WO2021008015A1

WO2021008015A1 - Intention recognition method, device and computer readable storage medium

Info

Publication number: WO2021008015A1
Application number: PCT/CN2019/116240
Authority: WO
Inventors: 石志娟; 徐小方
Original assignee: 平安科技（深圳）有限公司
Priority date: 2019-07-18
Filing date: 2019-11-07
Publication date: 2021-01-21
Also published as: CN110472027A

Abstract

An intention recognition method, device and a computer readable storage medium, applicable to the technical field of artificial intelligence. The method comprises: receiving a target search statement input by a user; performing word segmentation to the target search statement so as to obtain a word segmentation result of the target search statement; inputting the word segmentation result of the target search statement into a preset intention recognition model so as to obtain an intention recognition result corresponding to the target search statement, the intention recognition result being used for indicating whether the target search statement has question-answer properties; if the intention recognition result indicates that the target search statement has question-answer properties, outputting a search result of question-answer search result items corresponding to the target search statement. The method facilitates the improvement of intention recognition accuracy.

Description

Intention recognition method, equipment and computer readable storage medium

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on July 18, 2019, the application number is 201910653241.9, and the application name is "Intent Recognition Method, Equipment, and Computer-readable Storage Medium", the entire content of which is incorporated by reference In this application.

Technical field

This application relates to the field of artificial intelligence technology, and in particular to an intention recognition method, device, and computer-readable storage medium.

Background technique

Currently, search engines can recognize the intent of the search sentence based on the search sentence input by the user, so as to provide the user with search results based on the recognized intent. General search sentences include search sentences with question and answer intent and search sentences without question and answer intent. If it is recognized that a certain search sentence has question and answer intent, the search results of the search sentence can provide multiple question and answer data for users to view. In order to solve the user's problems as soon as possible and enhance the user experience. At present, judging whether a search sentence has question and answer intent is generally by judging whether the search sentence includes question words. If it includes question words, it is determined that the search sentence has question and answer intent, otherwise it is determined that the search sentence does not have question and answer intent. However, in fact, some search sentences with question and answer intent may not include question words, which leads to the unreliable way of identifying question and answer intent based on question words, and the accuracy of intent recognition is poor.

Summary of the invention

The embodiments of the present application provide an intent recognition method, device, and computer-readable storage medium, which can train an intent recognition model based on search event information associated with a search sentence set to perform question and answer intent recognition, which helps improve the accuracy of intent recognition.

In the first aspect, an embodiment of the present application provides an intention recognition method, including:

Receive the target search sentence entered by the user;

Performing word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence, and the word segmentation result of the target search sentence includes a plurality of word segmentation constituting the target search sentence;

Input the word segmentation result of the target search sentence into a preset intent recognition model to obtain the intent recognition result corresponding to the target search sentence, and the intent recognition model is based on multiple target search sentence sets and the multiple targets The search event information is obtained by training the search event information associated with each target search sentence set in the search sentence set, where each target search sentence set includes at least one search sentence, and the search event information includes information about each search sentence in the at least one search sentence. Search order and/or click information of the search result of each search sentence, and the intention recognition result is used to indicate whether the target search sentence has a question and answer attribute;

If the intent recognition result indicates that the target search sentence has a question and answer attribute, then output the search result including the question and answer type search result item corresponding to the target search sentence.

In a second aspect, an embodiment of the present application provides an intention recognition device, which includes a unit for executing the method of the first aspect.

In the third aspect, the embodiments of the present application provide another intention recognition device, including a processor and a memory, the processor and the memory are connected to each other, wherein the memory is used to store a computer program that supports the intention recognition device to execute the above method The computer program includes program instructions, and the processor is configured to invoke the program instructions to execute the method of the first aspect described above. Optionally, the intent recognition device may further include a user interface and/or a communication interface.

In a fourth aspect, an embodiment of the present application provides a computer non-volatile readable storage medium, the computer non-volatile readable storage medium stores a computer program, the computer program includes program instructions, and the program instructions When executed by a processor, the processor is caused to execute the method of the first aspect.

In the embodiment of the present application, an intent recognition model can be trained based on search event information associated with a search sentence set to perform question and answer intent recognition, so that the accuracy of intent recognition is improved, and the reliability of question and answer intent recognition is higher.

Description of the drawings

In order to more clearly describe the technical solutions of the embodiments of the present application, the following will describe the drawings that need to be used in the description of the embodiments.

FIG. 1 is a schematic flowchart of an intention recognition method provided by an embodiment of the present application;

FIG. 2 is a schematic flowchart of another intention identification method provided by an embodiment of the present application;

FIG. 3 is a schematic structural diagram of an intention recognition device provided by an embodiment of the present application;

Fig. 4 is a schematic structural diagram of another intention recognition device provided by an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

The technical solution of the present application can be applied to an intent recognition device, which may include a server, a terminal, a robot, or other recognition devices, for training an intent recognition model, recognizing the intent of a user's search sentence, and so on. The terminal involved in this application may be a mobile phone, computer, tablet, personal computer, smart watch, etc., which is not limited in this application.

Specifically, in this application, the intent recognition result of the target search sentence can be obtained by inputting the target search sentence for intent recognition to the intent recognition model trained based on multiple search sentence sets and their associated search event information. Determine whether the target search sentence has question and answer attributes, and then output the question and answer search results when the target search sentence has question and answer attributes, that is, the intent recognition model can be trained according to the search event information associated with the search sentence set for question and answer intent recognition. This improves the accuracy of intention recognition, and makes question-and-answer intention recognition more reliable. Detailed descriptions are given below.

Please refer to FIG. 1, which is a schematic flowchart of an intention recognition method provided by an embodiment of the present application. Specifically, the technical solution of this embodiment can be applied to the aforementioned intention recognition device. As shown in Figure 1, the intention recognition method may include the following steps:

101. Receive a target search sentence input by a user.

Wherein, the target search sentence is a search sentence for intent recognition. It can be understood that, in other embodiments, the target search sentence may also be obtained in other ways, such as from a search queue; the target search sentence may be input by text or voice, etc. This application does not limit the method of obtaining or inputting the target search sentence.

102. Perform word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence.

Wherein, the word segmentation result of the target search sentence may include multiple word segments (also referred to as words, words, entries, etc.) that make up the target search sentence. Optionally, the multiple participles may refer to all the participles of the target search sentence; or, the multiple participles may refer to the partial participles of all the participles, such as removing meaningless participles from all the participles (such as removing For example, a filter list can be preset, and the filter list can include various stop words or other meaningless words, such as "ah", "oh", " After the target search sentence is segmented, the stop words and other meaningless words in the query sentence can be determined by matching and comparison with the words in the filter list, and these words can be removed to Reduce the detection overhead of determining whether the search sentence has question and answer attributes; etc., which are not listed here.

Optionally, the word segmentation method corresponding to the word segmentation processing may be stuttering word segmentation or Stanford word segmentation or other word segmentation methods, which are not limited in this application.

103. Input the word segmentation result of the target search sentence into a preset intention recognition model to obtain the intention recognition result corresponding to the target search sentence. Among them, the intention recognition model can be used to identify whether the search sentence has question and answer attributes. The intent recognition model may be trained based on multiple target search sentence sets and search event information associated with each target search sentence set in the multiple target search sentence sets. Each target search sentence set may include at least one search sentence. The search event information may include the search order of each search sentence in the at least one search sentence and/or search result click information of each search sentence, and the intention recognition result may be used to indicate whether the target search sentence has a question and answer attribute.

