page displays the titles of 10 search results. If, over 7 days, all users click on these 10 search results titles 100, 200, 300, 500, 100, 800, 600, 1000, 200, and 300 respectively, then the click-through rate of the first title is: 100 / (100+200+300+500+100+800+600+1000+200+300) = 2.44%; and so on.

[0119] 202. Sort the click-through rates of the titles of at least one search result in descending order.

[0120] In the actual model, assuming the first search results page displays the titles of M search results, we can select only the K titles ranked by click-through rate as the target title input for the model, where K ≤ M, and K and M are positive integers. Therefore, after obtaining the click-through rate of the title of each search result in at least one search result, we can sort the click-through rates of the titles of at least one search result in descending order.

[0121] 203. Based on the ranking of the click-through rate of the titles of at least one search result, filter the titles of at least one search result whose click-through rate is less than or equal to a set value, and use them as the target titles of at least one target search result.

[0122] After sorting the click-through rates (CTRs) of at least one search result title in descending order, the titles of at least one search result with a CTR ranking less than or equal to a set value are selected as the target titles for at least one target search result. For example, assuming the first search results page displays the titles of M search results, the titles ranked among the top K by CTR are selected as the target titles for at least one target search result.

[0123] 204. Obtain the text feature vector of the first search term to obtain the search term features of the first search term.

[0124] The first search term, being a short text, can be used for feature extraction using existing pre-trained text extraction models.

[0125] For example, a pre-trained text extraction model can be used for feature extraction. This text extraction model can be, for example, any of the following: a bidirectional encoder representations from transformers (BERT) model, a robustly optimized BERT pretraining approach (RoBERTa), or a generative pre-training (GPT) model. The RoBERTa model is an upgraded version of BERT. For example, the formal representation of the first search term can be the word sequence q = w1w2…w L Where L represents the length of the first search term, w i Let represent the i-th character. The model output is represented as a d-dimensional feature vector Q∈R. d That is, the search term characteristics of the first search term.

[0126] like Figure 3 The intent recognition model shown takes a first search term as input, performs feature extraction through the pre-trained model, and outputs the search term feature vector Q of the first search term.

[0127] 205. Extract features from the target title of the at least one target search result to obtain a text feature vector of the target title of the at least one target search result.

[0128] The target title is also a type of short text. After determining the target title of at least one target search result, the pre-trained text extraction model described above can be used to extract features from the target title of each target search result to obtain the text feature vector of that target title.

[0129] For example, for any target search result's target title t i The text extraction model outputs a d-dimensional feature vector T. i ∈R d That is, the text feature vector of the target title of the target search result.

[0130] like Figure 3 The intent recognition model shown extracts features from the target titles of K target search results using a text extraction model, outputting title1 to title2. K Text feature vectors T1~T K .

[0131] 206. Using the click-through rate of the target title of at least one target search result as the weight, perform a weighted summation of the text feature vectors of the target titles of at least one target search result to obtain the target title features.

[0132] When obtaining the text feature vector of the target title of at least one target search result, considering that at least one target title is used, and the importance of each target title is not the same, and there may be target titles that users accidentally click on. Therefore, this embodiment can also consider an attention mechanism to highlight important target titles and reduce the impact of unimportant or accidentally clicked target titles. Therefore, the text feature vectors of the target titles of at least one target search result are weighted and summed using the click-through rate of the target titles of at least one target search result to obtain the target title features. The calculation method is shown in Formula 2 below:

[0133]

[0134] in, T represents the click-through rate corresponding to the i-th target title. i Let T represent the text feature vector of the i-th target title, where T is the target title feature.

[0135] like Figure 3 The intent recognition model shown applies to titles1 to titles... K Text feature vectors T1~T K After weighting, the target title feature T is obtained.

[0136] 207. Identify at least one entity from one or more of the first search term and the target title to obtain an entity set.

[0137] Besides the primary search term and target title, entities within these terms are also important information reflecting user intent. For example, the entity "Sito" in "Sito TV series ending" and "Da Zhuzai" in "the novel Da Zhuzai" contain crucial information about the user's intent. Therefore, entity features were introduced into the model.

[0138] Entities refer to common people, movies, songs, places, organizations, etc. Specifically, the first search term can contain entities, the target title can also contain entities, or both the first search term and the target title can contain entities.

