Intelligent AI-driven multi-round multi-intention recognition enhancement method and device

Abstract

The embodiment of the invention provides an intelligent AI-driven multi-round multi-intention recognition enhancement method, device and equipment and a computer readable storage medium. The method comprises the following steps: acquiring final query information of a user, and constructing a query combination set; performing prediction based on the query group to obtain a candidate intention set; retrieving a corresponding example set based on the candidate intention set; encoding the example set to obtain a vector of each example in the example set; splicing and coding all the obtained query information to obtain a target vector; respectively calculating cosine similarity scores of the target vector and the vector of each example in the example set, and constructing a target example set; based on the target instance set, an intent prediction for the target query is generated. In this way, the problems of corpus scarcity and high labeling difficulty in multi-round dialogue research are solved, and the training efficiency and performance of the model are improved.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of data processing, and in particular, to intelligent AI-driven multi-round multi-intent recognition enhancement methods, devices, equipment, and computer-readable storage media. Background Art

[0002] Currently, research on multi-round dialogues generally tends to conduct experimental verification on existing public datasets. This tendency is actually based on an implicit premise, that is, researchers assume that they already have sufficient and diverse multi-round dialogue corpus resources.

[0003] In fact, high-quality, authentic multi-round dialogue corpora are quite scarce. At the same time, the difficulty and complexity of effectively annotating these scarce corpora will consume a lot of manpower and material costs.

[0004] Therefore, how to adopt unsupervised / weakly supervised learning methods and introduce innovative means such as adaptive learning mechanisms and scene perception models to effectively deal with the problem of scarce multi-round dialogue materials and difficulty in annotation in reality is an issue that needs to be urgently addressed. Summary of the invention

[0005] According to the embodiments of the present application, an intelligent AI-driven multi-round multi-intent recognition enhancement solution is provided, which adopts unsupervised / weakly supervised learning methods and introduces innovative means such as adaptive learning mechanisms and scene perception models. It can effectively deal with the problems of scarce multi-round dialogue materials and difficult labeling in reality, and greatly improve user experience.

[0006] In a first aspect of the present application, an intelligent AI-driven multi-round multi-intention recognition enhancement method is provided. The method comprises: Get the user's final query information; Constructing a query combination set based on the current conversation context and the final query information; Using the trained single-round classification model, predict each combination in the query combination set to obtain a candidate intent set; Based on each candidate intent in the candidate intent set, a corresponding example set is retrieved from a single-round data set used to construct the single-round classification model; the example set is encoded to obtain a vector of each example in the example set; all the query information obtained is concatenated and encoded to obtain a target vector; Calculating the cosine similarity scores of the target vector and the vectors of each example in the example set respectively; summarizing the examples with cosine similarity scores greater than a threshold to obtain a target example set; Based on the target example set, a composite function is constructed; based on the composite function, an intent prediction for the target query is generated.

[0007] Furthermore, the trained single-round classification model is used to predict each combination in the query combination set, and the candidate intent set obtained includes: ; ; in, To query a single combination in a combination set; is the probability distribution of each layer; is the number of model layers.

[0008] Further, the respectively calculating the cosine similarity score of the target vector and the vector of each example in the example set includes: ; in, is the target vector; is the vector for each example in the example set.

[0009] Furthermore, it also includes: The candidate intent set includes a confidence scoring mechanism: ; in, For the At the level, the single-round classification model is used for intent probability; For a single candidate intent.

[0010] Furthermore, it also includes: The single-pass classification model is trained with the following intent classification loss: ; in, is the real label; is the probability predicted by the model.

[0011] Furthermore, it also includes: The conditional probability of the next round of the single-round classification model is calculated as follows: ; in, For the current query; For the next round of queries; is the vector representation of the current query; The weight matrix predicted for the next round of queries; Bias vector predicted for the next query round.

[0012] Furthermore, it also includes: Each possible next-round query category can be calculated as follows: ; Among them, K is the number of all possible query categories in the next round.

