Intention recognition method, device, electronic device and storage medium

Abstract

The embodiment of the present invention discloses an intent recognition method, device, electronic device and storage medium. The intent recognition method includes: obtaining first target intent sample data according to original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result ranking data; performing entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words; performing logical combination on the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary; constructing a first intent recognition model according to the intent matching generalized dictionary; when determining that the input data to be recognized is non-long-tail input data, inputting the input data to be recognized into the first intent recognition model; and outputting the intent recognition result of the input data to be recognized according to the first intent recognition model. The technical solution of the embodiment of the present invention can improve the accuracy of intent understanding.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology such as information processing, and in particular to an intent recognition method, device, electronic device and storage medium. Background Art

[0002] Intent recognition can also be called intent detection. It is used to determine which field and which operation the input information is used to perform. Its essence belongs to the multi-classification problem and is widely used in intelligent interaction technologies such as search and human-computer interaction. One manifestation of intelligent interaction is that intelligent products or applications can understand needs through intent recognition and provide appropriate responses based on the needs.

[0003] An important part of intent recognition is query processing. Each query hides the real query intent. When understanding the query, many different strategies need to be used to explore the needs behind it. Therefore, how to correctly identify the query intent, analyze the content of interest, and display the most interesting content in limited resources is of great significance to improving the experience of intelligent interactive functions.

[0004] In the process of realizing the present invention, the inventors found that the prior art has the following defects: At present, the existing intent recognition methods basically adopt a one-size-fits-all principle when processing queries, and do not distinguish between the two different types of queries, long-tail queries and non-long-tail queries. The concentration of long-tail queries is low, but the cumulative number is close to infinity. Although the search volume of a single long-tail query is not large, it has a long-tail effect, and the total search volume is comparable to the non-long-tail query volume at the head. If the processing methods of the two different types of queries, long-tail queries and non-long-tail queries, are not distinguished, and a unified processing method is used to understand the query, the accuracy of intent understanding will be low. Summary of the invention

[0005] Embodiments of the present invention provide an intent recognition method, device, electronic device, and storage medium, which can improve the accuracy of intent understanding.

[0006] According to one aspect of the present invention, there is provided a method for identifying intent, comprising:

[0007] Acquire first target intent sample data according to the original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result ranking data;

[0008] Performing entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words;

[0009] logically combining the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary;

[0010] Constructing a first intent recognition model according to the intent matching generalized dictionary;

[0011] In the case where it is determined that the input data to be identified is non-long-tail input data, inputting the input data to be identified into the first intent recognition model;

[0012] Output the intent recognition result of the input data to be recognized according to the first intent recognition model.

[0013] According to another aspect of the present invention, there is provided an intention recognition device, comprising:

[0014] A first sample data acquisition module is used to acquire first target intent sample data according to the original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result ranking data;

[0015] An abstract generalized entity word acquisition module is used to perform entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words;

[0016] An intent matching generalized dictionary generation module, used for logically combining the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary;

[0017] A first intent recognition model building module, used to build a first intent recognition model according to the intent matching generalized dictionary;

[0018] An input data to be identified input module, used for inputting the input data to be identified into the first intention recognition model when it is determined that the input data to be identified is non-long-tail input data;

[0019] An intention recognition result output module is used to output the intention recognition result of the input data to be recognized according to the first intention recognition model.

[0020] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0021] at least one processor; and

[0022] a memory communicatively connected to the at least one processor; wherein,

[0023] The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the intention recognition method described in any embodiment of the present invention.

[0024] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the intention recognition method described in any embodiment of the present invention when executed.

[0025] The embodiment of the present invention obtains the first target intent sample data including non-long-tail input sample data and intent matching result ranking data according to the original intent sample data, performs entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words, and logically combines the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary, thereby constructing a first intent recognition model according to the intent matching generalized dictionary, and using the first intent recognition model to perform intent recognition on input data to be recognized whose input data result is non-long-tail input data, and outputs the intent recognition result of the input data to be recognized, thereby solving the problem of low intent understanding accuracy in existing intent recognition methods and improving the accuracy of intent understanding.

[0026] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 is a flow chart of an intention recognition method provided in Embodiment 1 of the present invention;

[0029] Figure 2 is a flow chart of an intention recognition method provided by Embodiment 2 of the present invention;

[0030] Figure 3 is a schematic diagram of a BERT model training process provided by Embodiment 2 of the present invention;

[0031] Figure 4 is a schematic diagram of an intention recognition device provided by a fourth embodiment of the present invention;

[0032] Figure 5 A schematic diagram of the structure of an electronic device that can be used to implement an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] The embodiments of the present invention can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0036] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0037] Embodiment 1

[0038] Figure 1It is a flowchart of an intent recognition method provided in the first embodiment of the present invention. This embodiment can be applied to the situation where an intent recognition model is constructed based on non-long-tail input sample data to perform intent recognition on non-long-tail input data. The method can be executed by an intent recognition device, which can be implemented by software and / or hardware and can generally be integrated in an electronic device. The electronic device can be a terminal device or a server device. The embodiment of the present invention does not limit the specific device type of the electronic device.