[0139] In the preprocessing stage, entity recognition processing is performed on one or more of the first search term and the target title to identify the entities of one or more of the first search term and the target title, thus obtaining an entity set.

[0140] For example, a pre-trained entity recognition model can be used to identify one or more entities from the first search term and the target title to obtain an entity set.

[0141] 208. Determine at least one intent category to which each entity in the entity set belongs.

[0142] For the first search term entered by a user, there can be one or more intent categories. These intent categories can be derived based on content, entities, etc. For example, if a user enters the first search term "Jiang Ziya," "Jiang Ziya" can be considered an entity, and based on this entity, it can be divided into two intent categories: querying for movies and querying for people.

[0143] After obtaining the entity set, determine at least one intent category to which each entity in the entity set belongs.

[0144] 209. Obtain the prior probability of each intent category to which each entity belongs.

[0145] Unlike text, entities can be ambiguous. For example, the term "Jiang Ziya" can refer to both the movie "Jiang Ziya" and the person "Jiang Ziya," making it difficult to determine whether a user wants to search for the movie or the person using a single entity. This embodiment introduces prior probabilities for entities. These prior probabilities are historical statistics on the probability of an entity belonging to a particular intent category. The prior probability for each intent category to which each entity belongs is calculated as the ratio of the first occurrence count of each entity belonging to each intent category to the total number of times each entity belonging to at least one intent category appears in one or more of the first search term and target title. The first occurrence count of each entity belonging to each intent category is the number of times each entity belonging to each intent category appears in one or more of the first search term and target title.

[0146] For example, the prior probability of entity e can be formally represented as:

[0147]

[0148] Where N is the number of intent categories to which entity e belongs. It is the prior probability that entity e belongs to category i, calculated as shown in Formula 3 below:

[0149]

[0150] Among them, c iThis represents the number of times entity e, belonging to category i, appears in one or more of the first search term and target title. For example, if the movie "Jiang Ziya" appears 9 times and the character "Jiang Ziya" appears once, then the prior probability that Jiang Ziya belongs to the movie category is 0.9, and the prior probability that he belongs to the character category is 0.1. Prior probability, as a type of statistically derived prior information, can improve the accuracy of intent recognition.

[0151] Assuming the entity set comprises M entities, we can obtain the prior probability of each entity belonging to each intent category, denoted as .

[0152] 210. Generate entity features corresponding to the entity set based on the prior probability of each intent category to which each entity belongs.

[0153] After obtaining the prior probability of each entity in the entity set to each intent category, entity features corresponding to the entity set can be generated based on the prior probability of each entity to each intent category.

[0154] If the prior probability of each intent category to which each entity in the entity set belongs is directly used as the entity feature corresponding to the entity set and fused with the search term feature and target title feature, the influence of the search term feature and target title feature will be reduced due to the high dimensionality of the prior probability. Therefore, further, after step 209 and before step 210, the following can also be included:

[0155] The prior probability of each entity in the entity set to each intent category is reduced in dimensionality. After the reduction, the dimension of the prior probability of each entity in the entity set to each intent category is consistent with the number of target intent categories.

[0156] For example, the prior probability of each entity in the entity set to each intent category can be reduced in dimensionality according to Formula 4, and the pre-classification result of the prior probability of each entity in the entity set to each intent category is as follows:

[0157]

[0158] Among them, W H Let b be a trainable parameter matrix, which is a regular linear transformation matrix. H For bias vectors, For pre-classification features, C is the number of target intent categories.

[0159] After the dimensionality reduction process described above, the prior probability of each entity in the entity set belonging to each intent category is obtained. The dimensions are consistent with the number of target intent categories.

[0160] Furthermore, different entities appear at different frequencies in one or more of the first search term and the target title, and their importance also varies. This embodiment considers an attention mechanism during entity feature extraction.

[0161] Specifically, step 210 may include:

[0162] A1. Obtain the frequency of occurrence of each entity in the entity set.

[0163] The frequency of each entity is defined as the ratio of the number of times each entity appears in one or more of the first search term and target title to the total number of times the entity set appears in one or more of the first search term and target title.