[0013] In a second aspect of the present application, an intelligent AI-driven multi-round multi-intention recognition enhancement device is provided. The device comprises: The acquisition module is used to obtain the user's final query information; A construction module, used to construct a query combination set based on the current conversation context and the final query information; A first prediction module is used to predict each combination in the query combination set by using a trained single-round classification model to obtain a candidate intent set; An encoding module, configured to retrieve a corresponding example set from a single-round data set used to construct the single-round classification model based on each candidate intent in the candidate intent set; encode the example set to obtain a vector of each example in the example set; concatenate and encode all the query information obtained to obtain a target vector; A calculation module, used to calculate the cosine similarity score of the target vector and the vector of each example in the example set respectively; examples with cosine similarity scores greater than a threshold are summarized to obtain a target example set; The second prediction module is used to construct a composite function based on the target example set; based on the composite function, generate an intent prediction for the target query.

[0014] In a third aspect of the present application, an electronic device is provided, which includes a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the program, the method described above is implemented.

[0015] In a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method according to the first aspect of the present application is implemented.

[0016] The intelligent AI-driven multi-round multi-intention recognition enhancement method provided in the embodiment of the present application obtains the user's final query information; constructs a query combination set based on the current conversation context and the final query information; predicts each combination in the query combination set separately through the trained single-round classification model to obtain a candidate intent set; based on each candidate intent in the candidate intent set, retrieves the corresponding example set from the single-round data set used to construct the single-round classification model; encodes the example set to obtain a vector of each example in the example set; concatenates and encodes all the acquired query information to obtain a target vector; calculates the cosine similarity scores of the target vector and the vector of each example in the example set respectively; summarizes the examples with cosine similarity scores greater than a threshold to obtain a target example set; constructs a composite function based on the target example set; and generates an intent prediction for the target query based on the conforming function, which greatly improves the user experience and can provide strong support for current multi-round conversation research.

[0017] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present application, nor are they intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other features, advantages and aspects of the embodiments of the present application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein: Figure 1 is a flowchart of an intelligent AI-driven multi-round multi-intention recognition enhancement method according to an embodiment of the present application; Figure 2 is a block diagram of an intelligent AI-driven multi-round multi-intention recognition enhancement device according to an embodiment of the present application; Figure 3 It is a schematic diagram of the structure of a terminal device or a server suitable for implementing the embodiments of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0020] In addition, the term "and / or" in this article is only a description of the association relationship between the associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0021] Figure 1 A flow chart of a multi-round multi-intention recognition enhancement method driven by intelligent AI according to an embodiment of the present disclosure is shown. The method may include: S110, obtaining the user's final query information.

[0022] In this embodiment, the execution entity (such as a cloud server, etc.) of the intelligent AI-driven multi-round multi-intention recognition enhancement method can obtain the final query information through a wired or wireless connection.

[0023] Furthermore, the above-mentioned execution entity can obtain the final query information sent by the electronic device (such as a terminal device such as a mobile phone, iPad and / or laptop computer) that is connected to it for communication.

[0024] S120: Construct a query combination set based on the current dialog context and the final query information.

[0025] In some embodiments, based on the conversation context C and the last user query (final query information) , combine the last query with each historical query to build a query combination set: ; in, Splice for text; For each combined query .

[0026] S130, using the trained single-round classification model, predict each combination in the query combination set separately to obtain a candidate intent set.

[0027] In some embodiments, based on the constructed single-round dataset D, a single-round hierarchical text classification model is trained. Among them, the single-round dataset D can be a dataset composed of a large number of independent user queries and corresponding intent labels. Each data sample is a separate user query and does not contain conversation context information. It is mainly used in the training of the single-round intent classification model to provide supervision signals for the model so that it can learn to identify user intent from a single query. Specifically, the single-round dataset D includes: User query: a single query or request made by a user without context; for example, "check the weather", "play music", etc. Intent Path : The intent label corresponding to each user query, represented as an intent path: ; in, Corresponding to different levels of the intent classification tree.