[0039] Correspondingly, such as Figure 1 As shown, the method includes the following operations:

[0040] S110. Obtain first target intent sample data according to original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result sorting data.

[0041] Among them, the first target intent sample data may be sample data used to construct the first intent recognition model. The original intent sample data may be the full amount of historical intent data. Optionally, the intent data may be user intent data, such as user query data, or intent data automatically generated by a device or program, such as a data search instruction issued by a device or query data issued by a simulated real user, etc. The embodiment of the present invention does not limit the data type and generation method of the intent data. The input sample data is also the sample data that needs to be understood by the intent, for example, it may be query data input by a user, or it may be query data input by a device or program, etc. The input sample data may be text-type data or voice-type data, and the embodiment of the present invention does not limit the data type of the user input sample data. It is understandable that long-tail data refers to non-target data but related to target data, and may also bring combined data of search traffic. Non-long-tail data refers to target data. Non-long-tail input sample data may be input sample data in a non-long-tail form, that is, non-long-tail input sample data may be directly used as a keyword or directly segmented to obtain multiple keywords for intent understanding. The intent matching result ranking data may be a ranking result of feedback data obtained after understanding the intent of non-long-tail user input sample data.

[0042] In the embodiment of the present invention, the intent may be any type of intent, such as but not limited to search intent and interaction intent, etc. An object with intent, such as a user, a device, or a program, may be referred to as an intent output object.

[0043] Exemplarily, in the field of intelligent search technology, intent can be a search intent. When the intent output object needs to search for relevant content on the network or in an application, the search statement provided for the intent output object can identify the search intent of the intent output object based on the search statement, so as to recommend relevant content to the intent output object based on the search intent of the intent output object. Correspondingly, the non-long-tail input sample data can be the non-long-tail search statement of the intent output object, and the intent matching result ranking data can be the operation ranking data of the intent output object based on the feedback intention recognition result. In a specific example, assuming that the non-long-tail input sample data is the function module search data of the APP, the intent matching result ranking data can be the ranking data of the user's click frequency on the function module feedback of the APP for the function module search data.

[0044] Exemplarily, in the field of intelligent interactive technology, the intention may be a dialogue intention or an interaction intention. For example, in an intelligent question-and-answer system, the intention of the intended output object can be identified based on the sentence (which may be a text type or a voice type, etc.) input by the intended output object, and a suitable response can be provided for the intended output object. Accordingly, the non-long-tail input sample data may be the dialogue sentences of the intended output object, and the intention matching result ranking data may be the operation ranking data of the intention output object based on the feedback intention recognition result. In a specific example, assuming that the non-long-tail input sample data is the dialogue voice data input by the user to the intelligent question-and-answer system, the intention matching result ranking data may be the ranking data of the user's recognition of the response voice result fed back by the intelligent question-and-answer system for the dialogue voice data.

[0045] S120: Perform entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words.

[0046] The abstract generalized entity word may be a generalized data structure constructed by abstracting entity words from non-long-tail input sample data.

[0047] After obtaining the non-long-tail input sample data in the first target intent sample data, entity abstraction can be performed on the non-long-tail input sample data. The so-called entity abstraction is to extract entity words from the non-long-tail input sample data to construct abstract generalized entity words based on the extracted entity words.

[0048] In a specific example, taking an e-commerce APP as an example, when the non-long-tail input sample data input by the user is "original wood pulp toilet paper", the abstract generalized entity word that can be obtained based on the non-long-tail input sample data is "original wood pulp #commodity#".

[0049] S130: logically combine the abstract generalized entity words and the intention matching result ranking data to generate an intention matching generalized dictionary.

[0050] The intent matching generalized dictionary can provide intent matching result ranking data for non-long-tail input sample data to determine the final intent understanding result of the non-long-tail input sample data. That is, the intent matching generalized dictionary can be a structured dictionary used to search for intent understanding results for non-long-tail input sample data.

[0051] Correspondingly, after obtaining the corresponding abstract generalized entity words based on the non-long-tail input sample data, the data can be sorted according to the abstract generalized entity words corresponding to the non-long-tail input sample data and the intent matching results, and a matching dictionary query unit is constructed for each non-long-tail input sample data, and then an intent matching generalized dictionary is constructed based on the matching dictionary query units constructed for each non-long-tail input sample data.