[0164] For example, assuming the entity set includes M entities e, the frequency of occurrence of each of the M entities can be formally represented as a set. Among them, e i Represents the i-th entity. The frequency of the i-th entity e is represented by the following formula 5:

[0165]

[0166] in, This indicates the number of times the i-th entity e appears in one or more of the first search term and the target title; This represents the total number of times the entity set appears in one or more of the first search term and the target title.

[0167] A2. Using the frequency of each entity in the entity set as the weight, the prior probabilities of each intent category to which each entity in the entity set belongs are weighted and summed to obtain the entity features corresponding to the entity set.

[0168] For example, the prior probabilities of each intent category to which each entity belongs in the entity set are weighted and summed to obtain the entity features corresponding to the entity set as shown in Formula 6 below:

[0169]

[0170] in, This represents the frequency of the i-th entity e in the entity set; E represents the prior probability of each entity in the entity set after dimensionality reduction to belong to each intent category; E represents the entity feature corresponding to the entity set.

[0171] like Figure 3 The intent recognition model shown extracts entities entity1 to entity2. KWhen considering the corresponding entity features, prior entity probabilities are introduced to obtain the prior probabilities E1 to E2 of each intent category to which each entity in the entity set belongs. K Then, an attention mechanism is used to calculate the prior probabilities E1 to E2 of each intent category to which each entity in the entity set belongs. K We perform a weighted summation to obtain the entity feature E corresponding to the entity set.

[0172] 211. Concatenate the vectors of search term features, target title features, and entity features to obtain the fused features.

[0173] For example, the vectors of search term features, target title features, and entity features are concatenated to obtain a fused feature. This fused feature serves as the input feature for the final classification module, and is formally denoted as...

[0174] 212. Perform dimensionality reduction on the fused features so that the dimensionality of the fused features obtained after the process is consistent with the number of target intent categories.

[0175] For the post-concatenation operation, a fully connected layer can be used to change the dimensionality of the fused features to the number of intent categories. The fully connected layer is represented by the following formula 7:

[0176] F′=W F F+b F …….Formula 7

[0177] Among them, W F Let b be a trainable parameter matrix. F Let F be the bias vector, F be the fused feature, and F′ be the fused feature after dimensionality reduction.

[0178] 213. Based on the fusion features obtained after processing, obtain the classification probability corresponding to at least one intent category.

[0179] By inputting the fused feature F′ obtained after dimensionality reduction into the final classification module, the classification probability corresponding to at least one intent category can be obtained.

[0180] For example, the probability of finding the category for the movie "Jiang Ziya" is 0.9, the probability of finding the category for the character "Jiang Ziya" is 0.5, and the probability of finding the category for the TV series "Jiang Ziya" is 0.4.

[0181] 214. Normalize the classification probability corresponding to at least one intent category.

[0182] The sum of the classification probabilities corresponding to at least one of the above intent categories is not equal to 1, which is not conducive to intuitively distinguishing the intent category with the highest classification probability. Therefore, a regression model (softmax) can be used to normalize the classification probabilities corresponding to at least one intent category according to the following formula 8 to obtain the final classification probability distribution:

[0183] P = softmax(F′)…….Equation 8

[0184] in,

[0185]

[0186] P i This represents the classification probability corresponding to the i-th intent category after normalization.

[0187] 215. Among the classification probabilities corresponding to at least one intent category obtained after normalization, determine the intent category with the highest classification probability as the target intent category.

[0188] Among the classification probabilities corresponding to at least one intent category obtained after normalization, the intent category with the highest classification probability is determined as L = argmax(P), which is taken as the target intent category.

[0189] like Figure 3 The intent recognition model shown concatenates the target title feature T, search term feature Q, and entity feature E, and then performs dimensionality reduction and normalization to obtain the intent category L with the highest probability.

[0190] 216. Determine the target content corresponding to the target intent category. The target content is used to display in the search results page associated with the first search term.

[0191] After determining the target intent category, the target content corresponding to the target intent category can be obtained. This target content can then be displayed on the search results page associated with the first search term. For example, the target content can be displayed on the first search results page after the user enters the first search term; alternatively, the target content can be displayed on the second search results page after the user, based on the determined target intent category, enters a similar search term (a modified search term) to the first search term.

[0192] By combining the above steps, an accurate intent recognition model can be trained.