Before inputting the word segmentation result of the target search sentence into a preset intent recognition model to obtain the intent recognition result corresponding to the target search sentence, the intent recognition model can be obtained by pre-training. Specifically, multiple target search sentence sets and their associated search event information can be obtained, for example, from a preset search sentence database, and the search order of each search sentence in each target search sentence set and/or each The click information of the search result of the search sentence determines the intent of each target search sentence set. For example, it is determined whether the search sentence included in each target search sentence set has a question and answer attribute, and then according to the search sentence included in each target search sentence set and whether it has The determined result of the question and answer attribute is trained to obtain the intention recognition model.

Wherein, the search order can be used to indicate the search order of each search sentence, and the search result click information can be used to indicate the information of the search result item clicked by the user. Optionally, the search order can be text information or identification information (such as 1, 2, 3...), or the search event information can also include the search time of each search sentence, and the search of each search sentence The order can be indicated by the search time, etc., which is not limited in this application. The search result click information may include the total number of clicks on search result items, the number of clicks on Q&A search result items, and the number of clicks on non-Q&A search results (for example, by setting a label to indicate whether it is a Q&A or a non-Q&A) and / Or the browsing time of each search result item clicked, etc.

Optionally, the multiple target search sentence sets may be search sentence sets whose occurrence times in the search sentence database are greater than a first number threshold; or may be search sentences whose proportions in the search sentence database are greater than a preset proportion value Alternatively, the search sentence database may also record the search time of each search sentence, and the selected multiple target search sentence sets may be search sentence sets within the historical time window such as the previous month; or, multiple selected targets The search sentence set may also be determined in combination with the application field of the intent recognition model to be trained, or the selected multiple target search sentence sets may also be selected by combining any two or more selection methods mentioned above, etc. , I will not list them here. This helps to improve the reliability of the selected model training data.

Wherein, determining whether the search sentences included in each target search sentence set has a question and answer attribute can also be referred to as determining whether each target search sentence set has a question and answer attribute. Optionally, when determining whether a search sentence included in a target search sentence set has a question and answer attribute, it can be determined according to the click information of the search result of part of the search sentence in the key search event information of the target search sentence set, for example, according to the target search The search results of the first M search sentences (that is, the M search sentences with the most recent/latest search time) in the sentence set search order are determined by clicking on the information, and M is an integer greater than or equal to 1; alternatively, all search sentences can be set according to the target The search result of the search sentence is determined by click information; or, it can be determined according to the weighting coefficient of each search sentence in the target search sentence set and the search result click information of each search sentence, etc., which are not listed here.

104. If the intent recognition result indicates that the target search sentence has a question and answer attribute, output a search result including a question and answer type search result item corresponding to the target search sentence.

After determining that the target search sentence has a question and answer attribute, the question and answer data, that is, the search result item of the question and answer category corresponding to the target search result, can be obtained, and displayed to provide the user. For example, the Q&A search result items can be displayed in front of the output interface according to the generation time or the relevance to the target search sentence, and the non-Q&A search result items can be displayed after all the Q&A search result items; The Q&A search result item selects some search result items, such as the top N items with the latest generation time or the top M items with the highest correlation with the target search sentence, which will be displayed on the output interface, and will be displayed after the N items or M items. The non-Q&A search result items corresponding to the target search sentence (for example, the top E item with the latest generation time or the top F item with the highest correlation with the target search sentence is still displayed), where the N, M, E, and F are all It is an integer greater than 0; for another example, the output interface may only display the search result items of the question and answer category corresponding to the target search result, etc., which are not listed here.

In this embodiment, when acquiring the target search sentence input by the user, the intention recognition device can perform word segmentation processing on the target search sentence to obtain a word segmentation result, and input the word segmentation result into a set based on multiple search sentence sets and their associations. The intent recognition model trained on the search event information to obtain whether the target search sentence has the question and answer attribute, and then when the target search sentence has the question and answer attribute, the search result including the question and answer search result item is output, so that the search The search event information associated with the sentence set is trained to obtain an intent recognition model for question and answer intent recognition, which improves the accuracy of intent recognition, and the reliability of question and answer intent recognition is high.

Please refer to FIG. 2, which is a schematic flowchart of another intention recognition method provided by an embodiment of the present application. Specifically, as shown in FIG. 2, the intention recognition method may include the following steps:

201. Select a plurality of target search sentence sets from a search sentence database. The search sentence database records multiple search sentence sets and search event information associated with each search sentence set.

Wherein, each search sentence set includes one or more search sentences, that is, includes at least one search sentence, and the search event information includes the search order of each search sentence in the at least one search sentence and/or the search of each search sentence The result is clicked information, so I won’t repeat it here.

Optionally, if a search sentence set includes multiple search sentences, the search time interval between any two of the multiple search sentences does not exceed a preset time threshold, and any two of the multiple search sentences The overlap rate of keywords between search sentences (such as other word segmentation after removing meaningless words) is higher than the preset overlap rate threshold. That is to say, the search sentences included in the search sentence set may refer to similar search sentences within a preset time range (such as within 2 minutes of the first search), that is, keywords (such as removing modal particles and stop words from the search sentence). The word segmentation is used as a keyword) for search sentences whose overlap rate is higher than the preset overlap rate threshold. For example, the preset time threshold is 2min, the overlap rate threshold is 70%, the search time interval of two search sentences is 30s, that is, the preset time threshold is not exceeded, and the keywords of the two search sentences are 5 and 6 respectively. There are 4 (same) keywords overlapped by the two search sentences. The weight of each keyword is the same. There is no weighting coefficient, that is, the overlap rate is 4/5=80% (that is, two keywords can be selected In other embodiments, the number of keywords in the smaller one may be larger, or the average value of the two may also be used, etc., which are not listed here), which is greater than the overlap rate threshold, Then these two search sentences can be put into the same search sentence set. Because the user may change the sentence pattern or structure of the search sentence to search when the first search result is not ideal. Further optionally, weighting coefficients can be set for preset keywords (such as domain-specific words or words with a high frequency of occurrence) in advance, and the weighting coefficients corresponding to each preset keyword may be the same or different; When determining similar search sentences to determine the set of search sentences, you can match whether the specific keyword exists in the search sentence. If the specific keyword exists, then the specific keyword can be weighted according to the weighting coefficient of the specific keyword. In other words, the keyword overlap rate is weighted, that is, after the keyword overlap rate is increased, the similar search sentence is judged, and the search sentence set is determined according to the search time and the overlap rate of the weighted search sentence. This helps to improve the reliability of the search sentence set determination.

Further optionally, optionally, the multiple target search sentence sets may be search sentence sets in the search sentence database whose occurrence times are greater than a first number threshold; or, it may be that the proportion of the search sentence database in the search sentence database is greater than a preset Proportional search sentence sets; or, the search sentence database may also record the search time of each search sentence, and the selected multiple target search sentence sets may be the search sentence sets in the historical time window, such as the previous month; or, The selected multiple target search sentence sets may also be determined in combination with the application field of the intent recognition model to be trained, or the selected multiple target search sentence sets may also be combined through any two or more selection methods described above Selected, etc., not listed here. This helps to improve the reliability of the selected model training data.