[0052] In an embodiment of the present invention, the intent matching generalized dictionary can use the abstract generalized entity words obtained by entity abstraction of non-long-tail input sample data as the benchmark matching unit, and use the intent matching result ranking data of the same non-long-tail input sample data as the alternative intent understanding result of the benchmark matching unit, thereby combining the benchmark matching unit and each alternative intent understanding result of the same non-long-tail input sample data into a dictionary query unit of the non-long-tail input sample data in the intent matching generalized dictionary.

[0053] In a specific example, assuming that the non-long-tail input sample data of a medical APP is "What is the cause of leg cramps when sleeping at night?", the abstract generalized entity word obtained for the non-long-tail input sample data can be "What is the cause of #body part##disease# when sleeping at night?", and the intent matching result ranking data of the non-long-tail input sample data is: "Function Module 01": 30; "Function Module 02": 20; "Function Module 03": 10. Among them, the subsequent field value of each function module can be the number of times the user has clicked on the function module in history. For example, for "Function Module 01": 30, it means that when each user enters the search query "What is the cause of leg cramps when sleeping at night" in the medical APP, the user clicks on function module 01 30 times in the search results of each function module fed back by the medical APP. It can be understood that the more historical clicks, the more the function module matches the user's search intent. Correspondingly, the above-mentioned abstract generalized entity words and intent matching result ranking data are logically combined to obtain the dictionary query unit corresponding to the non-long-tail input sample data "What is the cause of leg cramps when sleeping at night", and its data structure is as follows:

[0054] “What is the cause of #bodypart#disease#when sleeping at night”:

[0055] {“Functional Module 01”:30

[0056] "Functional Module 02": 20

[0057] "Functional Module 03":10}

[0058] It is understandable that the intent matching generalized dictionary can be set according to the field, such as building a corresponding intent matching generalized dictionary for a technical field. Alternatively, the intent matching generalized dictionary can also involve multiple fields at the same time, which is not limited in the embodiment of the present invention.

[0059] S140: Construct a first intent recognition model according to the intent matching generalized dictionary.

[0060] The first intent recognition model is used to recognize the intent of non-long-tail input data.

[0061] In an embodiment of the present invention, the first intent recognition model is constructed according to the intent matching generalized dictionary. The intent matching generalized dictionary can be directly used as the first intent recognition model to perform intent recognition on non-long-tail input data to obtain the final intent understanding result.

[0062] S150. When it is determined that the input data to be identified is non-long-tail input data, input the input data to be identified into the first intent recognition model.

[0063] S160. Output the intent recognition result of the input data to be recognized according to the first intent recognition model.

[0064] Correspondingly, if it is determined that the input data to be identified is non-long-tail input data, the input data to be identified can be input into the constructed first intent recognition model to identify the non-long-tail input data to be identified through the first intent recognition model.

[0065] It can be seen that by using the intent matching generalized dictionary constructed by sorting the non-long-tail input sample data and its matching intent matching results as the first intent recognition model, the first intent recognition model can be used to perform intent recognition on the non-long-tail input data, thereby improving the accuracy of intent understanding of the non-long-tail input data.

[0066] The embodiment of the present invention obtains the first target intent sample data including non-long-tail input sample data and intent matching result ranking data according to the original intent sample data, performs entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words, and logically combines the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary, thereby constructing a first intent recognition model according to the intent matching generalized dictionary, and using the first intent recognition model to perform intent recognition on input data to be recognized whose input data result is non-long-tail input data, and outputs the intent recognition result of the input data to be recognized, thereby solving the problem of low intent understanding accuracy in existing intent recognition methods and improving the accuracy of intent understanding.

[0067] Embodiment 2

[0068] Figure 2 is a flowchart of an intent recognition method provided by the second embodiment of the present invention. This embodiment is specific based on the above embodiment. In this embodiment, a plurality of specific optional implementation methods are provided for obtaining the first target intent sample data according to the original intent sample data, performing entity abstraction on the non-long-tail input sample data, constructing the first intent recognition model according to the intent matching generalized dictionary, and constructing the second intent recognition model and the target intent recognition model. Figure 2 As shown, the method of this embodiment may include:

[0069] S210 , filtering the non-long-tail input sample data from the original intent sample data according to a non-long-tail input data filtering rule.

[0070] The non-long-tail input data screening rule is a rule for screening non-long-tail input data.

[0071] Optionally, the non-long-tail input data screening rule can be used to limit the number of keywords in the data, that is, when the number of keywords in the original intent sample data is greater than a certain threshold, the data is long-tail input sample data; when the number of keywords in the original intent sample data is less than or equal to a certain threshold, the data is non-long-tail input sample data. The threshold for dividing long-tail input data and non-long-tail input data can be set according to actual needs, such as 20, etc., and the embodiment of the present invention does not limit the specific value of the threshold.

[0072] S220: Obtain associated intention feedback data of the non-long-tail input sample data.