[0193] An intent recognition method provided in the embodiments of this application predicts search intent based on the fusion features of search term features, target title features, and entity features, thereby improving the accuracy of intent recognition.

[0194] Most search terms contain limited contextual information, while search result titles not only reflect user click characteristics but also contain richer textual features. Therefore, introducing target title features further improves the accuracy of intent recognition.

[0195] Another embodiment provides an alternative intent recognition method based on deep learning models. This type of method treats intent recognition as a classification model, training the deep learning model with a large amount of manually labeled data. The model typically uses a recurrent neural network (RNN) based approach. Deep learning-based methods are more advanced, have a wider range of applications, and higher recall rates compared to template-based methods. However, obtaining high-quality training data is one of the challenges. Secondly, most models only extract features from the search terms themselves. Search terms, as short texts, have few features and are prone to ambiguity, leading to less than ideal model performance. Most entities are ambiguous, especially in the entertainment field such as novels and films. The intent recognition model trained in this embodiment can disambiguate entities using prior probabilities and incorporate them as features, effectively introducing prior knowledge into the model and further improving the accuracy of intent recognition.

[0196] Based on the same concept as the above intent recognition methods, such as Figure 4 As shown in the figure, this application embodiment also provides an intent recognition device 400, which includes: a first determination unit 401, a first feature extraction unit 402, a second feature extraction unit 403, a third feature extraction unit 404, a feature fusion unit 405, a prediction unit 406, and a second determination unit 407; it may also include a first dimensionality reduction processing unit 408 (shown and connected by dashed lines in the figure). Wherein:

[0197] The first determining unit 401 is configured to determine the target title of at least one target search result from the titles of at least one search result in response to a user's triggering operation on the title of at least one search result in the first search result page; wherein the first search result page is configured to display the title of at least one search result obtained based on the first search term;

[0198] The first feature extraction unit 402 is used to perform a first feature extraction process on the first search term to obtain the search term features of the first search term;

[0199] The second feature extraction unit 403 is used to perform second feature extraction processing on the target title of the at least one target search result to obtain target title features;

[0200] The third feature extraction unit 404 is used to perform entity recognition processing on one or more of the first search term and the target title to obtain entity features;

[0201] The feature fusion unit 405 is used to fuse the search term features, the target title features, and the entity features to obtain fused features;

[0202] Prediction unit 406 is used to predict the search intent corresponding to the first search term based on the fusion features and determine the target intent category corresponding to the first search term;

[0203] The second determining unit 407 is used to determine target content corresponding to the target intent category, and the target content is used to be displayed in the search results page associated with the first search term.

[0204] In one possible implementation, the first determining unit 401 includes:

[0205] The first acquisition unit 4011 is configured to, in response to a user's triggering operation on the title of at least one search result in the first search results page, acquire the click-through rate of the title of the at least one search result, wherein the click-through rate of the title of each search result in the at least one search result is: the ratio of the number of clicks on the title of each search result by multiple users within a set time period to the total number of clicks on the title of the at least one search result by the multiple users, wherein the multiple users are users who input the first search term and information associated with the first search term;

[0206] The sorting unit 4012 is used to sort the click-through rates of the titles of the at least one search result in descending order.

[0207] The filtering unit 4013 is used to filter the titles of at least one search result whose click-through rate is less than or equal to a set value, based on the ranking result of the click-through rate of the titles of the at least one search result, and use them as the target titles of at least one target search result.

[0208] In yet another possible implementation, the second feature extraction unit 403 includes:

[0209] The second acquisition unit 4031 is used to extract features from the target title of the at least one target search result to obtain the text feature vector of the target title of the at least one target search result.

[0210] The first weighted summation unit 4032 is used to perform weighted summation on the text feature vector of the target title of the at least one target search result, using the click-through rate of the target title of the at least one target search result as the weight, to obtain the target title feature.