For example, in a possible implementation manner, when selecting the multiple target search sentence sets, the intent recognition device may determine the search sentence sets whose occurrence times are greater than the preset second number threshold from the search sentence database, and determine The set of search sentences with the number of occurrences greater than the second number threshold is used as the multiple target search sentence sets; or, the second set between the number of occurrences and the total number of search sentences in the search sentence database can be determined from the search sentence database. A search sentence set whose ratio is greater than a preset second proportion threshold, and the determined search sentence set whose second ratio is greater than the second proportion threshold is used as the multiple target search sentence sets; or, it can also be obtained from a search sentence database In determining the number of occurrences greater than the preset second number threshold, and the second ratio is greater than the preset second ratio threshold, the determined number of occurrences is greater than the second number threshold and the second ratio is greater than the preset The search sentence set of the second ratio threshold is set as the multiple target search sentence sets, etc., which are not listed here. Wherein, the number of occurrences of the search sentence set may be the sum of the number of occurrences of the search sentences included in the search sentence set, or the number of occurrences of the search sentence set may be the average number of occurrences of the search sentences included in the search sentence set, or , The number of occurrences of the search sentence set may be the highest number of occurrences of the search sentences included in the search sentence set, etc., which are not listed here. The number of occurrences of search sentences may refer to the number of search sentences in the search database or the number of search sentences in the search database whose similarity to the search sentence is higher than a threshold, and so on, which is not limited in this application.

For another example, in a possible implementation manner, when selecting the multiple target search sentence sets, the intention recognition device may determine the application field information of the intent recognition model to be trained, and obtain information from the search sentence database according to the application field information. The target sub-database is determined from the included multiple sub-databases, and then the multiple target search sentence sets are selected from the target sub-database. Among them, the sub-database has a one-to-one correspondence with the application field, and each sub-database includes multiple search sentence sets under the corresponding application field (the number of which is greater than the number of the selected target search sentence sets) and the search events associated with each search sentence set Information, the application field corresponding to the target sub-database is the same as the application field indicated by the application field information. That is to say, the search sentence database may include sub-databases under each application field, and each sub-database includes a search sentence set under an application field and the search order of each search sentence associated with each search sentence set and search result click information And so on, so when selecting the target search sentence set, you can determine the sub-database (such as the sub-database carrying the field label) by determining the application field information (such as the field label) of the intent recognition model to be trained, and select the target from it Search sentence set. Thereby, the reliability of the selected model training data can be further improved, and the training effect can be improved.

202. Perform word segmentation processing on the search sentences included in each target search sentence set of the multiple target search sentence sets respectively to obtain a word segmentation result of each target search sentence set.

Wherein, the word segmentation result of each target search sentence set includes multiple word breaks that make up the search sentence of the target search sentence set, and the multiple word breaks may refer to all the word breaks of the search sentence of the target search sentence set, or may refer to The partial participles in all the participles, for example, the participles after removing the meaningless participles (such as removing stop words or other meaningless participles) from all the participles. For example, a filter list can be preset, and the filter list can include Various stop words or other meaningless words, such as "ah", "oh", "的", etc., so that after the search sentence of the target search sentence set is segmented, it can pass the word in the filter list The method of matching and comparison determines the stop words and other meaningless words in the query sentence, and removes these words to reduce the detection overhead of determining whether the search sentence has the question and answer attribute; or, the multiple word segmentation can refer to the The word segmentation of the search sentence with the largest search order (that is, the most recent search) in the target search sentence set (may be all or part of the word segmentation of the search sentence with the largest search order, which will not be repeated here), etc., and will not be repeated here.

203. According to the search event information associated with each target search sentence set, determine whether the search sentence included in each target search sentence set has a question and answer attribute.

Optionally, the search result click information may include the total number of clicks of search result items and the number of clicks of Q&A search result items; when determining whether the search sentence has the Q&A attribute according to the search result click information of the search sentence, you can select The total number of clicks on the search result items included in the search result click information of the search sentence is compared with a preset first number threshold, and the number of clicks on the Q&A search result items included in the search result click information of the search sentence is calculated with The first ratio between the total number of clicks on the search result item, and compare the first ratio with the preset first ratio value; if the total number of clicks on the search result item is greater than the preset first number threshold, and the If the first ratio is greater than the preset first ratio threshold, it can be determined that the search sentence has a question and answer attribute; otherwise, it can be indicated that it does not have a question and answer attribute (or can be further determined in combination with other methods). Or, optionally, the search result click information may include the total number of clicks on search result items, the number of clicks on Q&A search result items, and the browsing time of each clicked search result item; in the search result click information based on the search sentence When determining whether a search sentence has a question and answer attribute, you can filter out search result items whose browsing duration is less than the preset duration threshold, and determine the total number of clicks of search result items remaining after filtering the search result item (that is, the search result click information includes the search The total number of clicks on the result item minus the number of search result items whose browsing duration is less than the preset duration threshold), to determine the number of clicks on the question and answer search result items remaining after filtering the search result item (ie the question and answer included in the search result click information The number of clicks on search result items of the category minus the number of search result items whose browsing duration is less than the preset duration threshold), and the number of clicks on the remaining Q&A search result items and the total number of clicks on the remaining search result items are calculated If the total number of clicks on the remaining search result items is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, it can be determined that the search sentence has a question and answer attribute . Or, optionally, when determining whether the search sentence has a question and answer attribute according to the click information of the search result of the search sentence, the number of clicks on the search result items of the question and answer type included in the search result click information of the search sentence and the preset If the number of clicks of the search result item of the question and answer category is greater than the other number threshold, it can be determined that the search sentence has question and answer attributes, etc., which are not listed here.

For example, in a possible implementation manner, when determining whether the search sentences included in the target search sentence set have question-and-answer attributes, the search order of each search sentence included in the search event information associated with the target search sentence set may be used. , Determine the search sentence corresponding to the largest search order in the at least one search sentence included in the target search sentence set; determine whether the search sentence included in the target search sentence set has the search result click information of the search sentence corresponding to the maximum search sequence Q&A attributes. According to the click information of the search result of the search sentence corresponding to the maximum search order, the method of determining whether the search sentence included in the target search sentence set has the question and answer attribute can be determined by referring to the above click information according to the search result of the search sentence to determine whether the search sentence has the question and answer attribute The method is not repeated here. If the search sentence corresponding to the maximum search order has the question and answer attribute, it can be determined that the search sentence included in the target search sentence set has the question and answer attribute. That is to say, when determining whether a target search sentence has a question and answer attribute, the most recent search event from the related search events, that is, the search event of the search sentence corresponding to the maximum number of searches, can be selected according to the maximum number of searches. The search result information of to determine whether the target search sentence has question and answer attributes, because the search results obtained from the previous searches may not be what the user wants, you can follow the subsequent clicks to improve the judgment efficiency and ensure the judgment accuracy .

For another example, in a possible implementation manner, when determining whether the search sentence included in the target search sentence set has a question and answer attribute, it can be determined according to the click information of the search results of all the search sentences in the target search sentence set. For details, please refer to The above method of determining whether the search sentence has a question and answer attribute according to the click information of the search result of the search sentence, for example, counting the sum of the number of clicks of the search result items of the question and answer type in the search result click information of all the search sentences, and judging the sum of the number of clicks Whether it exceeds the preset number threshold, if it exceeds, it can indicate that the search sentences included in the target search sentence set have question-and-answer attributes, etc., which will not be repeated here.