[0073] Among them, the associated intent feedback data can be data fed back on non-long-tail input sample data, as well as relevant statistical data of the fed back data, etc.

[0074] S230. Sort the associated intention feedback data to obtain the intention matching result sorting data.

[0075] After the non-long-tail input sample data is obtained by screening the original intent sample data, the associated intent feedback data of the non-long-tail input sample data can be further obtained, and the obtained associated intent feedback data can be sorted to obtain intent matching result sorting data. Optionally, when sorting the associated intent feedback data, it can be sorted in descending order.

[0076] In a specific example, assuming that the associated intent feedback data is the click frequency data of the application function module, the click frequency data of the application function module can be sorted in order from high to low click frequency to obtain the intent matching result sorting data. Assuming that the associated intent feedback data is the response interaction data of the intelligent interactive system to the input data of the intent output object, such as the machine interaction frequency data of the query in the human-computer dialogue, the response interaction data can be sorted in order from high to low interaction frequency to obtain the intent matching result sorting data. It can be understood that the higher the click frequency or machine interaction frequency, the higher the recognition of the intent output object for the intent recognition result.

[0077] In a specific example, assuming that the non-long-tail input sample data is "Why do my legs cramp when I sleep at night?", in the historical data, the APP can feedback 3 matching function modules as related intent feedback data for the non-long-tail input sample data, namely function module 01, function module 02 and function module 03, among which function module 01 has been clicked 30 times in history, function module 02 has been clicked 20 times in history, and function module 03 has been clicked 10 times in history. Then, the related intent feedback data is sorted, and the intent matching result sorting data can be: "Function module 01": 30; "Function module 02": 20; "Function module 03": 10.

[0078] S240 , performing entity abstraction on the non-long-tail input sample data according to an entity word dictionary to obtain initial abstract entity words.

[0079] The entity word dictionary may be a dictionary composed of entity words. The initial abstract entity words may be entity words abstracted from non-long-tail input sample data.

[0080] S250: Classify and group the initial abstract entity words to obtain the abstract generalized entity words.

[0081] In an embodiment of the present invention, entity abstraction can first be performed on non-long-tail input sample data according to an entity word dictionary, and the entity words constituting the non-long-tail input sample data can be used as initial abstract entity words. The initial abstract entity words obtained by the abstraction can be further classified and grouped to obtain abstract generalized entity words.

[0082] In a specific example, taking a medical APP as an example, the entity word dictionary may include, but is not limited to, entity words such as diseases, symptoms, departments, commodities, and body parts. Assuming that the input sample data is "virgin wood pulp toilet paper", the input sample data can be abstracted as "virgin wood pulp#commodity#", and assuming that the input sample data is "zinc gluconate oral solution", the input sample data can be abstracted as "drug##body part#oral solution". Furthermore, it is necessary to classify and group the initial abstract entity words after abstraction, for example: the initial abstract entity words of "drug##body part#oral solution" can be classified as "#drug#|#body part#". That is, "#drug#|#body part#" is a type of abstract generalized entity word, which can be used to match input sample data of types such as "zinc gluconate oral solution" and "wormwood and ginger foot patch".

[0083] S260: Logically combine the abstract generalized entity words and the intention matching result ranking data to generate an intention matching generalized dictionary.

[0084] In a specific example, let's take the user query sample data including body parts and disease entity words as an example. Assume that there are three user query sample data: "What is the reason for leg cramps when sleeping at night?", "What medicine is good for duodenal erosion?", "What is the reason for two bouts of fever in the back in one day? What medicine should I take to relieve it?" Based on these three user query sample data, the intent matching generalized dictionary that can be constructed is as follows:

[0085]

[0086] The data results of the above intent matching generalized dictionary have strong generalization for queries that have both body parts and disease entities.

[0087] The above-mentioned intent matching generalized dictionary can perform intent recognition on non-long-tail input data based on the DFA (Deterministic Finite Automaton) algorithm.

[0088] S270: Construct an input data edit distance calculation module according to the dictionary elements of the intent matching generalized dictionary.

[0089] Among them, the dictionary element is also the dictionary query unit of the intent matching generalized dictionary. Exemplarily, the dictionary element of the intent matching generalized dictionary can be, for example, "sleeping at night#body part##disease#what's going on" and "#body part#two bouts of#disease#a day what's going on what medicine to take to relieve" and so on. The input data edit distance calculation module can be used to calculate the edit distance between the input data and the dictionary element. It can be understood that the smaller the edit distance, the closer the input data is to the dictionary element, that is, the more the input data matches the dictionary element.