[0211] In yet another possible implementation, the third feature extraction unit 404 includes:

[0212] The identification unit 4041 is used to identify at least one entity from one or more of the first search term and the target title to obtain an entity set;

[0213] The third determining unit 4042 is used to determine at least one intent category to which each entity in the entity set belongs;

[0214] The third acquisition unit 4043 is used to acquire the prior probability of each intent category to which each entity belongs. The prior probability is historical statistical information on the probability of the intent category to which the entity belongs. The prior probability of each intent category to which each entity belongs is: the ratio of the first occurrence count corresponding to each entity belonging to each intent category to the total number of times each entity belonging to the at least one intent category appears in one or more of the first search term and the target title; wherein, the first occurrence count corresponding to each entity belonging to each intent category is the number of times each entity belonging to each intent category appears in one or more of the first search term and the target title.

[0215] The generation unit 4044 is used to generate entity features corresponding to the entity set based on the prior probability of each intent category to which each entity belongs.

[0216] In yet another possible implementation, the generation unit 4044 includes:

[0217] The fourth acquisition unit is used to acquire the frequency of occurrence of each entity in the entity set, wherein the frequency of occurrence of each entity is the ratio of the number of times each entity appears in one or more of the first search term and the target title to the total number of times the entity set appears in one or more of the first search term and the target title.

[0218] The second weighted summation unit is used to perform weighted summation on the prior probabilities of each intent category to which each entity in the entity set belongs, using the frequency of occurrence of each entity in the entity set as the weight, to obtain the entity features corresponding to the entity set.

[0219] In yet another possible implementation, the device further includes:

[0220] The first dimensionality reduction processing unit 408 is used to perform dimensionality reduction processing on the prior probability of each entity to each intent category in the entity set. The dimension of the prior probability of each entity to each intent category in the entity set after processing is consistent with the number of target intent categories.

[0221] In another possible implementation, the feature fusion unit 405 is used to concatenate the vectors of the search term features, the target title features, and the entity features to obtain the fused features.

[0222] In yet another possible implementation, the second determining unit 407 includes:

[0223] The second dimensionality reduction processing unit 4071 is used to perform dimensionality reduction processing on the fused features, and the dimension of the fused features obtained after processing is consistent with the number of target intent categories.

[0224] The fifth acquisition unit 4072 is used to obtain the classification probability corresponding to at least one intent category based on the fusion features obtained after processing.

[0225] The normalization processing unit 4073 is used to normalize the classification probability corresponding to the at least one intention category;

[0226] The fourth determining unit 4074 is used to determine the intention category with the highest classification probability among the classification probabilities corresponding to at least one intention category obtained after normalization processing, and use it as the target intention category.

[0227] For details on the specific implementation of each of the above units, please refer to the relevant descriptions in the foregoing method embodiments, which will not be repeated here.

[0228] An intent recognition device provided according to an embodiment of this application predicts search intent based on the fusion features of search term features, target title features, and entity features, thereby improving the accuracy of intent recognition;

[0229] Most search terms contain limited contextual information, while search result titles not only reflect user click characteristics but also contain richer textual features. Therefore, incorporating target title features further improves the accuracy of intent recognition.

[0230] Most entities suffer from ambiguity, especially in the entertainment field such as novels and movies. Prior probabilities can be used to disambiguate entities and serve as a feature, effectively introducing prior knowledge into the model and further improving the accuracy of intent recognition.

[0231] According to another embodiment of this application, Figure 4The various units or modules in the intent recognition device shown can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above units are based on logical function division. In practical applications, the function of one unit can also be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the media resource dynamic display device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0232] According to another embodiment of this application, a computer program (including program code) capable of performing the steps involved in the corresponding methods shown in the above method embodiments can be run on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), to construct a system such as... Figure 4 The intent recognition device shown herein, and the intent recognition method for implementing the embodiments of this application, are described. The computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned computing device via the computer-readable recording medium, and run therein.

[0233] Based on the description of the above method and device embodiments, this application also provides an intent recognition device. Please refer to... Figure 5 The device includes at least a processor 501, an input device 502, an output device 503, and a computer storage medium 504. The processor 501, input device 502, output device 503, and computer storage medium 504 within the device can be connected via a bus or other means.

[0234] The computer storage medium 504 can be stored in the device's memory. The computer storage medium 504 is used to store computer programs, which include program instructions. The processor 501 is used to execute the program instructions stored in the computer storage medium 504. The processor 501 (or CPU) is the computing and control core of the device, and it is adapted to implement one or more instructions, specifically adapted to load and execute one or more instructions to realize corresponding method flows or corresponding functions.

