A semantic understanding method and device fusing user features and electronic equipment

Abstract

The application relates to the field of semantic recognition, and provides a semantic understanding method and device fusing user features and electronic equipment. The method comprises the following steps: obtaining a text to be processed, inputting the text to be processed into a semantic understanding model, performing first semantic understanding on the text to be processed based on the semantic understanding model, and obtaining an initial semantic understanding result; determining whether second semantic understanding is needed according to the initial semantic understanding result; if it is determined that the second semantic understanding is needed, obtaining media information according to the initial semantic understanding result, obtaining content heat features of the media and user features of a target user according to the media information; inputting the content heat features and the user features into the semantic understanding model for processing, and obtaining a final semantic understanding result in combination with the initial semantic understanding result. The semantic understanding method fusing user features provided by the application can solve the problem that a general semantic understanding model cannot output an accurate recognition result for fuzzy text.

Description

Technical Field

[0001] This invention relates to the field of semantic recognition technology, and more specifically to a semantic understanding method, apparatus, and electronic device that integrates user features. Background Technology

[0002] Currently, semantic understanding is an important branch of NLP (Natural Language Processing). It is widely used in various human-computer dialogue systems, such as smart speakers and intelligent question-answering systems.

[0003] For user input obtained from various terminals, including text and voice, after being converted into standardized text data, semantic understanding is performed to parse out the user's domain, intent and key slots, and to provide corresponding services to the user.

[0004] Currently, the general implementation methods of semantic understanding in task-oriented dialogue systems can be divided into two types: First, constructing a multi-task joint model that combines domain recognition, intent recognition, and slot recognition, simultaneously identifying the domain, intent, and included slots of the target text. Second, constructing separate recognition models for each of the three tasks, performing domain recognition, intent recognition, and slot recognition separately for each target text. Current technologies generally only broadly integrate user features into the entire voice interaction process to provide personalized services, without considering integrating user features into the semantic understanding module to achieve personalized semantic understanding services for different users.

[0005] Analysis of existing technologies revealed the following shortcomings in current implementation solutions:

[0006] 1. User characteristics were not integrated into the semantic understanding process to achieve personalized semantic understanding capabilities.

[0007] 2. Identify waste of resources.

[0008] 3. The user segmentation granularity is not precise, and the user feature extraction does not conform to the actual business scenario.

[0009] 4. The semantic understanding was not combined with the popularity of the content. Summary of the Invention

[0010] This invention provides a semantic understanding method, apparatus, and electronic device that integrates user features to solve the technical problem that general semantic understanding models cannot output an accurate recognition result for fuzzy text.

[0011] In a first aspect, embodiments of the present invention provide a semantic understanding method that integrates user features, comprising:

[0012] The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result.

[0013] Determine whether a second semantic understanding is needed based on the initial semantic understanding results;

[0014] If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding result, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information;

[0015] The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding result.

[0016] In one embodiment, the step of inputting the content popularity features and user features into a semantic understanding model for processing, and combining the initial semantic understanding results to obtain the final semantic understanding result, specifically includes: inputting the content popularity features and user features into the semantic understanding model; superimposing the content popularity features and user features based on the semantic understanding model to obtain a superimposed bias term; and obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result.

[0017] In one embodiment, determining whether a second semantic understanding is needed based on the initial semantic understanding result specifically includes: determining whether the initial semantic understanding result is a preset intent and includes media asset information; if so, then searching in the media asset knowledge base based on the media asset information in the initial semantic understanding result to obtain search results; if the search results indicate that the slot identification result in the initial semantic understanding is fuzzy media asset, then it is determined that a second semantic understanding is needed.

[0018] In one embodiment, obtaining the content popularity features of the media assets and the user features of the target user based on the media asset information specifically includes: identifying the speech corresponding to the text to be processed based on voiceprint recognition to determine the target user; querying the content popularity features of the slot recognition results in the initial semantic understanding in different fields and the user features of the target user.

[0019] In one embodiment, the decoding layer of each subtask includes a domain decoding layer, an intent decoding layer, and a slot decoding layer. Obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result specifically includes: inputting the bias term into the semantic understanding model; superimposing the bias term and the domain decoding layer from the first semantic understanding to obtain a new domain decoding layer result; obtaining a new domain recognition result based on the new domain decoding layer result; generating an intent domain constraint vector and a slot domain constraint vector corresponding to the new domain recognition result; applying the intent domain constraint vector to the intent decoding layer and the slot domain constraint vector to the slot decoding layer to obtain the final semantic understanding result.

[0020] In one embodiment, determining whether a second semantic understanding is needed based on the initial semantic understanding result specifically includes: determining whether the initial semantic understanding result is a preset intent and includes media asset information; if it is determined that the initial semantic understanding result is a preset intent and includes media asset information, then based on the media asset information in the initial semantic understanding result, a search is performed in the media asset knowledge base to obtain the search result; if the search result indicates that the slot identification result in the initial semantic understanding is not a fuzzy media asset, then the identification ends and the identification result is output; if it is determined that the initial semantic understanding result is not a preset intent or does not include media asset information, then the identification ends and the identification result is output.