For another example, in a possible implementation manner, the weighting coefficient of each search sentence may be preset, for example, the weighting coefficient of the search sentence including the question word is higher than the weighting coefficient of the search sentence not including the question word, and/or, The weighting coefficient of the search sentence with a higher search order is higher than the weighting coefficient of the search sentence with a lower search order (that is, the higher the search order, the higher the weighting coefficient is), and/or the search result included in the click information for the search result The click item has a display result of a specific question and answer website or a search sentence that has a display result of a specific question and answer website in the search result, and its weighting coefficient is higher than the weight coefficient of a search sentence that does not have a display result of the specific question and answer website, and so on. When the search sentences included in the target search sentence set have question-and-answer attributes, the weight coefficient corresponding to each search sentence in the at least one search sentence included in the target search sentence set can be determined; according to the sum of the weight coefficients corresponding to each search sentence The search result click information in the search event information associated with the target search sentence set determines whether the search sentence included in the target search sentence set has a question and answer attribute. According to the weighting coefficient and the search result click information can refer to the weighting of the parameters of the search result click information, such as the number of search result click items of the question and answer category, the browsing time, etc., by the weighting coefficient. After weighting, the question and answer attribute determination method is specific You can refer to the above method of determining whether the search sentence has question and answer attributes based on the click information of the search result of the search sentence. For example, the number of clicks on the Q&A search result item corresponding to each search sentence can be weighted by the weighting coefficient of each search sentence (for example, the number of clicks on the Q&A search result item corresponding to each search sentence is 2, and the weighting factor is 1.5. , The number of clicks on the weighted question and answer search result items is 2*1.5=3); if the total number of clicks on the search result items of each search sentence in the target search sentence set is greater than the preset first number threshold, and each search The first ratio between the sum of the number of clicks on the search result items of the question and answer category (weighted) and the total number of clicks on the search result items of each search sentence is greater than the preset first ratio threshold, then the target can be determined The search sentences included in the search sentence set have question and answer attributes. That is to say, when determining whether a target search sentence set has the question and answer attribute, it can be determined whether the target search sentence has the question and answer attribute according to the search result information clicked by the user for each search number and the weight of each search result. This helps to improve the reliability of the question and answer attributes of the determined search sentence set.

Optionally, in this application, the execution order of

steps

202 and 203 is not limited. For example, step 203 can be performed first, and then step 202 can be performed, or

steps

202 and 203 can be performed simultaneously, which is not limited in this application. .

204. The word segmentation results of the target search sentence set with question and answer attributes in the multiple target search sentence sets are used as positive samples, and the word segmentation results of the target search sentence sets without question answering attributes in the multiple target search sentence sets are used as negative samples.

Optionally, the word segmentation results of the target search sentence set with the question and answer attribute in the multiple target search sentence sets may include the word segmentation of the search sentences with the question and answer attribute in the multiple target search sentence sets, which may be one or more target searches All participles of the sentence set can also be part of the participles. That is, the positive sample may include word segmentation of search sentences with question and answer attributes, and the negative sample may include word segmentation of search sentences without question and answer attributes. For example, when it is determined that the search sentence of a certain target search sentence set has a question and answer attribute, all the word segmentation (meaningless word segmentation can be removed) of the search sentence of the target search sentence set can be used as a positive sample.

In some embodiments, the word segmentation of the search sentence with the question and answer attribute may be used as a positive training sample, and the word segmentation of the search sentence without the question and answer attribute may be used as the negative training sample, so as to obtain the intention recognition based on the positive training sample and the negative training sample. Model, so that the subsequent intent recognition model can quickly identify whether the input search sentence has question and answer attributes. Furthermore, it is possible to return information for the user according to whether the recognized input search sentence has a Q&A attribute recognition result. For example, the search sentence with question and answer attributes may indicate that there is a question and answer requirement, and the search sentence without question and answer attributes may indicate that there is no question and answer requirement, so that different pages (interfaces) can be returned to the user according to whether the search sentence has question and answer requirements. Different demand content.

Optionally, the description of the steps 201-204 and the related description of the embodiment shown in FIG. 1 may refer to each other, which is not repeated here.

205. Calculate the absolute value of the difference between the number of search sentences corresponding to the positive sample and the number of search sentences corresponding to the negative sample.

206. Determine whether the absolute value exceeds a preset number threshold.

207. If the absolute value exceeds the number threshold, process the positive sample and/or the negative sample according to a preset sample balance rule to obtain processed positive samples and negative samples.

Optionally, after determining the positive samples and negative samples corresponding to the multiple target search sentence sets, it can be further determined whether the number of positive and negative samples is balanced, for example, it is determined that the number of search sentences corresponding to the positive sample corresponds to the negative sample Whether the absolute value of the difference between the number of search sentences exceeds a preset number threshold, such as a preset third number threshold, if it exceeds, it may indicate that the number of positive and negative samples is unbalanced. When training the model, the positive and negative samples are often unbalanced, resulting in poor recognition accuracy of the trained model. It is easy to overfit a large sample, that is to say, the prediction is easy to be biased to the classification with a large number of samples. This greatly reduces the normalization ability of the model, resulting in unreliable recognition results. Therefore, before training, the number of positive and negative samples can be counted separately. When the difference between the two is too large, such as the difference exceeds the preset third number threshold, the number of positive and negative samples can be balanced according to the preset sample balance rule. training. The preset sample balance rule can be multiple, and can be selected according to the number of positive and negative samples, or according to the training scenario. For example, in the case of fewer positive samples, you can increase the positive samples to balance the positive and negative samples; another example, in the case of fewer negative samples, you can increase the negative samples to balance the positive and negative samples; another example, for needs For scenes where a large number of samples are trained (for example, the scene label is multi-sample), the positive and negative samples can be balanced by the way of synthetic samples; another example, for scenes with high reliability requirements (for example, the scene label is high reliability), changes can be adopted The method of sample weight balances positive and negative samples, and the specific balance rules for each sample and the selected scenes can be preset. Optionally, the way to balance positive and negative samples can be as follows:

1) Upsampling: To increase the samples with a small number of samples, the method is to directly copy the original samples. For example, it can be used when the sample is small.

2) Downsampling: To reduce samples with a large number of samples, the way is to discard these redundant samples. For example, it can be used when there are many samples. For example, the target search sentences can be sorted according to the total number of clicks, and samples corresponding to the target search sentences with a small total number of clicks are discarded.

3) Synthetic samples: increase the type of samples with a small number of samples. Synthesis refers to the combination of various features of existing samples to generate new samples. Specifically, the method of generating a new sample can be to randomly select some features from each feature or select some specific features through some methods (such as features with a number of occurrences higher than a threshold, or sample similarity higher than a threshold, such as Europe The features between samples whose distance is less than the threshold, etc.) are then spliced into a new sample, thereby increasing the number of samples in the category with a smaller number of samples. Unlike upsampling, which is simply copying samples, here is the splicing to obtain new samples, which can further improve the reliability of model training.

4) Change the sample weight: increase the weight of the key word segmentation. If for a positive sample, the word segmentation with obvious question and answer attributes can be multiplied by a weight to improve the reliability of judgment.

After the positive and negative samples are obtained, the intent recognition model can be trained, so that the subsequent intent recognition model can quickly identify whether the input search sentence has question and answer intent.

208. Use the processed positive samples and negative samples to train to obtain the intention recognition model.

Among them, the intention recognition model can be used to identify whether the input search sentence has question and answer attributes. The model may be a model based on a binary tree, a model based on a multi-tree, or a neural network model, etc., which is not limited in this application.

Optionally, after the search sentence (set) with question and answer intent is determined, the question and answer category of the search sentence can be further determined, for example, whether it belongs to an explicit question answer search sentence that includes question words or does not include question words. Implicit question and answer search statement. The intent recognition model can be trained based on the positive and negative samples and the question and answer category (such as category label) to which each positive sample belongs. Therefore, when the intention recognition model is used to identify the question and answer attribute of the search sentence, not only the question and answer type search sentence, that is, the search sentence with the question and answer intention, can be identified, but also the question and answer category to which the question and answer intention belongs can be determined. Further optionally, the corresponding relationship between each question and answer category and display content/page (keyword or content title format) can be preset, and then the content can be displayed for users according to the question and answer category, which improves the flexibility of page display .