[0235] In one embodiment, the processor 501 described in this application can be used to load and execute such... Figure 1 or Figure 2 The method steps in the illustrated embodiment.

[0236] It should be noted that one or more of the above units can be implemented by software, hardware, or a combination of both. When any of the above units is implemented by software, the software exists as computer program instructions and is stored in memory. The processor can be used to execute the program instructions and implement the above method flow. The processor can be built into a system-on-chip (SoC) or ASIC, or it can be a separate semiconductor chip. In addition to the core that executes the software instructions for computation or processing, the processor may further include necessary hardware accelerators, such as field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), or logic circuits that implement dedicated logic operations.

[0237] When the above units or components are implemented in hardware, the hardware can be any one or any combination of a CPU, microprocessor, digital signal processing (DSP) chip, microcontroller unit (MCU), artificial intelligence processor, ASIC, SoC, FPGA, PLD, application-specific digital circuit, hardware accelerator, or non-integrated discrete device, which can run the necessary software or perform the above method flow independently of software.

[0238] Optionally, embodiments of this application also provide a chip system, including: at least one processor and an interface, wherein the at least one processor is coupled to a memory via the interface, and when the at least one processor executes a computer program or instructions in the memory, the chip system performs the method in any of the above method embodiments. Optionally, the chip system may be composed of chips, or may include chips and other discrete devices; embodiments of this application do not specifically limit this.

[0239] It should be understood that in the description of this application, unless otherwise stated, " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B can represent A or B; where A and B can be singular or plural. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple. Additionally, to facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" do not necessarily imply difference. In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being better or more advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0240] This application embodiment also provides a computer storage medium, which is a memory device in the device used to store programs and data. It is understood that the computer storage medium here may include the built-in storage medium in the device, or it may include extended storage media supported by the device. The computer storage medium provides storage space that stores the operating system of the device. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 301, which may be one or more computer programs (including program code). It should be noted that the computer storage medium here may be a high-speed RAM memory, or a non-volatile memory, such as at least one disk storage device; optionally, it may also be at least one computer storage medium located remotely from the aforementioned processor.

[0241] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0242] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the division of units is merely a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling, direct coupling, or communication connection shown or discussed between each other may be indirect coupling or communication connection through some interfaces, apparatuses, or units, and may be electrical, mechanical, or other forms.

[0243] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0244] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be read-only memory (ROM), random access memory (RAM), or magnetic media, such as floppy disks, hard disks, magnetic tapes, magnetic disks, or optical media, such as digital versatile discs (DVDs), or semiconductor media, such as solid-state disks (SSDs).

Claims

1. An intent recognition method, characterized in that, The method includes: In response to a user's trigger action on the title of at least one search result in the first search results page, the click-through rate of the title of the at least one search result is obtained; The click-through rates of the titles of the at least one search result are sorted in descending order. Based on the ranking results of the click-through rate of the titles of the at least one search result, the titles of at least one search result whose click-through rate is less than or equal to a set value are filtered as the target titles of at least one target search result; wherein, the first search result page is used to display the titles of at least one search result obtained based on the first search term; Perform a first feature extraction process on the first search term to obtain the search term features of the first search term; Feature extraction is performed on the target title of the at least one target search result to obtain the text feature vector of the target title of the at least one target search result; Using the click-through rate of the target title of the at least one target search result as a weight, the text feature vector of the target title of the at least one target search result is weighted and summed to obtain the target title feature; Entity recognition processing is performed on one or more of the first search term and the target title to obtain entity features; The search term features, the target title features, and the entity features are fused together to obtain fused features; Based on the fusion features, the search intent corresponding to the first search term is predicted, and the target intent category corresponding to the first search term is determined; Determine the target content corresponding to the target intent category, the target content being displayed in the search results page associated with the first search term.

2. The method according to claim 1, characterized in that, The click-through rate of the title of each search result in the at least one search result is: the ratio of the number of clicks on the title of each search result by multiple users within a set time period to the total number of clicks on the title of the at least one search result by the multiple users, where the multiple users are the users who input the first search term and the information associated with the first search term.