[0090] It is understandable that dictionary elements can be used as a query matching benchmark to match input data. Exemplarily, the dictionary element constructed based on the user's historical behavior can be: "What is the matter with sleeping at night#body part##disease#". When the input non-long-tail data is "What's wrong with leg cramps when sleeping at night" or "What's the matter with leg cramps when sleeping at night", it can be considered as a generalization of the dictionary element "What is the matter with sleeping at night#body part##disease#". That is, the dictionary elements matched by "What is the matter with leg cramps when sleeping at night" or "What's the matter with leg cramps when sleeping at night" in the intent matching generalized dictionary are: "What is the matter with sleeping at night#body part##disease#".

[0091] After obtaining the intent matching generalized dictionary, an input data edit distance calculation module can be constructed according to the element structure of the dictionary elements of the intent matching generalized dictionary to calculate the edit distance (also known as similarity) between the input data and the dictionary elements through the input data edit distance calculation module.

[0092] S280. Construct the first intent recognition model according to the input data edit distance calculation module and the intent matching generalized dictionary.

[0093] Among them, the first intent recognition model is used to identify the intent of non-long-tail input data.

[0094] In an embodiment of the present invention, the first intent recognition model may include two modules: an input data edit distance calculation module and an intent matching generalization dictionary. In addition, the first intent recognition model may also include an entity word dictionary for entity abstraction. The first intent recognition model can be used to perform intent recognition on non-long-tail input data.

[0095] Specifically, the first intent recognition model can first perform entity abstraction on the input query of the intent output object based on the entity word dictionary to obtain the initial abstract entity word, and perform entity classification and grouping on the initial abstract entity word to obtain the final abstract generalized entity word. Then, the first intent recognition model can calculate the edit distance between the abstract generalized entity word and each dictionary element in the intent matching generalized dictionary. Finally, the first intent recognition model uses the intent included in the dictionary element with the smallest edit distance between the intent matching generalized dictionary and the abstract generalized entity word as the intent of the input query. It can be understood that if there are multiple intents included in the dictionary element with the smallest edit distance between the intent matching generalized dictionary and the abstract generalized entity word, the first intent can also be selected as the intent of the input query based on the ranking results of each intent.

[0096] S290: Construct a second intention recognition model.

[0097] Among them, the second intent recognition model can be used to perform intent recognition on long-tail input data to obtain the final intent understanding result.

[0098] In an optional embodiment of the present invention, constructing a second intent recognition model may include: pre-training a preset neural network model based on pre-trained sample data to obtain a pre-trained neural network model; obtaining second target intent sample data based on the original intent sample data; wherein the second target intent sample data includes long-tail input sample data and intent labeling result data; training the pre-trained neural network model based on the second target intent sample data to obtain a second intent recognition model.

[0099] Among them, the preset neural network model can be any type of neural network model that can realize the intent recognition function. The pre-trained neural network model can be a neural network model obtained by pre-training the preset neural network model. The second target intent sample data can be sample data used to formally train the second intent recognition model. The long-tail input sample data can be input sample data in a long-tail form, which has the characteristics of semantic complexity. The intent labeling result data can be data that is pre-labeled with the intent that matches the long-tail input sample data.

[0100] In an embodiment of the present invention, a neural network model can be used to perform intent recognition for long-tail input sample data. Specifically, the pre-trained neural network model can be pre-trained using pre-trained sample data to train the data understanding ability of the pre-trained neural network model to obtain a pre-trained neural network model. After the pre-training is completed, the second target intent sample data pre-trained neural network model can be further used for training to obtain a second intent recognition model.

[0101] In a specific example, assuming that the preset neural network model is a BERT (Bidirectional Encoder Representation from Transformers, a language representation model) model, pre-training the preset neural network model may include two pre-training tasks, one is an MLM (Masked Language Model) pre-training task, and the other is an NSP (Next Sentence Prediction) pre-training task. Among them, the MLM pre-training task can be understood as a cloze task, which can randomly mask a certain number of words (such as 15% of the words in each sentence) in each sentence, and use its context to make predictions. For example, the pre-training sample data "my dog ​​is hairy" is converted to "my dog ​​is [MASK]". Here, "hairy" is masked. Then, an unsupervised learning method is used to predict what the word at the mask position is. The NSP pre-training task can be understood as a text matching task. Specifically, some sentence pairs A and B can be selected, of which 50% of the data B is one of the sentences of A, and the remaining 50% of the data B is randomly selected from the corpus, so that the network can learn the correlation therein. For example, assume that sentence A is: Simplify the workflow of search operation and maintenance personnel and improve the work efficiency of operation and maintenance personnel. One of the short sentences of sentence B can be one of the short sentences of sentence A, and another short sentence of sentence B can be a randomly selected sentence segment. For example, sentence B can be: Simplify the workflow of search operation and maintenance personnel and need to eat on time. The above pre-training process can enable the pre-trained neural network model to understand the relationship between the two sentences, so that the pre-trained neural network model can better adapt to the above data processing tasks.