209. Receive the target search sentence input by the user.

210. Perform word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence.

211. Input the word segmentation result of the target search sentence into a preset intention recognition model to obtain the intention recognition result corresponding to the target search sentence.

212. If the intent recognition result indicates that the target search sentence has a question and answer attribute, output a search result including a question and answer type search result item corresponding to the target search sentence.

Optionally, the description of the steps 209-212 and the related description of the steps 101-104 in the embodiment shown in FIG. 1 may refer to each other, and details are not repeated here.

In this embodiment, the intent recognition device can perform word segmentation processing on the search sentences included in the selected multiple target search sentence sets to obtain the word segmentation result of each target search sentence set, and according to the associated search sentence set of each target search sentence set. The search event information determines whether the search sentences included in each target search sentence set have question and answer attributes, and then the word segmentation results of search sentences with question and answer attributes in the multiple target search sentence sets can be used as positive samples and search sentences that do not have question and answer attributes The word segmentation result of is used as a negative sample, and after the positive and negative samples are balanced according to the preset sample balance rule, the intent recognition model is trained based on the balanced positive and negative samples to perform question and answer intent recognition, so that the obtained search sentence Input to the intent recognition model to identify whether the target search sentence has question and answer attributes, and then when the target search sentence has question and answer attributes, the output includes the search result items of the question and answer category, which improves the accuracy of intent recognition, so that The reliability and recall rate of Q&A intention recognition are high.

The foregoing method embodiments are all examples of the intention identification method of the present application, and the description of each embodiment has its own focus. For parts that are not described in detail in an embodiment, reference may be made to related descriptions of other embodiments.

Please refer to FIG. 3, which is a schematic structural diagram of an intention recognition device provided by an embodiment of the present application. The intention recognition device of the embodiment of the present application includes a unit for executing the above-mentioned intention recognition method. Specifically, the intention recognition device 300 of this embodiment may include: an acquiring unit 301 and a processing unit 302. among them,

The obtaining unit 301 is configured to receive the target search sentence input by the user;

The processing unit 302 is configured to perform word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence, and the word segmentation result of the target search sentence includes multiple word segmentation that constitute the target search sentence;

The processing unit 302 is further configured to input the word segmentation result of the target search sentence into a preset intent recognition model to obtain an intent recognition result corresponding to the target search sentence, and the intent recognition model is based on multiple target search sentences Set and the search event information associated with each target search sentence set in the multiple target search sentence sets, each target search sentence set includes at least one search sentence, and the search event information includes the at least one search The search order of each search sentence in the sentence and/or the click information of the search result of each search sentence, and the intention recognition result is used to indicate whether the target search sentence has a question and answer attribute;

The processing unit 302 is further configured to output a search result including a question and answer type search result item corresponding to the target search sentence if the intention recognition result indicates that the target search sentence has a question and answer attribute.

Optionally, the acquiring unit 301 is further configured to select multiple target search sentence sets from a search sentence database; wherein, the search sentence database records multiple search sentence sets and search event information associated with each search sentence set Wherein each search sentence set includes at least one search sentence, and the search event information includes a search order of each search sentence in the at least one search sentence and/or search result click information of each search sentence;

The processing unit 302 is further configured to separately perform word segmentation processing on the search sentences included in each target search sentence set in the multiple target search sentence sets to obtain the word segmentation result of each target search sentence set. The word segmentation result of the search sentence set includes multiple word segmentation that constitute the search sentence of the target search sentence set;

The processing unit 302 may be further configured to determine whether the search sentences included in each target search sentence set have question and answer attributes according to the search event information associated with each target search sentence set; The word segmentation result of the target search sentence set of the question and answer attribute is used as a positive sample, and the word segmentation result of the target search sentence set that does not have the question and answer attribute in the plurality of target search sentence sets is used as a negative sample, and the multiple target search sentence sets are used Corresponding positive samples and negative samples are trained to obtain an intention recognition model; wherein, the intention recognition model is used to identify whether the input search sentence has question and answer attributes.

Optionally, the processing unit 302 may be specifically configured to determine the at least one search sentence included in the target search sentence set according to the search order of each search sentence included in the search event information associated with the target search sentence set The search sentence corresponding to the largest search order in the middle; according to the click information of the search result of the search sentence corresponding to the largest search order, it is determined whether the search sentence included in the target search sentence set has a question and answer attribute.

Optionally, the search result click information includes the total number of clicks on search result items and the number of clicks on Q&A search result items;

The processing unit 302, when determining whether the search sentence included in the target search sentence set has a question and answer attribute according to the click information of the search result of the search sentence corresponding to the maximum search order, may be specifically used to: calculate the maximum search The first ratio between the number of clicks on the search result items of the question and answer category and the total number of clicks on the search result items included in the search result click information of the search sentence corresponding to the order; if the total number of clicks on the search result item is greater than the preset A first number threshold, and the first ratio is greater than a preset first ratio threshold, it is determined that the search sentences included in the target search sentence set have question and answer attributes.

Optionally, the processing unit 302 may be specifically configured to determine a weighting coefficient corresponding to each search sentence in the at least one search sentence included in the target search sentence set, and a search sentence with a higher search order among the at least one search sentence The weighting coefficient of is higher than the weighting coefficient of search sentences with a small search order; the target search sentence set is determined according to the weighting coefficient corresponding to each search sentence and the search result click information in the search event information associated with the target search sentence set Whether the included search sentence has question and answer attributes.

Optionally, the acquiring unit 301 may be specifically configured to: determine from the search sentence database a set of search sentences with a number of occurrences greater than a preset second number threshold, and determine that the number of occurrences is greater than the second number threshold As the multiple target search sentence sets; or, determining from the search sentence database that the second ratio between the number of occurrences and the total number of search sentences in the search sentence database is greater than a preset second ratio threshold Search sentence sets, and use the determined search sentence sets whose second ratio is greater than the second ratio threshold as the multiple target search sentence sets.

Wherein, the number of occurrences of the search sentence set is the sum of the number of occurrences of the search sentences included in the search sentence set, or the number of occurrences of the search sentence set is the average number of occurrences of the search sentences included in the search sentence set.

Optionally, the acquiring unit 301 may be specifically configured to: determine the application field information of the intent recognition model to be trained; determine the target sub-database from multiple sub-databases included in the search sentence database according to the application field information; The target sub-database selects the multiple target search sentence sets.

Wherein, the sub-databases have a one-to-one correspondence with application fields, each sub-database includes multiple search sentence sets under the corresponding application field and search event information associated with each search sentence set, and the application field corresponding to the target sub-database is The application fields indicated by the application field information are the same.

Optionally, the processing unit 302 may also be configured to calculate the number of search sentences corresponding to the positive samples and the total number of search sentences corresponding to the positive samples before the intent recognition model is obtained by training using the positive samples and negative samples corresponding to the multiple target search sentence sets. The absolute value of the difference between the number of search sentences corresponding to the negative sample; determine whether the absolute value exceeds a preset third number threshold; if the absolute value exceeds the third number threshold, according to the preset The sample balance rule processes the positive sample and/or the negative sample to obtain processed positive sample and negative sample;

The processing unit 302 may be specifically configured to use the processed positive samples and negative samples to train to obtain the intention recognition model.