3. The method according to claim 1, characterized in that, The entity recognition processing of one or more of the first search term and the target title to obtain entity features includes: Identify at least one entity from one or more of the first search term and the target title to obtain an entity set; Determine at least one intent category to which each entity in the entity set belongs; Obtain the prior probability of each entity to each intent category. The prior probability is historical statistical information on the probability of an entity to its intent category. The prior probability of each entity to each intent category is the ratio of the first occurrence count of each entity belonging to each intent category to the total number of times each entity belonging to at least one intent category appears in one or more of the first search term and the target title. Wherein, the first occurrence count of each entity belonging to each intent category is the number of times each entity belonging to each intent category appears in one or more of the first search term and the target title. Entity features corresponding to the entity set are generated based on the prior probability of each intent category to which each entity belongs.

4. The method according to claim 3, characterized in that, The step of generating entity features corresponding to the entity set based on the prior probability of each intent category to which each entity belongs includes: The frequency of each entity in the entity set is obtained, wherein the frequency of each entity is the ratio of the number of times each entity appears in one or more of the first search term and the target title to the total number of times the entity set appears in one or more of the first search term and the target title. Using the frequency of occurrence of each entity in the entity set as the weight, the prior probabilities of each intent category to which each entity in the entity set belongs are weighted and summed to obtain the entity features corresponding to the entity set.

5. The method according to claim 3 or 4, characterized in that, The method further includes: The prior probability of each entity in the entity set to each intent category is reduced in dimensionality, and the resulting prior probability dimension of each entity in the entity set to each intent category is consistent with the number of the target intent categories.

6. The method according to any one of claims 1-4, characterized in that, The step of predicting the search intent corresponding to the first search term based on the fusion features and determining the target intent category corresponding to the first search term includes: The fusion features are then subjected to dimensionality reduction processing, and the dimensionality of the resulting fusion features is consistent with the number of target intent categories. Based on the fusion features obtained after the processing, the classification probability corresponding to at least one intent category is obtained; The classification probabilities corresponding to the at least one intent category are normalized. Among the classification probabilities corresponding to at least one intent category obtained after normalization, the intent category with the highest classification probability is determined as the target intent category.

7. An intent recognition device, characterized in that, The device includes: A first determining unit is configured to, in response to a user's triggering operation on the title of at least one search result in a first search result page, determine the target title of at least one target search result from the titles of the at least one search result; wherein the first search result page is configured to display the title of at least one search result obtained based on a first search term; The first determining unit includes: The first acquisition unit is configured to acquire the click-through rate of the title of the at least one search result in response to a user's trigger operation on the title of at least one search result in the first search result page. A sorting unit is used to sort the click-through rates of the titles of the at least one search result in descending order. A filtering unit is configured to filter the titles of at least one search result whose click-through rate is less than or equal to a set value, based on the ranking result of the click-through rate of the titles of the at least one search result, and use them as the target titles of at least one target search result. The first feature extraction unit is used to perform a first feature extraction process on the first search term to obtain the search term features of the first search term; The second feature extraction unit is used to perform second feature extraction processing on the target title of the at least one target search result to obtain target title features; The second feature extraction unit includes: The second acquisition unit is used to extract features from the target titles of the at least one target search result to obtain the text feature vector of the target title of the at least one target search result. The first weighted summation unit is used to perform a weighted summation of the text feature vectors of the target titles of the at least one target search result, with the click-through rate of the target title of the at least one target search result as the weight, to obtain the target title features; The third feature extraction unit is used to perform entity recognition processing on one or more of the first search term and the target title to obtain entity features; The feature fusion unit is used to fuse the search term features, the target title features, and the entity features to obtain fused features; The prediction unit is used to predict the search intent corresponding to the first search term based on the fusion features, and determine the target intent category corresponding to the first search term; The second determining unit is used to determine the target content corresponding to the target intent category, and the target content is used to be displayed in the search results page associated with the first search term.

8. An intent recognition device, characterized in that, Including input devices and output devices, and also including: A processor, adapted to implement one or more instructions; and, A computer storage medium storing one or more instructions adapted to be loaded by the processor and executed as described in any one of claims 1-6.

9. A computer storage medium, characterized in that, The computer storage medium stores one or more instructions, which are adapted to be loaded by a processor and executed as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes computer program instructions adapted to be loaded by a processor and to perform the method as described in any one of claims 1-6.

Images