[0102] In a specific example, assuming that the preset neural network model is a BERT model, when the BERT model is trained according to the second target intent sample data, the CrossEntropy loss function and BP (BackPropagation) propagation mechanism can be used to allow the model to autonomously learn and update the network weight parameters to implement the training process. The trained BERT model is used as the second intent recognition model. The training process of the BERT model can be referred to Figure 3 As shown in the figure. BERT is a powerful pre-trained model that uses Transformers as a feature extractor. Given its huge number of parameters and superb feature representation capabilities, it can learn deep semantic information in text. Using BERT to embed long-tail input data can map text information to a high-dimensional vector space, and use an embedding vector to represent the semantic information of the long-tail input data.

[0103] S2110. Construct a target intent recognition model based on the first intent recognition model and the second intent recognition model.

[0104] Among them, the target intent recognition model is a model that can perform intent recognition on any type of input data.

[0105] S2120: Obtain input data to be identified, classify the input data to be identified, and obtain an input data classification result.

[0106] The input data classification result includes long-tail input data and non-long-tail input data.

[0107] The input data to be identified may be input data that needs to be identified. For example, it may be query data input by a user in real time, or it may be query data input by a device or program in real time. The input data to be identified may be text data or voice data. The embodiment of the present invention does not limit the data type of the input data to be identified. The classification result of the input data is also the classification result of the input data to be identified.

[0108] After obtaining the input data to be identified, in order to determine the model for performing intent recognition on the input data to be identified, the input data to be identified may be first classified to determine whether the input data to be identified is long-tail input data or non-long-tail input data.

[0109] In an optional embodiment of the present invention, classifying the input data to be identified may include: when it is determined that the data length of the input data to be identified is less than or equal to a preset data length threshold, determining that the input data classification result of the input data to be identified is non-long-tail input data; when it is determined that the data length of the input data to be identified is greater than the preset data length threshold, determining that the input data classification result of the input data to be identified is long-tail input data.

[0110] Among them, the preset data length threshold can be a length threshold used to divide long-tail data and non-long-tail data. Exemplarily, the preset data length threshold can be set to 20 or 25, etc., and can be set according to actual needs. The embodiment of the present invention does not limit the specific value of the preset data length threshold.

[0111] Specifically, the data length of the input data to be identified can be determined so as to classify the input data to be identified by the data length. The data length can be the number of words or characters in the input data to be identified, such as the data length of the input data to be identified, "What causes leg cramps when sleeping at night?", is 12. Accordingly, if it is determined that the data length of the input data to be identified is less than or equal to the preset data length threshold, the input data classification result of the input data to be identified can be determined as non-long-tail input data; otherwise, the input data classification result of the input data to be identified is determined as long-tail input data.

[0112] S2130, determine whether the input data to be identified is non-long-tail input data, if so, execute S2140, otherwise, execute S2150.

[0113] S2140. Input the input data to be identified into the first intent recognition model to output an intent recognition result of the input data to be identified according to the first intent recognition model.

[0114] S2150. Input the input data to be recognized into the second intent recognition model to output the intent recognition result of the input data to be recognized according to the second intent recognition model.

[0115] Specifically, the first intention recognition model of the target intention recognition model can be used to perform intent recognition on non-long-tail input data, and the second intention recognition model of the target intention recognition model can be used to perform intent recognition on long-tail input data, thereby obtaining the intent recognition result of the input data.

[0116] Among them, the first intent recognition model can be based on the DFA algorithm of a large number of behaviors, which can not only improve the efficiency of intent recognition, but also has good generalization. The second intent recognition model can well extract the semantic information implicit in the long-tail input data and better represent its semantic features. By differentially processing the long-tail and non-long-tail input data, the intent can be more accurately identified and understood.

[0117] By adopting the above technical solution, by respectively constructing the first intent recognition model and the second intent recognition model, and jointly forming a target intent recognition model based on the first intent recognition model and the second intent recognition model, different models can be used for intent recognition for long-tail input data and non-long-tail input data, which can improve the accuracy and efficiency of intent understanding, thereby improving user experience.

[0118] It should be noted that any arrangement and combination of the technical features in the above embodiments also falls within the protection scope of the present invention.

[0119] Embodiment 3

[0120] Figure 4is a schematic diagram of an intention recognition device provided by Embodiment 4 of the present invention, such as Figure 4 As shown, the device includes: a first sample data acquisition module 410, an abstract generalized entity word acquisition module 420, an intent matching generalized dictionary generation module 430, a first intent recognition model construction module 440, an input data input module 450 to be recognized, and an intent recognition result output module 460, wherein:

[0121] A first sample data acquisition module 410 is used to acquire first target intent sample data according to the original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result ranking data;

[0122] An abstract generalized entity word acquisition module 420 is used to perform entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words;

[0123] An intention matching generalized dictionary generation module 430 is used to logically combine the abstract generalized entity words and the intention matching result ranking data to generate an intention matching generalized dictionary;

[0124] A first intent recognition model building module 440, configured to build a first intent recognition model according to the intent matching generalized dictionary;

[0125] The input data to be identified input module 450 is used to input the input data to be identified into the first intention recognition model when it is determined that the input data to be identified is non-long-tail input data;

[0126] The intention recognition result output module 460 is used to output the intention recognition result of the input data to be recognized according to the first intention recognition model.