Specifically, the intention recognition device can implement part or all of the steps in the intention recognition method in the embodiment shown in FIG. 1 to FIG. 2 through the foregoing unit. It should be understood that the embodiment of the present application is an apparatus embodiment corresponding to the method embodiment, and the description of the method embodiment is also applicable to the embodiment of the present application, and will not be repeated here.

Please refer to FIG. 4, which is a schematic structural diagram of another intention recognition device provided by an embodiment of the present application. The intention recognition device is used to perform the above-mentioned method. As shown in FIG. 4, the intention recognition device 400 in this embodiment may include: one or more processors 401 and a memory 402. Optionally, the intention recognition device may further include one or more user interfaces 403 and/or one or more communication interfaces 404. The above-mentioned processor 401, user interface 403, communication interface 404, and memory 402 may be connected through a bus 405, or may be connected in other ways, as illustrated in FIG. 4 by way of a bus. The memory 402 is used to store a computer program, and the computer program includes program instructions, and the processor 401 is used to execute the program instructions stored in the memory 402.

The processor 401 may be used to call the program instructions to perform the following steps: call the user interface 403 to receive the target search sentence input by the user; perform word segmentation processing on the target search sentence to obtain the word segmentation result of the target search sentence, so The word segmentation result of the target search sentence includes multiple word segmentation that make up the target search sentence; the word segmentation result of the target search sentence is input into a preset intention recognition model to obtain the intention recognition result corresponding to the target search sentence, The intent recognition model is trained based on multiple target search sentence sets and search event information associated with each target search sentence set in the multiple target search sentence sets, and each target search sentence set includes at least one search sentence The search event information includes the search order of each search sentence in the at least one search sentence and/or search result click information of each search sentence, and the intention recognition result is used to indicate whether the target search sentence is It has a question and answer attribute; if the intent recognition result indicates that the target search sentence has a question and answer attribute, the user interface 403 is called to output a search result including a question and answer type search result item corresponding to the target search sentence.

Optionally, the processor 401 may perform the following steps before executing the inputting the word segmentation result of the target search sentence into a preset intent recognition model to obtain the intent recognition result corresponding to the target search sentence: Multiple target search sentence sets are selected from the search sentence database; wherein, the search sentence database records multiple search sentence sets and search event information associated with each search sentence set, and each search sentence set includes at least one search Sentence, the search event information includes the search order of each search sentence in the at least one search sentence and/or the click information of the search result of each search sentence; each target in the plurality of target search sentences The search sentences included in the search sentence set are subjected to word segmentation processing to obtain the word segmentation result of each target search sentence set, and the word segmentation result of each target search sentence set includes multiple word segmentation that constitute the search sentence of the target search sentence set According to the search event information associated with each target search sentence set, it is determined whether the search sentences included in each target search sentence set have question and answer attributes; The word segmentation result is taken as a positive sample, and the word segmentation result of the search sentence that does not have the question and answer attribute in the multiple target search sentence sets is taken as a negative sample, and the intention is obtained by training with the positive samples and negative samples corresponding to the multiple target search sentence sets Recognition model; wherein the intention recognition model is used to recognize whether the input search sentence has question and answer attributes.

Optionally, when the processor 401 executes the search event information associated with the target search sentence set to determine whether the search sentence included in the target search sentence set has a question and answer attribute, the processor 401 may specifically execute the following steps: The search order of each search sentence included in the search event information associated with the target search sentence set determines the search sentence corresponding to the largest search order in the at least one search sentence included in the target search sentence set; and according to the maximum search order The search result click information of the corresponding search sentence determines whether the search sentence included in the target search sentence set has a question and answer attribute.

Optionally, when the processor 401 executes the search event information associated with the target search sentence set to determine whether the search sentence included in the target search sentence set has a question and answer attribute, the processor 401 may specifically perform the following steps: determine the The target search sentence set includes a weighting coefficient corresponding to each search sentence in the at least one search sentence; according to the weighting coefficient corresponding to each search sentence and the search result click information in the search event information associated with the target search sentence set, It is determined whether the search sentences included in the target search sentence set have question and answer attributes.

Further optionally, the weighting coefficient of the search sentence with a higher search order in the at least one search sentence is higher than the weighting coefficient of the search sentence with a lower search order, and/or the search sentence including the question word in the at least one search sentence The weighting coefficient of is higher than the weighting coefficient of search sentences that do not include question words, etc., which will not be repeated here.

Optionally, the search result click information includes the total number of clicks of search result items and the number of clicks of Q&A search result items; the processor 401 is executing the search result clicks of the search sentence corresponding to the maximum search order Information, when determining whether the search sentences included in the target search sentence set have question and answer attributes, the following steps may be specifically executed: the search result of the search sentence corresponding to the maximum search order is calculated. The click information includes the click of the search result item of the question and answer category The first ratio between the number and the total number of clicks on the search result item; if the total number of clicks on the search result item is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, It is determined that the search sentences included in the target search sentence set have question and answer attributes.

Optionally, when the processor 401 executes the selection of multiple target search sentence sets from the search sentence database, it may specifically perform the following steps: determine from the search sentence database the search sentences whose occurrence times are greater than the preset second number threshold Set, and use the determined set of search sentences with the number of occurrences greater than the second number threshold as the multiple target search sentence sets; or, determine the number of occurrences from the search sentence database and search in the search sentence database Search sentence sets whose second ratio between the total number of sentences is greater than a preset second proportion threshold, and use the determined search sentence set whose second ratio is greater than the second proportion threshold as the multiple target searches Statement set

Optionally, when the processor 401 executes the selection of multiple target search sentence sets from the search sentence database, it may specifically perform the following steps: determine the application field information of the intent recognition model to be trained; A target sub-database is determined from multiple sub-databases included in the search sentence database; and the multiple target search sentence sets are selected from the target sub-database.

Optionally, before the processor 401 executes the training to obtain the intent recognition model by using the positive samples and negative samples corresponding to the multiple target search sentence sets, it may also perform the following steps: calculate the search sentence corresponding to the positive sample The absolute value of the difference between the number and the number of search sentences corresponding to the negative sample; determine whether the absolute value exceeds a preset third number threshold; if the absolute value exceeds the third number threshold, follow The preset sample balance rule processes the positive sample and/or the negative sample to obtain processed positive sample and negative sample;

When the processor 401 executes the training using the positive samples and negative samples corresponding to the multiple target search sentence sets to obtain the intention recognition model, the processor 401 may specifically perform the following steps: train using the processed positive samples and negative samples to obtain the intention Identify the model.

Wherein, the processor 401 may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), and application specific integrated circuits (Application Specific Integrated Circuits). Circuit, ASIC), ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.

The user interface 403 may include an input device and an output device. The input device may include a touch panel, a microphone, etc., and the output device may include a display (LCD, etc.), a speaker, etc.

The communication interface 404 may include a receiver and a transmitter for communicating with other devices.

The memory 402 may include a read-only memory and a random access memory, and provides instructions and data to the processor 401. A part of the memory 402 may also include a non-volatile random access memory. For example, the memory 402 may also store the aforementioned multiple search sentence sets, search event information associated with each search sentence set, and so on.

In specific implementation, the processor 401 described in the embodiment of the present application, etc., can execute the implementation described in the method embodiments shown in FIG. 1 to FIG. The implementation of the unit will not be repeated here.

The embodiment of the present application also provides a computer non-volatile readable storage medium, the computer non-volatile readable storage medium stores a computer program, and when the computer program is executed by a processor, it can realize FIGS. 1 to Part or all of the steps in the intention recognition method described in the corresponding embodiment of 2 can also realize the function of the intention recognition device in the embodiment shown in FIG. 3 or FIG. 4 of this application, and will not be repeated here.