[0028] Furthermore, the single-round dataset D generally has the following characteristics: Independence: Each query is independent and does not depend on previous conversations or context; Hierarchical intent labels: Intent labels are hierarchical according to a predefined classification tree, which refines the classification of intents and helps the model learn more sophisticated intent recognition; Large amount of data: Since single-round data is easier to obtain and annotate, the data set size is usually large, providing sufficient data support for model training.

[0029] In some embodiments, when training a single round of classification models When performing multi-task learning, we can introduce multi-task learning to take into account both intent classification and next round query prediction.

[0030] Intent Classification Loss: ; in, is the true label (one-hot encoding); is the probability predicted by the model.

[0031] Define the next round The conditional probability is: ; in, For the current query; For the next round of queries; It is a vector representation of the current query, usually generated by an encoder function or model, such as BERT, RoBERTa, recurrent neural network (RNN), convolutional neural network (CNN), etc., which is used to convert the input text Convert to a fixed-length vector; The weight matrix for the next round of query prediction, the parameter matrix of the model when performing the next round of query prediction task, is used to represent the current query Projected into the latent space of the next query; for example, The dimension is , and the number of possible classifications for the next round of queries is K, then The dimension is K× ; The bias vector predicted for the next round of queries corresponds to The bias term is used to help the model perform better linear transformation; The dimension is K, which is consistent with the number of categories in the next round of queries; It is a linear change, used to map the representation of the current query to the score (logits) of the next round of queries. It is calculated by linearly transforming the representation vector and adding a bias term to obtain the unnormalized score for each possible next round of query categories. is the activation function, which is used to convert the output of the linear transformation into a probability distribution; it is calculated as follows. For each possible next-round query category k, the probability is calculated as follows: ; Among them, K is the number of all possible query categories for the next round; is the Kth element of the vector.

[0032] Defining query prediction loss , which is used to measure the difference between the next round of query probability distribution predicted by the model and the true label: ; in, The one-hot encoding of the next round of query labels is a vector.

[0033] Furthermore, the total loss function is: ; in, To balance the hyperparameters of the two tasks, they can be set according to the actual application scenario.

[0034] In some embodiments, a single-round classification model trained in the above manner is used. , get the probability distribution of each layer , and then select the top intent of each layer according to these probabilities to form a set of candidate intents: ; Aggregate candidate intents for all combined queries: ; in, To query a single combination in a combination set; is the probability distribution of each layer; is the number of model layers.

[0035] S140, based on each candidate intent in the candidate intent set, retrieve a corresponding example set from the single-round data set used to build the single-round classification model; encode the example set to obtain a vector for each example in the example set; concatenate and encode all the query information that has been obtained to obtain a target vector.

[0036] In some embodiments, based on the contextual learning capability of the large language model (LLM), relevant examples (i.e., demonstration samples) are retrieved for the current conversation from the single-turn dataset D. All queries in the conversation are concatenated into a whole text. : ; right Encode it and get its representation vector (target vector): ; For each candidate intent , retrieve the corresponding example set from the single-round dataset D .calculate Each example The representation vector of is: ; in, is the vector for each example in the example set.

[0037] S150, respectively calculating the cosine similarity scores of the target vector and the vector of each example in the example set; summarizing the examples with cosine similarity scores greater than a threshold to obtain a target example set.

[0038] In some embodiments, the cosine similarity scores of the target vector and the vector of each example in the example set may be calculated by the following formula: ; in, is the target vector; is the vector for each example in the example set.

[0039] Furthermore, for each candidate intent , select the top K examples based on their similarity scores: ; ; Among them, E is the final example set (target example set).

[0040] S160, constructing a composite function based on the target example set; and generating an intent prediction for the target query based on the composite function.