[0021] Secondly, embodiments of the present invention provide a semantic understanding apparatus that integrates user features, comprising:

[0022] The first semantic understanding module is used to acquire the text to be processed. The text to be processed is input into the semantic understanding model, and the semantic understanding model performs the first semantic understanding on the text to be processed to obtain the initial semantic understanding result.

[0023] The judgment module is used to determine whether a second semantic understanding is needed based on the initial semantic understanding results.

[0024] The second semantic understanding module is used to obtain media asset information based on the initial semantic understanding result if it is determined that a second semantic understanding is needed; obtain the content popularity features of the media asset and the user features of the target user based on the media asset information; input the content popularity features and the user features into the semantic understanding model for processing; and combine the initial semantic understanding result to obtain the final semantic understanding result.

[0025] Thirdly, embodiments of the present invention provide a terminal, including a memory, a transceiver, and a processor;

[0026] A memory for storing computer programs; a transceiver for sending and receiving data under the control of the processor; and a processor for reading the computer programs from the memory and performing the following operations:

[0027] The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result.

[0028] Determine whether a second semantic understanding is needed based on the initial semantic understanding results;

[0029] If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding result, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information;

[0030] The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding result.

[0031] Fourthly, embodiments of the present invention provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the semantic understanding method for fusing user features as described in the first aspect.

[0032] Fifthly, embodiments of the present invention provide a non-volatile storage medium storing computer-executable instructions, which are executed by a processor to implement the steps of the semantic understanding method for fusing user features as described in the first aspect.

[0033] This invention provides a semantic understanding method, apparatus, and electronic device that integrates user features. By combining user voiceprint information and user historical behavior, user features are extracted and integrated into a semantic understanding model to assist in the semantic understanding of user text. This solves the problem that general semantic understanding models cannot output an accurate recognition result for fuzzy text, achieving a personalized semantic understanding effect for different users. Simultaneously, content popularity features are incorporated into the semantic understanding model. When the semantic understanding result contains media asset information (such as media asset content names), and the media asset content name corresponds to multiple fields (e.g., "Three Lives Three Worlds" is both music and video), the content popularity features of the media asset are obtained. These content popularity features are then integrated into the semantic understanding model as a bias term for secondary semantic understanding to obtain the final result. This approach makes the model's recognition more intelligent and meets the needs of most users who seek popularity. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0035] Figure 1 This is one of the flowcharts illustrating the semantic understanding method that integrates user features provided in this embodiment of the invention;

[0036] Figure 2 This is the second flowchart of the semantic understanding method that integrates user features provided in this embodiment of the invention;

[0037] Figure 3 This is a schematic diagram of a semantic understanding model that integrates user features and content popularity features, provided in an embodiment of the present invention.

[0038] Figure 4 This is a schematic diagram of the structure of the semantic understanding device that integrates user features provided in an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram of the structure of a terminal provided according to an embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0042] Semantic understanding is an important branch of NLP (Natural Language Processing). It is currently widely used in various human-computer dialogue systems, such as smart speakers and intelligent question-answering systems.

[0043] User input from various terminals, including text and voice, is converted into standardized text data. Semantic understanding is then performed to analyze the user's domain, intent, and key slots, providing corresponding services. For example, a user might upload a voice command, "I want to listen to Jay Chou's 'Sunny Day'," which is then converted into standardized text using ASR (Automatic Speech Recognition). The semantic understanding model then identifies that a service needs to be provided to the user within the music domain, and that the user's intent is to request a song by the artist "Jay Chou," with the song title slot being "Sunny Day." Finally, the business module plays the corresponding song based on the recognition results. This example illustrates a complete human-computer interaction process and the role of semantic understanding in current human-computer interaction systems.

[0044] Currently, the general implementation methods of semantic understanding in task-oriented dialogue systems can be divided into two types: First, constructing a multi-task joint model that combines domain recognition, intent recognition, and slot recognition to simultaneously identify the domain, intent, and slots contained in the target text. Second, constructing separate recognition models for each of the three tasks (domain recognition, intent recognition, and slot recognition) and performing domain recognition, intent recognition, and slot recognition separately for each target text. Based on the patent search results and the four patents mentioned above, it can be seen that currently published patents generally only broadly integrate user features into the entire voice interaction process to provide personalized services for users, without considering integrating user features into the semantic understanding module to achieve personalized semantic understanding services for different users.

