Intention recognition method based on deep learning

Abstract

The invention discloses an intention recognition method based on deep learning. The intention recognition method comprises the following steps that firstly, text, voice, images and other different modal data are subjected to fusion processing and output; then, constructing a context sensing module based on a Transform model, receiving the fused data, and fusing context information of a dialogue before and after; then determining a field corresponding to intention recognition, and automatically adjusting the weight of a field specific vocabulary in an input text feature vector; and finally, through sharing a part of network layers, intention identification and slot filling tasks are synchronously carried out, so that the identification model precision is improved, and model overfitting is reduced. According to the method, through multi-modal data fusion, context sensing and targeted weight adjustment and task cooperation, the intention recognition precision is effectively improved, the over-fitting risk is reduced, a more accurate and reliable technical scheme is provided for intention understanding of an intelligent interaction system, and improvement of the man-machine interaction quality is facilitated.

Description

Technical Field

[0001] This invention relates to an intent recognition method, and more specifically to an intent recognition method based on deep learning. Background Technology

[0002] With the rapid development of Natural Language Processing (NLP) technology, intent recognition, as an important component of semantic understanding, has been widely applied in fields such as virtual assistants, customer service systems, and smart homes. However, traditional intent recognition algorithms mostly rely solely on text input for analysis, lacking the integration and reference of other modal information. This often proves inadequate when dealing with ambiguities, implicit semantics, or complex contexts within text. For example, when text is concise, contains polysemous words, or requires prior information to clarify the true intent, surface analysis of the current text alone is insufficient to accurately capture the user's true needs, leading to inaccuracies in practical applications. Furthermore, this single-text input-only analysis method struggles to effectively connect logical relationships and semantic coherence between contexts, failing to dynamically and coherently understand user intent by incorporating dialogue history or scene background, thus limiting its ability to grasp contextual information. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a multimodal input fusion method that jointly processes data from different modalities such as text, speech, and images. By fusing deep learning models of multimodal data, it is possible to more accurately understand user intent, especially in intent recognition involving complex tasks and contexts.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intent recognition method based on deep learning, comprising the following steps: Step 1: Fuse data from different modalities, such as text, speech, and images, and output the fused data. Step 2: Build a context-aware module based on the Transformer model, which receives the fused data output from Step 1 and then adds context information from the preceding and following dialogues. Step 3: Determine the domain corresponding to the intent to be identified, and automatically adjust the weights of domain-specific words in the feature vector of the input text. Step four involves simultaneously performing intent recognition and slot filling tasks by sharing some network layers to improve the accuracy of the recognition model and reduce overfitting.

[0005] As a further improvement of the present invention, the specific method for fusing multimodal inputs in step one is as follows: Assuming the input data includes text and images, the text and images are extracted using convolutional neural networks and recurrent neural networks, respectively, and the final outputs are text features and image features. Then, a weighted average is used to achieve the fusion of multimodal inputs.

[0006] As a further improvement of the present invention, in step two, the context awareness module is modeled based on a self-attention mechanism. Given a series of input word vectors, each word vector is calculated to obtain a weighted output through the attention mechanism, thereby completing the addition of context information.

[0007] As a further improvement of the present invention, the specific method for automatically adjusting the weights for domain-specific words in step three is as follows: given a domain vocabulary set D and the feature vector of the input text, calculate the weight of each word, and adjust the feature vector based on these weights.

[0008] As a further improvement of the present invention, the specific method of performing intent recognition and slot filling tasks simultaneously by sharing part of the network layer in step four is as follows: set the loss functions of the two tasks of intent recognition and slot filling as and , respectively, then calculate the overall loss function of the multi-task learning framework, and optimize the model according to the overall loss function.

[0009] The beneficial effects of this invention are that the intent recognition algorithm significantly improves the accuracy and robustness of recognition, especially in complex scenarios and multimodal environments. Multimodal input fusion and context-aware mechanisms effectively enhance the system's understanding of user intent, while dynamically weighted domain-specific vocabulary processing and an adaptive multi-task learning framework enhance the model's generalization ability. Detailed Implementation

[0010] The present invention will now be described in further detail with reference to the embodiments.

[0011] This embodiment of an intent recognition method based on deep learning mainly includes the following five parts: 1. Fusion of multimodal inputs Traditional intent recognition algorithms mostly rely solely on text input for analysis. This embodiment innovatively proposes a multimodal input fusion method that jointly processes data from different modalities, including text, speech, and images. By fusing deep learning models of multimodal data, it is possible to more accurately understand user intent, especially in intent recognition involving complex tasks and contexts.

[0012] 2. Context-aware intent recognition This embodiment proposes a context-aware mechanism that improves traditional intent recognition based on a single input by incorporating contextual information from preceding and following dialogues. By constructing a context-aware module based on the Transformer model, it can better understand long sequences of user dialogues and improve the accuracy of intent recognition.

[0013] 3. Dynamic weighting of domain-specific terms When dealing with intent recognition across different domains, domain-specific vocabulary often contains more semantic information. To improve recognition accuracy, this embodiment proposes a dynamic weighting mechanism based on domain-specific vocabulary. In the feature vector of the input text, the weights of domain-specific vocabulary are automatically adjusted, allowing these words to contribute more significantly to the final intent recognition result.