The embodiments of the present application also provide a computer program product containing instructions, which when run on a computer, cause the computer to execute part or all of the steps in the above method.

The computer non-volatile readable storage medium may be the internal storage unit of the intent identification device described in any of the foregoing embodiments, such as the hard disk or memory of the intent identification device. The computer non-volatile readable storage medium may also be an external storage device of the intent identification device, for example, a plug-in hard disk equipped on the intent identification device, a smart media card (SMC), and a safe Digital (Secure Digital, SD) card, Flash Card (Flash Card), etc.

In this application, the term "and/or" is only an association relationship describing the associated objects, indicating that there can be three types of relationships, for example, A and/or B, which can mean: A alone exists, and both A and B exist. , There are three cases of B alone. In addition, the character "/" in this text generally indicates that the associated objects before and after are in an "or" relationship. In the various embodiments of the present application, the size of the sequence number of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not correspond to the implementation process of the embodiments of the present application. Constitute any limitation.

The above are only part of the implementation of this application, but the protection scope of this application is not limited to this. Anyone familiar with the technical field can easily think of various equivalents within the technical scope disclosed in this application. Modifications or replacements, these modifications or replacements shall be covered within the protection scope of this application.

Claims

An intention recognition method, characterized in that it includes:

Receive the target search sentence entered by the user;

Performing word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence, and the word segmentation result of the target search sentence includes a plurality of word segmentation constituting the target search sentence;

Input the word segmentation result of the target search sentence into a preset intent recognition model to obtain the intent recognition result corresponding to the target search sentence, and the intent recognition model is based on multiple target search sentence sets and the multiple targets The search event information is obtained by training the search event information associated with each target search sentence set in the search sentence set, where each target search sentence set includes at least one search sentence, and the search event information includes information about each search sentence in the at least one search sentence. Search order and/or click information of the search result of each search sentence, and the intention recognition result is used to indicate whether the target search sentence has a question and answer attribute;

If the intent recognition result indicates that the target search sentence has a question and answer attribute, then output the search result including the question and answer type search result item corresponding to the target search sentence.
The method according to claim 1, characterized in that, before inputting the word segmentation result of the target search sentence into a preset intention recognition model to obtain the intention recognition result corresponding to the target search sentence, the Methods also include:

Select multiple target search sentence sets from the search sentence database; wherein, the search sentence database records multiple search sentence sets and search event information associated with each search sentence set, and each search sentence set includes at least one Search sentence, the search event information includes a search order of each search sentence in the at least one search sentence and/or search result click information of each search sentence;

Word segmentation processing is performed on the search sentences included in each target search sentence set of the multiple target search sentence sets to obtain the word segmentation result of each target search sentence set, and the word segmentation result of each target search sentence set includes Multiple word segmentation of the search sentence constituting the target search sentence set;

According to the search event information associated with each target search sentence set, determine whether the search sentence included in each target search sentence set has a question and answer attribute;

Taking the word segmentation results of the target search sentence set with question and answer attributes in the multiple target search sentence sets as positive samples, and taking the word segmentation results of the target search sentence sets without question answering attributes in the multiple target search sentence sets as negative samples, The positive samples and negative samples corresponding to the multiple target search sentence sets are used for training to obtain an intention recognition model; wherein, the intention recognition model is used to identify whether the input search sentence has question and answer attributes.
The method according to claim 2, wherein the determining whether the search sentences included in the target search sentence set have question and answer attributes according to the search event information associated with the target search sentence set comprises:

Determine the search sentence corresponding to the largest search order in the at least one search sentence included in the target search sentence set according to the search order of each search sentence included in the search event information associated with the target search sentence set;

According to the click information of the search result of the search sentence corresponding to the maximum search order, it is determined whether the search sentence included in the target search sentence set has a question and answer attribute.
The method according to claim 3, wherein the search result click information includes the total number of clicks on search result items and the number of clicks on Q&A search result items; the search sentence corresponding to the maximum search order Click the information in the search result to determine whether the search sentence included in the target search sentence set has question and answer attributes, including:

Calculating the first ratio between the number of clicks on the search result items of the question and answer category and the total number of clicks on the search result items included in the search result click information of the search sentence corresponding to the maximum search order;

If the total number of clicks on the search result item is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, it is determined that the search sentence included in the target search sentence set has a question and answer attribute.
The method according to claim 2, wherein the determining whether the search sentences included in the target search sentence set have question and answer attributes according to the search event information associated with the target search sentence set comprises:

Determine the weighting coefficient corresponding to each search sentence in the at least one search sentence included in the target search sentence set, and the weighting coefficient of the search sentence with a higher search order in the at least one search sentence is higher than that of the search sentence with a lower search order Weighting factor

According to the weighting coefficient corresponding to each search sentence and the search result click information of each search sentence in the search event information associated with the target search sentence set, it is determined whether the search sentence included in the target search sentence set has a question and answer attribute.
The method according to claim 2, wherein the search event information includes search result click information of each search sentence in the at least one search sentence, and the search result click information includes the total number of clicks of search result items , The number of clicks on Q&A search result items and the browsing time of each clicked search result item;

The determining whether the search sentences included in the target search sentence set have question and answer attributes according to the search event information associated with the target search sentence set includes:

Filter out search result items whose browsing duration is less than the preset duration threshold, and determine the total number of clicks on the search result items remaining after filtering the search result item, and determine the number of clicks on the question-and-answer search result items remaining after filtering the search result item , And calculating the first ratio between the number of clicks on the remaining Q&A search result items and the total number of clicks on the remaining search result items;

If the total number of clicks on the remaining search result items is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, it is determined that the search sentences included in the target search sentence set have question and answer attributes.
The method according to any one of claims 2-6, wherein the selecting multiple target search sentence sets from a search sentence database comprises:

Determine the application domain information of the intent recognition model to be trained;

According to the application field information, a target sub-database is determined from a plurality of sub-databases included in the search sentence database, the sub-databases have a one-to-one correspondence with application fields, and each sub-database includes multiple search sentence sets under the corresponding application field And search event information associated with each search sentence set, the application field corresponding to the target sub-database is the same as the application field indicated by the application field information;

The multiple target search sentence sets are selected from the target sub-database.
An intention recognition device, which is characterized by comprising: an acquisition unit and a processing unit;

The acquiring unit is used to receive the target search sentence input by the user;

A processing unit, configured to perform word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence, and the word segmentation result of the target search sentence includes a plurality of word segmentation forming the target search sentence;

The processing unit is further configured to input the word segmentation result of the target search sentence into a preset intent recognition model to obtain an intent recognition result corresponding to the target search sentence, and the intent recognition model is based on multiple target searches Sentence sets and search event information associated with each target search sentence set in the multiple target search sentence sets, each target search sentence set includes at least one search sentence, and the search event information includes the at least one The search order of each search sentence in the search sentence and/or the click information of the search result of each search sentence, and the intention recognition result is used to indicate whether the target search sentence has a question and answer attribute;

The processing unit is further configured to output a search result including a question and answer type search result item corresponding to the target search sentence if the intention recognition result indicates that the target search sentence has a question and answer attribute.
The device according to claim 8, wherein:

The acquiring unit is also used to select multiple target search sentence sets from a search sentence database; wherein, the search sentence database records multiple search sentence sets and search event information associated with each search sentence set. Each search sentence set includes at least one search sentence, and the search event information includes the search order of each search sentence in the at least one search sentence and/or search result click information of each search sentence;