[0045] Further analysis reveals the following shortcomings in the existing technical solutions:

[0046] 1. Lack of integration of user features into the semantic understanding process to achieve personalized semantic understanding capabilities. Current semantic understanding methods generally do not integrate user features into the semantic understanding process to achieve personalized semantic understanding services for different users. The use of user features is generally limited to adding them to the overall voice interaction process, without considering combining user features with atomic capabilities, including semantic understanding capabilities, to improve the effectiveness of each atomic capability. Semantic understanding capabilities often need to output different recognition results for different users when recognizing fuzzy text. General semantic understanding capabilities, lacking user feature information, cannot provide personalized output. For example, if a user inputs the text "Play A Chinese Ghost Story," which is both a song and a movie, and both are well-known resources, using general semantic understanding capabilities to recognize this text, because it only contains text information, it cannot output an accurate result, but only multiple possible semantic understanding results for subsequent modules to choose the best one. If user features are integrated into the semantic understanding model, adding one-dimensional user feature information, then based on the user's preference for listening to music or watching videos, an optimal output can be provided at the semantic understanding level.

[0047] 2. Waste of resources. Currently, the use of user features in voice interaction systems often involves first recognizing the user input and then outputting multiple recognition results. Subsequently, the optimal result is selected from these multiple results based on the user features and returned. As can be seen, multiple intermediate results are recognized before returning the final result, resulting in a certain degree of resource waste.

[0048] 3. Inaccurate user segmentation granularity and user feature extraction not aligned with actual business scenarios. Current user feature extraction methods largely fail to consider the situation in modern home environments where a single device is used by multiple users. Extracting user features solely based on voice or historical user behavior cannot accurately depict different user profiles. Inaccurate user profiles inevitably compromise the quality of service provided to users. For example, in a home with a single speaker, all family members may use it for different purposes. Segmenting users solely by device granularity and performing user behavior analysis and feature extraction fails to provide precise services to different users and may even prevent the accurate extraction of user characteristics. This is because the age range of users in a home setting is vast, making it difficult to identify patterns. Without patterns, accurate user characteristics cannot be derived.

[0049] 4. Lack of semantic understanding based on content popularity. For some ambiguous text, semantic understanding based solely on user characteristics still fails to meet the actual needs of users. For example, a user's characteristics include a preference for music and a liking for trending content. Given the input text "Play Three Lives Three Worlds," "Three Lives Three Worlds" is a common expression among users of the currently popular TV series "Three Lives Three Worlds, Ten Miles of Peach Blossoms," and it is also a song title. If only user characteristics are used to assist semantic understanding, the result would be to play the song. However, considering the information from the entire voice interaction system, most users currently want the TV series to play when they request "Three Lives Three Worlds." Therefore, the above semantic understanding result of playing the song is highly likely to be a misidentification.

[0050] To address the aforementioned shortcomings, this invention proposes a semantic understanding method, apparatus, and electronic device that integrates user features. By incorporating user features into the semantic understanding process, personalized semantic understanding for different users is achieved. User features are primarily determined based on users' historical behavioral habits, while also considering the situation of multiple users on a single device. Voiceprint information is used to distinguish different users on the same device, refining the granularity of user feature extraction and improving the accuracy of user profiling. Furthermore, this invention adds content popularity features to user features to assist semantic understanding, thus avoiding information silos to some extent while satisfying users' historical behavioral habits. Because user features are integrated into the semantic understanding process, only one final semantic understanding result is output after the semantic understanding is completed. This avoids the resource waste problem inherent in typical voice interaction systems that use user features, which first identify multiple semantic understanding results, perform subsequent operations, and then use user features to select the best result from among the multiple results before outputting a single result. A detailed explanation follows.

[0051] Figure 1 This is one of the flowcharts illustrating a semantic understanding method that integrates user features, provided in an embodiment of the present invention. (Refer to...) Figure 1 This invention provides a semantic understanding method that integrates user features, which may include:

[0052] S110: Obtain the text to be processed, input the text to be processed into the semantic understanding model, perform the first semantic understanding on the text to be processed based on the semantic understanding model, and obtain the initial semantic understanding result.

[0053] Furthermore, such as Figure 2 The diagram shown is a second flowchart illustrating the semantic understanding method that integrates user features provided in this embodiment of the invention. First, the user's speech is converted into user text input into the semantic understanding model. The model then performs a first semantic processing step to obtain an initial semantic understanding result. This first semantic processing step does not include user features or content popularity features.

[0054] S120: Determine whether a second semantic understanding is needed based on the initial semantic understanding results.

[0055] In one embodiment, determining whether a second semantic understanding is needed based on the initial semantic understanding result specifically includes: determining whether the initial semantic understanding result is a preset intent and includes media asset information; if so, then searching in the media asset knowledge base based on the media asset information in the initial semantic understanding result to obtain the search result; if the search result indicates that the slot identification result in the initial semantic understanding is fuzzy media asset, then it is determined that a second semantic understanding is needed.