[0014] 4. Adaptive Multi-Task Learning Framework This embodiment also proposes an adaptive multi-task learning framework capable of simultaneously performing intent recognition and slot filling tasks. By sharing some network layers, the algorithm can be trained jointly on multiple tasks, thereby improving recognition accuracy and reducing model overfitting.

[0015] Furthermore, the fusion of the aforementioned multimodal inputs specifically involves: Assuming the input data includes text and images Text and images are processed using convolutional neural networks (CNN) and recurrent neural networks (RNN) for feature extraction, respectively, ultimately outputting text features. and image features Multimodal input fusion can be performed using a weighted average: Here, α and β are the weights of the text and image features, respectively, satisfying the condition α + β = 1. This allows for the reasonable fusion of information from different modalities, improving the effectiveness of intent recognition.

[0016] By integrating multiple modalities such as text and images, the algorithm can more accurately understand the user's intent in complex scenarios, especially in multi-tasking or complex situations, and can effectively improve the accuracy of recognition.

[0017] Furthermore, the aforementioned context-aware intent recognition specifically refers to: The context-aware module is modeled based on a self-attention mechanism. Given a series of input word vectors... Each word vector is processed through an attention mechanism to obtain a weighted output: in The attention weights are calculated based on the correlation between input words and can be obtained by calculating dot product attention. in, and These are the query vector and the key vector, respectively.

[0018] It is to calculate the query vector. With key vector The exponent of the dot product.

[0019] For all query vectors With key vector Perform the same calculation and sum the results.

[0020] Context-aware mechanisms enable models to understand dynamic changes in long dialogues and multi-turn interactions, especially through self-attention mechanisms to process long sequences, which improves recognition accuracy. Understanding changes in user intent reduces ambiguity or misunderstanding caused by single inputs, providing more accurate and unambiguous responses for scenarios such as multi-turn dialogues and customer service interactions, and enhancing the system's adaptability to complex dialogues.

[0021] Furthermore, the dynamic weighting of the aforementioned domain-specific terms is as follows: Given a domain vocabulary D and feature vectors of the input text The weight of each word can be calculated. And adjust the feature vector based on these weights: in, It represents the weight of the domain term d, and the weight can be learned through a deep neural network.

[0022] By enhancing the weight of specific words according to the specific needs of different fields, the system ensures that domain-specific terminology is accurately processed. This approach improves the system's recognition accuracy and adaptability across various industries (such as law, healthcare, and finance), increases the model's application value in multiple fields, and enhances the system's adaptability to industry-specific tasks.

[0023] Furthermore, the aforementioned adaptive multi-task learning framework is specifically as follows: Let the loss functions for the two tasks be respectively (Intent recognition) and (Slot filling), the overall loss function of the multi-task learning framework is: in, and It refers to task weights, which can adaptively adjust based on the importance of the task. .

[0024] By employing a multi-task learning framework, the system can handle multiple tasks simultaneously, such as intent recognition and slot filling, reducing training time and resource consumption while improving system accuracy. Simultaneously, it avoids overfitting, resulting in more stable system performance, especially in scenarios with limited training data.

[0025] In summary, the deep learning-based intent recognition method in this embodiment significantly improves the accuracy and robustness of intent recognition by combining contextual information, domain-specific vocabulary, and multimodal data.

[0026] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based intent recognition method, characterized in that: Includes the following steps: Step 1: Fuse data from different modalities, such as text, speech, and images, and output the fused data. Step 2: Build a context-aware module based on the Transformer model, which receives the fused data output from Step 1 and then adds context information from the preceding and following dialogues. Step 3: Determine the domain corresponding to the intent to be identified, and automatically adjust the weights of domain-specific words in the feature vector of the input text. Step four involves simultaneously performing intent recognition and slot filling tasks by sharing some network layers to improve the accuracy of the recognition model and reduce overfitting.

2. The deep learning-based intent recognition method according to claim 1, characterized in that: The specific method for fusing multimodal inputs in step one is as follows: Assuming the input data includes text and images Text and images are processed using convolutional neural networks and recurrent neural networks, respectively, to extract features, ultimately outputting text features. and image features Then, the fusion of multimodal inputs is achieved through a weighted average method.

3. The deep learning-based intent recognition method according to claim 1 or 2, characterized in that: In step two, the context-aware module is modeled based on a self-attention mechanism, given a series of input word vectors. Each word vector is processed through an attention mechanism to obtain a weighted output, thereby adding contextual information.

4. The deep learning-based intent recognition method according to claim 1 or 2, characterized in that: The specific method for automatically adjusting weights for domain-specific vocabulary in step three is as follows: given a domain vocabulary set D and the feature vector of the input text... Calculate the weight of each word. And adjust the feature vector based on these weights.

5. The deep learning-based intent recognition method according to claim 1 or 2, characterized in that: The specific method for performing intent recognition and slot filling tasks simultaneously through a shared portion of the network layer in step four is as follows: The loss functions for the two tasks, intent recognition and slot filling, are set as follows: and Then, the overall loss function of the multi-task learning framework is calculated, and the model is optimized based on the overall loss function.

Images