The processing unit is further configured to separately perform word segmentation processing on the search sentences included in each target search sentence set in the multiple target search sentence sets, so as to obtain the word segmentation results of each target search sentence set. The word segmentation result of the target search sentence set includes multiple word segmentation that constitute the search sentence of the target search sentence set;

The processing unit is further configured to determine, according to the search event information associated with each target search sentence set, whether the search sentence included in each target search sentence set has a question and answer attribute; and collect the multiple target search sentences The word segmentation result of the target search sentence set with the question and answer attribute is used as a positive sample, and the word segmentation result of the target search sentence set without the question and answer attribute in the plurality of target search sentence sets is used as a negative sample, and the multiple target search sentences are used The positive samples and negative samples corresponding to the set are trained to obtain an intent recognition model; wherein, the intent recognition model is used to identify whether the input search sentence has question and answer attributes.
The device according to claim 9, wherein:

The processing unit is specifically configured to determine the maximum search order among the at least one search sentence included in the target search sentence set according to the search order of each search sentence included in the search event information associated with the target search sentence set The corresponding search sentence;

According to the click information of the search result of the search sentence corresponding to the maximum search order, it is determined whether the search sentence included in the target search sentence set has a question and answer attribute.
The device according to claim 10, wherein the search result click information includes the total number of clicks on search result items and the number of clicks on Q&A search result items; When the search result click information of the search sentence corresponding to the search order is used to determine whether the search sentence included in the target search sentence set has a question and answer attribute, it is specifically used for:

Calculating the first ratio between the number of clicks on the search result items of the question and answer category and the total number of clicks on the search result items included in the search result click information of the search sentence corresponding to the maximum search order;

If the total number of clicks on the search result item is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, it is determined that the search sentence included in the target search sentence set has a question and answer attribute.
The device according to claim 9, wherein:

The processing unit is specifically configured to determine a weighting coefficient corresponding to each search sentence in the at least one search sentence included in the target search sentence set, and a search sentence with a higher search order in the at least one search sentence has a higher weighting coefficient The weighting coefficient of the search sentence with a small search order; the search result click information of each search sentence in the search event information associated with each search sentence and the search event information associated with the target search sentence set determines the target search Whether the search sentences included in the sentence set have question and answer attributes.
The device according to claim 9, wherein the search event information includes search result click information of each search sentence in the at least one search sentence, and the search result click information includes the total number of clicks of search result items , The number of clicks on Q&A search result items and the browsing time of each clicked search result item;

The processing unit is specifically configured to filter out search result items whose browsing duration is less than a preset duration threshold, determine the total number of clicks on the search result items remaining after filtering the search result item, and determine the question and answer remaining after filtering the search result item The number of clicks on search result items of the category, and the first ratio between the number of clicks on the remaining Q&A search result items and the total number of clicks on the remaining search result items; if the remaining search result items are clicked The total number is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, it is determined that the search sentences included in the target search sentence set have question and answer attributes.
The device according to any one of claims 9-13, wherein:

The acquiring unit is specifically configured to determine the application field information of the intention recognition model to be trained; according to the application field information, a target sub-database is determined from a plurality of sub-databases included in the search sentence database, and the sub-database and the application Fields have one-to-one correspondence, each sub-database includes multiple search sentence sets under the corresponding application field and search event information associated with each search sentence set, and the application field corresponding to the target sub-database corresponds to the application indicated by the application field information The fields are the same; the multiple target search sentence sets are selected from the target sub-database.
An intention recognition device, characterized by comprising a processor and a memory, the processor and the memory are connected to each other, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured Used to call the program instructions, perform the following steps:

Receive the target search sentence entered by the user;

Performing word segmentation processing on the target search sentence to obtain a word segmentation result of the target search sentence, and the word segmentation result of the target search sentence includes a plurality of word segmentation constituting the target search sentence;

Input the word segmentation result of the target search sentence into a preset intent recognition model to obtain the intent recognition result corresponding to the target search sentence, and the intent recognition model is based on multiple target search sentence sets and the multiple targets The search event information is obtained by training the search event information associated with each target search sentence set in the search sentence set, where each target search sentence set includes at least one search sentence, and the search event information includes information about each search sentence in the at least one search sentence. Search order and/or click information of the search result of each search sentence, and the intention recognition result is used to indicate whether the target search sentence has a question and answer attribute;

If the intent recognition result indicates that the target search sentence has a question and answer attribute, then output the search result including the question and answer type search result item corresponding to the target search sentence.
The device according to claim 15, wherein the processor is executing the input of the word segmentation result of the target search sentence into a preset intention recognition model to obtain the intention recognition corresponding to the target search sentence Before the result, perform the following steps:

Select multiple target search sentence sets from the search sentence database; wherein, the search sentence database records multiple search sentence sets and search event information associated with each search sentence set, and each search sentence set includes at least one Search sentence, the search event information includes a search order of each search sentence in the at least one search sentence and/or search result click information of each search sentence;

Word segmentation processing is performed on the search sentences included in each target search sentence set of the multiple target search sentence sets to obtain the word segmentation result of each target search sentence set, and the word segmentation result of each target search sentence set includes Multiple word segmentation of the search sentence constituting the target search sentence set;

According to the search event information associated with each target search sentence set, determine whether the search sentence included in each target search sentence set has a question and answer attribute;

Taking the word segmentation results of the target search sentence set with question and answer attributes in the multiple target search sentence sets as positive samples, and taking the word segmentation results of the target search sentence sets without question answering attributes in the multiple target search sentence sets as negative samples, The positive samples and negative samples corresponding to the multiple target search sentence sets are used for training to obtain an intention recognition model; wherein, the intention recognition model is used to identify whether the input search sentence has question and answer attributes.
The device according to claim 16, wherein when the processor executes the search event information associated with the target search sentence set to determine whether the search sentence included in the target search sentence set has a question and answer attribute , Perform the following steps:

Determine the search sentence corresponding to the largest search order in the at least one search sentence included in the target search sentence set according to the search order of each search sentence included in the search event information associated with the target search sentence set;

According to the click information of the search result of the search sentence corresponding to the maximum search order, it is determined whether the search sentence included in the target search sentence set has a question and answer attribute.
The device according to claim 17, wherein the search result click information includes the total number of clicks on search result items and the number of clicks on Q&A search result items; When the search result click information of the search sentence corresponding to the search order is clicked to determine whether the search sentence included in the target search sentence set has a question and answer attribute, the following steps are specifically performed:

Calculating the first ratio between the number of clicks on the search result items of the question and answer category and the total number of clicks on the search result items included in the search result click information of the search sentence corresponding to the maximum search order;

If the total number of clicks on the search result item is greater than the preset first number threshold, and the first ratio is greater than the preset first ratio threshold, it is determined that the search sentence included in the target search sentence set has a question and answer attribute.
The device according to claim 16, wherein when the processor executes the search event information associated with the target search sentence set to determine whether the search sentence included in the target search sentence set has a question and answer attribute , Perform the following steps:

Determine the weighting coefficient corresponding to each search sentence in the at least one search sentence included in the target search sentence set, and the weighting coefficient of the search sentence with a higher search order in the at least one search sentence is higher than that of the search sentence with a lower search order Weighting factor

According to the weighting coefficient corresponding to each search sentence and the search result click information of each search sentence in the search event information associated with the target search sentence set, it is determined whether the search sentence included in the target search sentence set has a question and answer attribute.
A computer nonvolatile readable storage medium, wherein the computer nonvolatile readable storage medium stores a computer program, the computer program includes program instructions, and the program instructions when executed by a processor The processor is caused to execute the method according to any one of claims 1-7.