[0056] Furthermore, if the initial semantic understanding result indicates a specific intent and the media asset knowledge base retrieval confirms the inclusion of a fuzzy media asset content name (a fuzzy media asset content name refers to a content name that can refer to multiple contents, such as "A Chinese Ghost Story" being both a movie and a song, in which case "A Chinese Ghost Story" is a fuzzy media asset content name), then user features (determined by combining voiceprint information and user historical behavior) and content popularity features are obtained for a second semantic understanding, and the final semantic understanding result is output.

[0057] In one embodiment, determining whether a second semantic understanding is needed based on the initial semantic understanding result specifically includes: determining whether the initial semantic understanding result is a preset intent and includes media asset information; if it is determined that the initial semantic understanding result is a preset intent and includes media asset information, then a search is performed in the media asset knowledge base based on the media asset information in the initial semantic understanding result to obtain the search result; if the search result indicates that the slot identification result in the initial semantic understanding is not a fuzzy media asset, then the identification ends and the identification result is output; if it is determined that the initial semantic understanding result is not a preset intent or does not include media asset information, then the identification ends and the identification result is output.

[0058] Furthermore, if the first semantic recognition result does not contain a specific intent or a media asset content name, it means that the user's semantics can be understood accurately, and there is no need to perform a second semantic understanding. The recognition ends directly, and the recognition result is output. If the first semantic recognition result is a specific intent and contains a media asset content name, the media asset knowledge base is searched to determine whether the media asset content name is a fuzzy media asset. If it is not a fuzzy media asset, it means that the user's intent can also be understood accurately, and there is no need to perform a second semantic understanding. The recognition ends directly, and the recognition result is output.

[0059] S130: If it is determined that a second semantic understanding is required, then obtain media asset information based on the initial semantic understanding results, and obtain the content popularity characteristics of the media assets and the user characteristics of the target users based on the media asset information.

[0060] In one embodiment, obtaining the content popularity features of media assets and the user features of target users based on media asset information specifically includes: identifying the speech corresponding to the text to be processed based on voiceprint recognition to determine the target user; querying the content popularity features of slot recognition results in the initial semantic understanding in different fields and the user features of target users.

[0061] S140: Input the content popularity features and user features into the semantic understanding model for processing, and combine them with the initial semantic understanding results to obtain the final semantic understanding result.

[0062] In one embodiment, content popularity features and user features are input into a semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding result. Specifically, this includes: inputting content popularity features and user features into a semantic understanding model; superimposing the content popularity features and user features based on the semantic understanding model to obtain a superimposed bias term; and obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result.

[0063] In one embodiment, the decoding layer of each subtask includes a domain decoding layer, an intent decoding layer, and a slot decoding layer. The final semantic understanding result is obtained based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result. Specifically, this includes: inputting the bias term into the semantic understanding model; superimposing the bias term and the domain decoding layer from the first semantic understanding to obtain a new domain decoding layer result; obtaining a new domain recognition result based on the new domain decoding layer result; generating an intent domain constraint vector and a slot domain constraint vector corresponding to the new domain recognition result; applying the intent domain constraint vector to the intent decoding layer and the slot domain constraint vector to the slot decoding layer to obtain the final semantic understanding result.

[0064] The present invention will be further described below with reference to specific embodiments:

[0065] 1) Integration and use of user features and content popularity features in semantic understanding models

[0066] To address the issue that current mainstream semantic understanding models fail to provide personalized recognition based on different user behavioral preferences and thus lack more intelligent semantic understanding services, this paper proposes a method to integrate user features and content popularity features into the semantic understanding model. First, user features and content popularity features are vectorized, and then superimposed as bias terms during the decoding process of the semantic understanding model. This improves the accuracy of the semantic understanding model in recognizing fuzzy text, providing users with a personalized semantic understanding service. The user feature bias term and content popularity bias term are only applied during the second semantic understanding recognition stage. The diagram below illustrates their use in the model decoding process. Figure 3As shown, after the text is segmented into words, each subtask is decoded, and the results of each subtask decoding layer are retained, and the vectors of each subtask decoding layer are output.

[0067] Before introducing the overall process, we will briefly introduce the concept of domain constraint vectors, as this semantic understanding model involves their use. Domain constraint vectors are used to define the domains of intent and slot information. There are two types of domain constraint vectors: one for intents, with a length equal to the number of intent categories N; and the other for slots, with a length equal to the number of slot categories M. Domain constraint vectors contain only 0 and 1 values. The value of a domain constraint vector is determined by the arrangement of intents and slots. The following explanation uses a one-dimensional array to illustrate the calculation of the intent domain constraint vector. Assume there are two domains (A, B) and five intents (a1, b1, a2, a3, b2), where intents a1, a2, and a3 belong to domain A, and b1 and b2 belong to domain B. Furthermore, the output order of intents at the intent decoding layer is a1, b1, a2, a3, b2. Based on the above information, the domain constraint vector is calculated. When the domain is A, the domain constraint vector for the intent is [1,0,1,1,0], and when the domain is B, the domain constraint vector for the intent is [0,1,0,0,1]. The calculation method for the domain constraint vector for slots is the same as that for the intent. Simply put, the domain constraint vector is calculated by setting the probability of intents not belonging to the domain to zero in the intent recognition task and setting the probability of slots not belonging to the domain to zero in the slot recognition task.