[0041] In some embodiments, based on the target example set, a composite function P is constructed: ; in, Describe the function for the task; It is the instance collection processing function; It is the conversation history context processing function; It is the target query processing function; Request processing function for output.

[0042] In some embodiments, based on the composite function, an intent prediction for the target query is generated. The constructed P is input into the LLM, and the LLM generates an intent prediction for the target query based on the prompt content. The specific process can be modeled as follows: =LLM(P); in, Target query Predicted intention path.

[0043] Specifically, the complete prompt P is input into the LLM, and an understanding function Understand(P) is defined, which includes multiple sub-steps: ; in, for task understanding; Learn by example mode; Analyze the conversation context; Target query.

[0044] Furthermore, the intention prediction can be expressed as: ; Define the post-processing function PostProcess( ): ; The final intention prediction result is obtained through the above formula.

[0045] Furthermore, it also includes: Confidence scores can be introduced in subsequent intents to distinguish the importance of different intents. , the confidence is calculated based on the probability of the single-round classification model : ; in, For the At the level, the single-round classification model is used for intent probability; For a single candidate intent.

[0046] When constructing prompts or selecting examples, Weight the intents to increase the influence of high-confidence intents.

[0047] Further, it also includes: In the process of sample selection, a dynamic adjustment mechanism is introduced. According to the distribution of similarity scores, the number of selected samples K is adjusted. If the highest similarity score Below a certain threshold τ, the value of K is reduced to select only the most relevant examples and avoid introducing noise.

[0048] Furthermore, it also includes: For multilingual scenarios, we introduce cross-language embedding alignment technology. We use contrastive learning or translation alignment methods to ensure that similar semantics in different languages ​​are close in the embedding space, thereby improving the performance of the model in multilingual environments.

[0049] According to the embodiments of the present disclosure, the following technical effects are achieved: It solves the problems of scarce corpus and difficult annotation in multi-round dialogue research, improves the training efficiency and performance of the model, lowers the threshold for system use, and provides strong support for the widespread application and development of multi-round dialogue technology.

[0050] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0051] The above is an introduction to the method embodiment. The following is a further explanation of the scheme described in this application through an apparatus embodiment.

[0052] Figure 2A block diagram of a multi-round multi-intention recognition enhancement device driven by an intelligent AI according to an embodiment of the present application is shown, Figure 2 Shown include: The acquisition module 210 is used to acquire the final query information of the user; A construction module 220, configured to construct a query combination set based on the current conversation context and the final query information; A first prediction module 230 is used to predict each combination in the query combination set by using a trained single-round classification model to obtain a candidate intent set; The encoding module 240 is used to retrieve a corresponding example set from the single-round data set used to construct the single-round classification model based on each candidate intent in the candidate intent set; encode the example set to obtain a vector of each example in the example set; concatenate and encode all the query information obtained to obtain a target vector; A calculation module 250 is used to calculate the cosine similarity score of the target vector and the vector of each example in the example set respectively; examples with cosine similarity scores greater than a threshold are aggregated to obtain a target example set; The second prediction module 260 is used to construct a composite function based on the target example set; and generate an intent prediction for the target query based on the composite function.

[0053] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0054] Figure 3 A schematic diagram of the structure of a terminal device or server suitable for implementing an embodiment of the present application is shown.

[0055] like Figure 3 As shown, the terminal device or server includes a CPU 301, which can perform various appropriate actions and processes according to the program stored in the ROM 302 or the program loaded from the storage part 308 into the RAM 303. In the RAM 303, various programs and data required for the operation of the terminal device or server are also stored. The CPU 301, the ROM 302 and the RAM 303 are connected to each other via a bus 304. An I / O interface 305 is also connected to the bus 304.

[0056] The following components are connected to the I / O interface 305: an input section 306 including a keyboard, a mouse, etc.; an output section 307 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. A removable medium 311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 310 as needed, so that a computer program read therefrom is installed into the storage section 308 as needed.