[0127] The embodiment of the present invention obtains the first target intent sample data including non-long-tail input sample data and intent matching result ranking data according to the original intent sample data, performs entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words, and logically combines the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary, thereby constructing a first intent recognition model according to the intent matching generalized dictionary, and using the first intent recognition model to perform intent recognition on input data to be recognized whose input data result is non-long-tail input data, and outputs the intent recognition result of the input data to be recognized, thereby solving the problem of low intent understanding accuracy in existing intent recognition methods and improving the accuracy of intent understanding.

[0128] Optionally, the first sample data acquisition module 410 is specifically used to: filter the non-long-tail input sample data from the original intention sample data according to the non-long-tail input data filtering rules; obtain the associated intention feedback data of the non-long-tail input sample data; sort the associated intention feedback data to obtain the intention matching result sorting data.

[0129] Optionally, the abstract generalized entity word acquisition module 420 is specifically used to: perform entity abstraction on the non-long-tail input sample data according to an entity word dictionary to obtain initial abstract entity words; and classify and group the initial abstract entity words to obtain the abstract generalized entity words.

[0130] Optionally, the first intent recognition model construction module 440 is specifically used to: construct an input data edit distance calculation module according to the dictionary elements of the intent matching generalized dictionary; and construct the first intent recognition model according to the input data edit distance calculation module and the intent matching generalized dictionary.

[0131] Optionally, the intention recognition device also includes: a preset neural network model pre-training module, which is used to pre-train the preset neural network model according to the pre-training sample data to obtain a pre-trained neural network model; a second target intention sample data acquisition module, which is used to obtain second target intention sample data according to the original intention sample data; wherein the second target intention sample data includes long-tail input sample data and intention marking result data; a second intention recognition model acquisition module, which is used to train the pre-trained neural network model according to the second target intention sample data to obtain a second intention recognition model; and a target intention recognition model construction module, which is used to construct a target intention recognition model based on the first intention recognition model and the second intention recognition model.

[0132] Optionally, the intention recognition device also includes: an input data acquisition module to be identified, used to obtain the input data to be identified; an input data classification module to be identified, used to classify the input data to be identified to obtain an input data classification result; wherein the input data classification result includes long-tail input data and non-long-tail input data.

[0133] Optionally, the input data classification module to be identified is specifically used to: when it is determined that the data length of the input data to be identified is less than or equal to a preset data length threshold, determine that the input data classification result of the input data to be identified is non-long-tail input data; when it is determined that the data length of the input data to be identified is greater than the preset data length threshold, determine that the input data classification result of the input data to be identified is long-tail input data.

[0134] The above-mentioned intention recognition device can execute the intention recognition method provided by any embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the intention recognition method provided by any embodiment of the present invention.

[0135] Since the intention recognition device introduced above is a device that can execute the intention recognition method in the embodiment of the present invention, based on the intention recognition method introduced in the embodiment of the present invention, a person skilled in the art can understand the specific implementation of the intention recognition device of the present embodiment and its various variations, so how the intention recognition device implements the intention recognition method in the embodiment of the present invention will not be described in detail here. As long as a person skilled in the art implements the device used in the intention recognition method in the embodiment of the present invention, it belongs to the scope of protection of this application.

[0136] Embodiment 4

[0137] Figure 5 A schematic diagram of the structure of an electronic device 10 that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0138] like Figure 5 As shown, the electronic device 10 includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 to the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0139] A number of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0140] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 executes the various methods and processes described above, such as the intent recognition method.

[0141] In some embodiments, the intent recognition method may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the intent recognition method described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to perform the intent recognition method in any other appropriate manner (e.g., by means of firmware).

[0142] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0143] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0144] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0145] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0146] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0147] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0148] Embodiment 5

[0149] Embodiment 5 of the present invention further provides a computer storage medium storing a computer program, wherein the computer program, when executed by a computer processor, is used to execute the intention recognition method described in any of the above embodiments of the present invention.

[0150] Among them, the intention recognition method includes: obtaining first target intention sample data according to original intention sample data; wherein, the first target intention sample data includes non-long-tail input sample data and intention matching result ranking data; performing entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words; performing logical combination on the abstract generalized entity words and the intention matching result ranking data to generate an intention matching generalized dictionary; constructing a first intention recognition model according to the intention matching generalized dictionary; when it is determined that the input data to be recognized is non-long-tail input data, inputting the input data to be recognized into the first intention recognition model; and outputting the intention recognition result of the input data to be recognized according to the first intention recognition model.