[0068] The following is combined Figure 2 and below Figure 3 The document provides detailed examples to illustrate the entire identification process and the roles of user characteristics and content popularity characteristics in the identification process.

[0069] It is assumed that there are currently a total of 5 domains, 6 intents, and 7 slots, wherein the 5 domains are smart home, video, music, weather, and news respectively. First, user A inputs the text "Play *Feng Qiu Huang*", which is transmitted to a semantic understanding model for recognition. The first semantic understanding result is output, and the results of decoding layers of each subtask are retained for secondary semantic understanding: wherein the result of the domain decoding layer is [0, 0.5, 0.45, 0, 0.05], and the corresponding domain recognition result is video; the result of the intent decoding layer is [0, 0.6, 0.3, 0, 0.1, 0], after applying the domain restriction vector, the corresponding intent recognition result obtained is video on demand; the results of the slot decoding layer are [0.99, 0, 0, 0, 0, 0, 0.01], [0, 0, 0, 0.98, 0.02, 0, 0], [0, 0, 0, 0.98, 0.02, 0, 0], [0, 0, 0, 0.98, 0.02, 0, 0], after applying the domain restriction vector, the corresponding slot recognition result obtained is the video name "*Feng Qiu Huang*".

[0070] Checking the preset rules, it is required to perform media asset knowledge retrieval for the recognition result where the intent is video on demand and the slot recognition result includes media asset information. After retrieval, it is obtained that "*Feng Qiu Huang*" is both a song name and a video name, which belongs to an ambiguous media asset content name. Then the heat information corresponding to the media asset content name is queried, and it is obtained that the content heat bias term of "*Feng Qiu Huang*" is [0, 0.05, 0.2, 0, 0], the specific meaning of which is that "*Feng Qiu Huang*" has a certain degree of heat in both the video domain and the music domain, and the heat in the music domain is significantly higher at 0.2.

[0071] Subsequently, the user characteristic bias term corresponding to user A is continuously queried, and the obtained user characteristic bias term result is [0, 0.15, 0.1, 0, 0], that is, the user has a certain preference for video on demand. Then the user characteristic bias term and the content heat bias term are superimposed to obtain the superimposed bias term [0, 0.2, 0.3, 0, 0], and secondary recognition is performed by the semantic understanding model. Since the decoding layer results of each subtask in the first semantic understanding are retained, it is only necessary to superimpose the bias term with the domain decoding layer of the first semantic understanding to obtain a brand-new domain decoding layer [0, 0.7, 0.75, 0, 0.05], and a new domain recognition result is obtained as music according to the new domain decoding layer result. After obtaining the new domain information, the intent domain restriction vector [0, 0, 1, 0, 0, 0] and the slot domain restriction vector [0, 0, 0, 0, 1, 0, 1] corresponding to the music domain are generated, then they are respectively applied to the results of the intent decoding layer and the slot decoding layer, and finally a complete secondary recognition result is obtained. It can be seen that the use of the domain restriction vector makes the second semantic understanding recognition basically have no computational consumption and will not affect the performance.

[0072] The user feature bias and content popularity bias are determined based on user historical behavior and content popularity, respectively. The dimension size is consistent with the number of domains. For example, if semantic understanding currently only supports 5 domains, then the dimension size of both the user feature bias and content popularity bias is 5. The size of the user feature bias is mainly determined by the frequency of user usage in different domains. Additionally, domains such as home furnishings do not contain ambiguous media assets, so the corresponding value for that domain is directly set to 0. The size of the content popularity bias is mainly determined by the number of times the content is viewed, and also by whether the user prefers to pursue popularity, with a certain penalty applied to this bias. To prevent the results of the two biases from excessively influencing the model results, the size range of each dimension of the bias is limited to [0, 0.2]. This range is mainly established based on statistical analysis; currently, for more than 80% of ambiguous texts, the probability difference between the top two in the first semantic understanding domain decoding layer is between [0, 0.2].

[0073] 2) Refined definition of user characteristics

[0074] Current user feature extraction methods typically only segment users down to the device level. However, considering the common scenario of multiple users per device in a home environment, with varying ages and usage habits, further refinement of user segmentation is necessary. This invention primarily uses voiceprint classification based on voice information uploaded by devices to differentiate and record the historical behaviors of different users on the same device. User features are then extracted based on these refined historical behaviors. These refined user features better reflect real-world usage scenarios, avoiding the problems of inconsistent user interaction habits and unclear features that arise when feature extraction is performed on a single device and user.