[0057] In particular, according to an embodiment of the present application, the above method flow steps can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a machine-readable medium, and the computer program contains a program code for executing the method shown in the flow chart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 309, and / or installed from the removable medium 311. When the computer program is executed by the CPU 301, the above-mentioned functions defined in the system of the present application are executed.

[0058] It should be noted that the computer-readable medium shown in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: 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 the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal 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 computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. 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, RF, etc., or any suitable combination of the above.

[0059] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of the code, and the aforementioned module, program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0060] The units or modules involved in the embodiments described in the present application may be implemented by software or hardware. The units or modules described may also be arranged in a processor. The names of these units or modules do not, in some cases, constitute limitations on the units or modules themselves.

[0061] As another aspect, the present application further provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable storage medium stores one or more programs, and when the above programs are used by one or more processors to execute the method described in the present application.

[0062] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of application involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the aforementioned application concept. For example, the above features are replaced with (but not limited to) technical features with similar functions applied in the present application.

Claims

1. An intelligent AI-driven multi-round multi-intention recognition enhancement method, characterized in that: include: Get the user's final query information; Constructing a query combination set based on the current conversation context and the final query information; Using the trained single-round classification model, predict each combination in the query combination set to obtain a candidate intent set; Based on each candidate intent in the candidate intent set, a corresponding example set is retrieved from a single-round data set used to construct the single-round classification model; the example set is encoded to obtain a vector of each example in the example set; all the query information obtained is concatenated and encoded to obtain a target vector; Calculating the cosine similarity score of the target vector and the vector of each example in the example set respectively; Summarize the examples whose cosine similarity scores are greater than the threshold to obtain the target example set; Based on the target example set, construct a composite function; Based on the composite function, an intent prediction for the target query is generated.

2. The method according to claim 1, characterized in that The trained single-round classification model is used to predict each combination in the query combination set, and the candidate intent set obtained includes: ； ； in, To query a single combination in a combination set; is the probability distribution of each layer; is the number of model layers.

3. The method according to claim 2, characterized in that The calculating the cosine similarity score of the target vector and each example vector in the example set comprises: ； in, is the target vector; is the vector for each example in the example set.

4. The method according to claim 1, characterized in that: Also includes: The candidate intent set includes a confidence scoring mechanism: ； in, For the At the level, the single-round classification model is used for intent probability; For a single candidate intent.

5. The method according to claim 4, characterized in that Also includes: The single-pass classification model is trained with the following intent classification loss: ； in, is the real label; is the probability predicted by the model.

6. The method according to claim 5, characterized in that Also includes: The conditional probability of the next round of the single-round classification model is calculated as follows: ； in, For the current query; For the next round of queries; is the vector representation of the current query; The weight matrix predicted for the next round of queries; The bias vector predicted for the next query round.

7. The method according to claim 6, characterized in that Also includes: Each possible next-round query category can be calculated as follows: ； Among them, K is the number of all possible query categories in the next round.

8. An intelligent AI-driven multi-round multi-intention recognition enhancement device, characterized in that: include: The acquisition module is used to obtain the user's final query information; A construction module, used to construct a query combination set based on the current conversation context and the final query information; A first prediction module is used to predict each combination in the query combination set by using a trained single-round classification model to obtain a candidate intent set; An encoding module, configured to retrieve a corresponding example set from a single-round data set used to construct the single-round classification model based on each candidate intent in the candidate intent set; encode the example set to obtain a vector of each example in the example set; concatenate and encode all the query information obtained to obtain a target vector; A calculation module, used to calculate the cosine similarity score of the target vector and the vector of each example in the example set respectively; Summarize the examples whose cosine similarity scores are greater than the threshold to obtain the target example set; A second prediction module, configured to construct a composite function based on the target example set; Based on the composite function, an intent prediction for the target query is generated.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images