[0151] The computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM, or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0152] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, which carry computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0153] The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wire, optical cable, radio frequency (RF), etc., or any suitable combination of the foregoing.

[0154] Computer program code for performing the operations of the present invention may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0155] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0156] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intention recognition, characterized in that: include: Acquire first target intent sample data according to the original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result ranking data; Performing entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words; logically combining the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary; Constructing a first intent recognition model according to the intent matching generalized dictionary; Pre-training a preset neural network model according to pre-training sample data to obtain a pre-trained neural network model; Acquire second target intent sample data according to the original intent sample data; wherein the second target intent sample data includes long-tail input sample data and intent labeling result data; Training the pre-trained neural network model according to the second target intent sample data to obtain a second intent recognition model; Building a target intent recognition model according to the first intent recognition model and the second intent recognition model; In the case where it is determined that the input data to be identified is non-long-tail input data, the input data to be identified is input into a first intention recognition model in the target intention recognition model; and according to the first intention recognition model, an intention recognition result of the input data to be identified is output; When it is determined that the input data to be identified is long-tail input data, the input data to be identified is input into a second intention recognition model in the target intention recognition model; and the intention recognition result of the input data to be identified is output according to the second intention recognition model.

2. The method according to claim 1, characterized in that The obtaining of first target intent sample data according to the original intent sample data includes: Filtering the non-long-tail input sample data from the original intent sample data according to a non-long-tail input data filtering rule; Obtaining associated intention feedback data of the non-long-tail input sample data; The associated intent feedback data is sorted to obtain the intent matching result sorting data.

3. The method according to claim 1, characterized in that: The performing entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words includes: Perform entity abstraction on the non-long-tail input sample data according to the entity word dictionary to obtain initial abstract entity words; The initial abstract entity words are classified and grouped to obtain the abstract generalized entity words.

4. The method according to claim 1, characterized in that: The step of constructing a first intent recognition model according to the intent matching generalized dictionary includes: Constructing an input data edit distance calculation module according to the dictionary elements of the intent matching generalized dictionary; The first intent recognition model is constructed according to the input data edit distance calculation module and the intent matching generalization dictionary.

5. The method according to claim 1, characterized in that Before inputting the input data to be recognized into the first intention recognition model, the method further includes: Obtain input data to be identified; The input data to be identified is classified to obtain an input data classification result; wherein the input data classification result includes long-tail input data and non-long-tail input data.

6. The method according to claim 5, characterized in that The classifying the input data to be identified includes: In a case where it is determined that the data length of the input data to be identified is less than or equal to a preset data length threshold, determining that the input data classification result of the input data to be identified is non-long-tail input data; When it is determined that the data length of the input data to be identified is greater than the preset data length threshold, it is determined that the input data classification result of the input data to be identified is long-tail input data.

7. An intention recognition device, characterized in that: include: A first sample data acquisition module is used to acquire first target intent sample data according to the original intent sample data; wherein the first target intent sample data includes non-long-tail input sample data and intent matching result ranking data; An abstract generalized entity word acquisition module is used to perform entity abstraction on the non-long-tail input sample data to obtain abstract generalized entity words; An intent matching generalized dictionary generation module, used for logically combining the abstract generalized entity words and the intent matching result ranking data to generate an intent matching generalized dictionary; A first intent recognition model building module, used to build a first intent recognition model according to the intent matching generalized dictionary; A pre-trained neural network model acquisition module is used to pre-train a preset neural network model according to pre-training sample data to obtain a pre-trained neural network model; A second target intent sample data acquisition module, used to acquire second target intent sample data according to the original intent sample data; A second intent recognition model acquisition module, used to train the pre-trained neural network model according to the second target intent sample data to obtain a second intent recognition model; A target intention recognition model construction module, used to construct a target intention recognition model according to the first intention recognition model and the second intention recognition model; A non-long-tail input data input module to be identified, used for inputting the input data to be identified into a first intent recognition model in the target intent recognition model when determining that the input data to be identified is non-long-tail input data; A non-long-tail input data intention recognition result output module, used to output the intention recognition result of the input data to be recognized according to the first intention recognition model; A module for inputting long-tail input data to be identified, used for inputting the input data to be identified into a second intent recognition model in the target intent recognition model when determining that the input data to be identified is long-tail input data; The long-tail input data intention recognition result output module is used to output the intention recognition result of the input data to be recognized according to the second intention recognition model.

8. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can perform the intention recognition method according to any one of claims 1 to 6.

9. A computer storage medium, wherein the computer readable storage medium stores computer instructions, wherein the computer instructions are used to enable a processor to implement the intention recognition method according to any one of claims 1 to 6 when executed.

Images