[0075] Based on the above embodiments, this invention endows the semantic understanding model with personalized recognition capabilities by incorporating user features and content popularity features into the semantic understanding model. The key points of this invention are as follows:

[0076] 1) User features are incorporated into the semantic understanding model as bias terms to aid in the semantic understanding of user interaction text. These features are extracted from users' historical interaction data and voice information, primarily reflecting user behavioral preferences, including domain usage preferences (whether they prefer listening to music or watching videos) and their liking for trending content. Domain usage preferences are used to perform biased recognition when performing semantic understanding on fuzzy text, and the liking for trending content helps calibrate and balance the user's choices when accessing trending content with their domain usage preferences. The use of user features in the semantic understanding model makes the model's recognition more intelligent and personalized.

[0077] 2) In the extraction and confirmation of user features, unlike the general approach of segmenting users solely based on the device dimension, this invention takes into account the current home scenario where one device has multiple users, and adds user voice features to further refine user differentiation. Regarding the use of voice features, considering the current poor performance of voiceprint recognition, the user voice feature extraction mentioned in this invention only targets the voice classification and recognition of multiple users on one device, to distinguish different users on the same device, ultimately determining the interaction habits of different users on the same device, and thus refining the user differentiation level.

[0078] 3) Content popularity features are incorporated into the semantic understanding model. When the model's first semantic recognition result contains media asset content names that correspond to multiple fields, the popularity information of the content will be combined, and the content popularity features will be incorporated into the model as a bias term for secondary recognition. This method makes the model's recognition more intelligent and meets the needs of most users who pursue popularity.

[0079] Based on the above embodiments and key points, the advantages of the present invention over the prior art are as follows:

[0080] Existing technologies lack a method to directly integrate user features into a semantic understanding model, thereby endowing it with personalized recognition capabilities. Instead, user features are broadly incorporated into the entire voice interaction process to assist in voice interaction. The steps involve first performing speech recognition, then semantic understanding. For ambiguous text, multiple possible alternative results are typically output for subsequent business processing to obtain a final result. Finally, user features are combined to select the optimal result from these multiple business processing results and output the final output. This approach of outputting multiple semantic understanding results and then using user features to determine the optimal response is resource-intensive due to the multiple alternative results involved in business processing. To address this issue, this invention directly integrates user features into the semantic understanding model. This ensures that for ambiguous text, semantic understanding outputs only a single recognition result that aligns with the user's usage habits, saving intermediate computational resources while guaranteeing a superior user experience.

[0081] Existing technical solutions do not integrate content popularity features into the semantic understanding process or the entire voice interaction process. While relying solely on user features for voice interaction takes into account user selection habits, they fail to consider the simultaneous demand from most users for trending content. This invention integrates content popularity features into the semantic understanding model. When the model's initial recognition result contains media asset names corresponding to multiple media assets, it incorporates the content popularity information as a bias term into the model for secondary recognition and outputs the final result. The integration of content popularity features into the semantic understanding model allows the entire system to consider users' actual needs from more dimensions, balancing the issue of limited selection that arises when relying solely on user features for auxiliary recognition.

[0082] Existing technical solutions only segment users based on devices and then extract user features. This approach fails to consider the scenario where a single device is used by multiple users in a real-world home environment. Furthermore, home users vary widely in age and usage habits, and the user features extracted from the device perspective are too coarse-grained to reflect the true usage habits of different users. This invention classifies voice information uploaded by devices using voiceprints to distinguish different users on the same device. After segmenting different users on the same device, it records user behavior and extracts user features separately. This method extracts more accurate user features, enabling subsequent capabilities that incorporate these features to provide more precise services.

[0083] The following describes a semantic understanding device that integrates user features according to an embodiment of the present invention. The semantic understanding device that integrates user features described below and the semantic understanding method that integrates user features described above can be referred to and correspond to each other.

[0084] Figure 4 A schematic diagram of a semantic understanding system that integrates user features according to an embodiment of the present invention, comprising:

[0085] The first semantic understanding module 410 is used to acquire the text to be processed, input the text to be processed into the semantic understanding model, and perform the first semantic understanding on the text to be processed based on the semantic understanding model to obtain the initial semantic understanding result; the judgment module 420 is used to determine whether a second semantic understanding is needed based on the initial semantic understanding result; the second semantic understanding module 430 is used to acquire media asset information based on the initial semantic understanding result if it is determined that a second semantic understanding is needed, acquire the content popularity features of the media asset and the user features of the target user based on the media asset information, input the content popularity features and user features into the semantic understanding model for processing, and combine them with the initial semantic understanding result to obtain the final semantic understanding result.

[0086] In one embodiment, the second semantic understanding module includes a semantic understanding unit, which is used to input content popularity features and user features into the semantic understanding model, and superimpose the content popularity features and user features based on the semantic understanding model to obtain the superimposed bias term; and obtain the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result.

[0087] In one embodiment, the judgment module 420 is specifically used to determine whether the initial semantic understanding result is a preset intention and includes media asset information. If so, it performs a search in the media asset knowledge base based on the media asset information in the initial semantic understanding result to obtain the search result. If the search result indicates that the slot identification result in the initial semantic understanding is a fuzzy media asset, it determines that a second semantic understanding is required.

[0088] In one embodiment, the second semantic understanding module 430 includes a feature acquisition unit; the feature acquisition unit is used to identify the speech corresponding to the text to be processed based on voiceprint recognition, determine the target user, and query the content popularity features of the slot recognition results in the initial semantic understanding in different fields and the user features of the target user.

[0089] In one embodiment, the decoding layer of each subtask includes a domain decoding layer, an intent decoding layer, and a slot decoding layer; the semantic understanding unit includes a decoding subunit; the decoding subunit is used to input the bias term into the semantic understanding model, superimpose the bias term and the domain decoding layer during the first semantic understanding to obtain a new domain decoding layer result, obtain a new domain recognition result based on the new domain decoding layer result, generate the corresponding domain intent domain constraint vector and slot domain constraint vector based on the new domain recognition result, apply the intent domain constraint vector to the intent decoding layer, apply the slot domain constraint vector to the slot decoding layer, and obtain the final semantic understanding result.

[0090] In one embodiment, the judgment module 420 is specifically used to determine whether the initial semantic understanding result is a preset intention and includes media asset information. If it is determined that the initial semantic understanding result is a preset intention and includes media asset information, then based on the media asset information in the initial semantic understanding result, a search is performed in the media asset knowledge base to obtain the search result. If the search result indicates that the slot identification result in the initial semantic understanding is not a fuzzy media asset, then the recognition ends and the recognition result is output. If it is determined that the initial semantic understanding result is not a preset intention or does not include media asset information, then the recognition ends and the recognition result is output.

[0091] The terminal involved in the embodiments of the present invention may be a device that provides voice and / or data connectivity to a user, a handheld device with wireless connectivity, or other processing devices connected to a wireless modem. The name of the terminal device may differ in different systems; for example, in a 5G system, the terminal device may be called a User Equipment (UE).

[0092] Figure 5 This is a schematic diagram of the structure of a terminal according to an embodiment of the present invention, with reference to... Figure 5 The present invention also provides a terminal, which may include: a memory 510, a transceiver 520 and a processor 530;

[0093] The memory 510 is used to store computer programs; the transceiver 520 is used to send and receive data under the control of the processor 530; the processor 530 is used to read the computer program in the memory 510 and perform the following operations:

[0094] The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result.

[0095] Determine whether a second semantic understanding is needed based on the initial semantic understanding results;

[0096] If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding results, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information.

[0097] The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding results.

[0098] Among them, Figure 5 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 530) and memory (memory 510). The bus architecture can also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 520 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. For different user equipment, the user interface 540 can also be an interface capable of connecting external or internal devices as needed.

[0099] The processor 530 is responsible for managing the bus architecture and general processing, while the memory 510 can store the data used by the processor 530 when performing operations.

[0100] The processor 530 executes any of the methods provided in the embodiments of the present invention according to the obtained executable instructions by calling a computer program stored in the memory 510. The processor and the memory may also be physically separated.

[0101] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call a computer program in the memory 630 to execute steps of a semantic understanding method that fuses user features, such as:

[0102] The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result.

[0103] Determine whether a second semantic understanding is needed based on the initial semantic understanding results;

[0104] If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding results, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information.

[0105] The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding results.

[0106] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0107] On the other hand, embodiments of the present invention also provide a processor-readable storage medium storing a computer program for causing a processor to perform the steps of the methods provided in the above embodiments, such as including:

[0108] The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result.

[0109] Determine whether a second semantic understanding is needed based on the initial semantic understanding results;

[0110] If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding results, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information.

[0111] The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding results.

[0112] Processor-readable storage media can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semantic understanding method that integrates user features, characterized in that, include: The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result. Determine whether a second semantic understanding is needed based on the initial semantic understanding results; If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding result, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information; The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding result. The step of inputting the content popularity features and user features into a semantic understanding model for processing, and combining the initial semantic understanding results to obtain the final semantic understanding result, specifically includes: inputting the content popularity features and user features into a semantic understanding model; superimposing the content popularity features and user features based on the semantic understanding model to obtain a superimposed bias term; and obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result. The decoding layers for each subtask include a domain decoding layer, an intent decoding layer, and a slot decoding layer. The process of obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result specifically includes: inputting the bias term into the semantic understanding model; superimposing the bias term and the domain decoding layer from the first semantic understanding to obtain a new domain decoding layer result; obtaining a new domain recognition result based on the new domain decoding layer result; generating an intent domain constraint vector and a slot domain constraint vector corresponding to the new domain recognition result; applying the intent domain constraint vector to the intent decoding layer and the slot domain constraint vector to the slot decoding layer to obtain the final semantic understanding result.

2. The semantic understanding method integrating user features according to claim 1, characterized in that, The step of determining whether a second semantic understanding is needed based on the initial semantic understanding result specifically includes: Determine whether the initial semantic understanding result is a pre-set intent and includes media asset information. If so, search the media asset knowledge base based on the media asset information in the initial semantic understanding result to obtain the search results. If the retrieval result indicates that the slot identification result in the initial semantic understanding is fuzzy media, then it is determined that a second semantic understanding is required.

3. The semantic understanding method integrating user features according to claim 1, characterized in that, The step of obtaining the content popularity characteristics of media assets and the user characteristics of target users based on the media asset information specifically includes: Based on voiceprint recognition, the speech corresponding to the text to be processed is identified to determine the target user; The query retrieves the slot identification results from the initial semantic understanding, including the content popularity features in different fields and the user characteristics of the target user.

4. The semantic understanding method for fusing user features according to claim 1, characterized in that, The step of determining whether a second semantic understanding is needed based on the initial semantic understanding result specifically includes: Determine whether the initial semantic understanding result is a preset intent and includes media asset information. If it is determined that the initial semantic understanding result is a preset intent and includes media asset information, then search the media asset knowledge base based on the media asset information in the initial semantic understanding result to obtain the search results. If the retrieval result represents the slot identification result in the initial semantic understanding and is not a fuzzy medium, then the identification ends and the identification result is output. If it is determined that the initial semantic understanding result is not the preset intent or does not include media asset information, the recognition ends and the recognition result is output.

5. A semantic understanding device that integrates user features, characterized in that, include: The first semantic understanding module is used to acquire the text to be processed. The text to be processed is input into the semantic understanding model, and the semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result. The judgment module is used to determine whether a second semantic understanding is needed based on the initial semantic understanding results. The second semantic understanding module is used to obtain media asset information based on the initial semantic understanding result if it is determined that a second semantic understanding is needed; obtain the content popularity features of the media asset and the user features of the target user based on the media asset information; input the content popularity features and the user features into the semantic understanding model for processing; and combine the initial semantic understanding result to obtain the final semantic understanding result. The step of inputting the content popularity features and user features into a semantic understanding model for processing, and combining the initial semantic understanding results to obtain the final semantic understanding result, specifically includes: inputting the content popularity features and user features into a semantic understanding model; superimposing the content popularity features and user features based on the semantic understanding model to obtain a superimposed bias term; and obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result. The decoding layers for each subtask include a domain decoding layer, an intent decoding layer, and a slot decoding layer. The process of obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result specifically includes: inputting the bias term into the semantic understanding model; superimposing the bias term and the domain decoding layer from the first semantic understanding to obtain a new domain decoding layer result; obtaining a new domain recognition result based on the new domain decoding layer result; generating an intent domain constraint vector and a slot domain constraint vector corresponding to the new domain recognition result; applying the intent domain constraint vector to the intent decoding layer and the slot domain constraint vector to the slot decoding layer to obtain the final semantic understanding result.

6. A terminal, characterized in that, Includes memory, transceiver, and processor; A memory for storing computer programs; a transceiver for sending and receiving data under the control of the processor; and a processor for reading the computer programs from the memory and performing the following operations: The text to be processed is obtained, and the text to be processed is input into the semantic understanding model. The semantic understanding model performs the first semantic understanding of the text to be processed to obtain the initial semantic understanding result. Determine whether a second semantic understanding is needed based on the initial semantic understanding results; If it is determined that a second semantic understanding is required, then media asset information is obtained based on the initial semantic understanding result, and the content popularity characteristics of the media asset and the user characteristics of the target user are obtained based on the media asset information; The content popularity features and user features are input into the semantic understanding model for processing, and the final semantic understanding result is obtained by combining the initial semantic understanding result. The step of inputting the content popularity features and user features into a semantic understanding model for processing, and combining the initial semantic understanding results to obtain the final semantic understanding result, specifically includes: inputting the content popularity features and user features into a semantic understanding model; superimposing the content popularity features and user features based on the semantic understanding model to obtain a superimposed bias term; and obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result. The decoding layers for each subtask include a domain decoding layer, an intent decoding layer, and a slot decoding layer. The process of obtaining the final semantic understanding result based on the bias term and the results of the encoding layers of each subtask in the initial semantic understanding result specifically includes: inputting the bias term into the semantic understanding model; superimposing the bias term and the domain decoding layer from the first semantic understanding to obtain a new domain decoding layer result; obtaining a new domain recognition result based on the new domain decoding layer result; generating an intent domain constraint vector and a slot domain constraint vector corresponding to the new domain recognition result; applying the intent domain constraint vector to the intent decoding layer and the slot domain constraint vector to the slot decoding layer to obtain the final semantic understanding result.

7. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the semantic understanding method for fusing user features as described in any one of claims 1 to 4.

8. A processor-readable storage medium storing a computer program, characterized in that, A computer program is used to cause a processor to perform the steps of the semantic understanding method that integrates user features as described in any one of claims 1 to 4.

Images