Information identification methods, devices, equipment and storage media

By separating the encoding of input information and the encoding of intent, features are extracted and fused separately, solving the problem of diluted intent features and improving the accuracy of key information extraction.

CN123115899APending Publication Date: 2026-11-17CHINA PING AN PROPERTY INSURANCE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611328872.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, intent features are diluted or submerged during the process of intent recognition and key information extraction, resulting in omissions or mismatches of key information items and low accuracy.

Method used

By encoding the input information to generate coded representations and intent codes, feature extraction is performed on the coded representations, intent codes, and combined results to generate fused features to identify key information.

Benefits of technology

It improves the accuracy of key information extraction, avoids omissions or mismatches, and ensures that intent features are not diluted during processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115899A_ABST
    Figure CN123115899A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of artificial intelligence technology and discloses a method, apparatus, device, and storage medium for information recognition. The method includes: acquiring input information; encoding the input information to generate an encoded representation of the input information; generating an intent encoding of the input information based on the encoded representation; combining the encoded representation and the intent encoding to obtain a combined result; extracting features from the encoded representation, the intent encoding, and the combined result to obtain a first feature, a second feature, and a third feature; fusing the first feature, the second feature, and the third feature to obtain a fused feature; and identifying key information related to intent in the input information based on the fused feature. This invention can be applied to business management systems in fintech, healthcare, and other fields, solving the technical problem of low accuracy in extracting key information from user input information in existing technologies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and is applied in the fields of financial technology and healthcare. In particular, it relates to an information identification method, device, equipment, and storage medium. Background Technology

[0002] In the field of information recognition, accurately identifying user intent and related key information from user-input text is a core technology in application scenarios such as intelligent customer service, agent assistance, and self-service systems. Taking the insurance business scenario as an example, if a user asks through online customer service or voice call, "I was in a car accident last week, how do I apply for compensation now?", the system needs to identify the user intent as "claims consultation" and extract key information such as "accident time: last week," "accident type: car accident," and "business type: claims" to direct the user to the appropriate claims service channel or assist an agent in quickly understanding the user's needs. Taking the healthcare scenario as an example, if a user enters "I've had a cough for several days and want to prescribe some medicine" on an internet hospital platform, the system needs to identify the user intent as "online consultation" and extract key information such as "symptoms: cough," "duration: several days," and "request type: prescription medicine" to direct the user to the appropriate consultation service channel or assist a doctor in quickly understanding the patient's complaint.

[0003] Currently, several technical solutions have been proposed to achieve the aforementioned intent recognition and key information extraction. For example, a deep learning-based joint modeling scheme can be used, where the intent recognition and key information extraction tasks are modeled simultaneously through a shared encoding layer. The intent recognition result is typically used as an auxiliary supervision signal for model training, or simply concatenated to the decoder's input during the decoding phase to aid in key information extraction. However, in this approach, the intent recognition result is merely used as an auxiliary supervision signal (only effective during training) or as a simple input feature to the decoder. During the transmission through the shared encoding layer, intent-related features and key information-related features become mixed, leading to the dilution or overwhelming of intent features. Due to this dilution or overwhelming, when the input information contains multiple information items, the intent features cannot provide sufficient guidance for key information extraction, resulting in significant issues with the omission or mismatch of key information items in the extraction results.

[0004] Therefore, the accuracy of extracting key information from user input is low in existing technologies. Summary of the Invention

[0005] This invention provides an information identification method, apparatus, device, and storage medium, which can solve the technical problem of low accuracy in extracting key information from user input information in the prior art.

[0006] In a first aspect, the present invention provides a method for identifying information, comprising: Obtain input information; The input information is encoded to generate an encoded representation of the input information; Based on the encoded representation, the intent encoding of the input information is generated; The encoded representation and the intent encoding are combined to obtain a combined result; Feature extraction is performed on the encoded representation, the intent encoding, and the combined result to obtain a first feature, a second feature, and a third feature; The first feature, the second feature, and the third feature are fused to obtain a fused feature; Based on the fusion features, key information in the input information that is associated with the intent is identified.

[0007] Secondly, the present invention provides an information identification device, comprising: The information acquisition module is used to acquire input information; The first encoding module is used to encode the input information and generate an encoded representation of the input information; The second encoding module is used to generate the intent encoding of the input information based on the encoding representation; The encoding combination module is used to combine the encoded representation and the intent encoding to obtain a combination result; The feature extraction module is used to extract features from the encoded representation, the intent encoding, and the combination result respectively to obtain a first feature, a second feature, and a third feature; The feature fusion module is used to fuse the first feature, the second feature and the third feature to obtain a fused feature; The information recognition module is used to identify key information in the input information that is associated with the intent based on the fusion features.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for recognizing the aforementioned information.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the method for recognizing the aforementioned information.

[0010] It is understood that the present invention encodes the input information to generate an encoded representation, thereby converting the input information into an encoded representation and providing basic features for subsequent processing.

[0011] Furthermore, this invention generates intent codes based on the encoded representation, thereby explicitly extracting intent-related features from the encoded representation to form independent intent codes. In this way, by specifically generating intent codes, intent features have independent feature representations in subsequent processing, rather than being merely partial information mixed in with the shared encoding layer. This directly addresses the problem of intent feature dilution by explicitly separating intent features from the encoded representation.

[0012] Furthermore, the present invention combines the encoded representation and the intent encoding to obtain a combined result, thereby combining the encoded representation and the intent encoding to generate a combined result that simultaneously contains both types of information.

[0013] Furthermore, this invention extracts features from the encoded representation, intent encoding, and combined result separately, obtaining a first feature, a second feature, and a third feature. In this way, feature extraction is performed independently on the encoded representation, intent encoding, and combined result. By extracting features separately, the complete contextual features in the encoded representation, the independent intent features in the intent encoding, and the interactive features resulting from their combination are processed through independent feature extraction channels without interference. This directly addresses the problem of intent features being mixed with key information features, ensuring that intent features have an independent feature extraction channel and are not diluted throughout the processing.

[0014] Furthermore, this invention fuses the first, second, and third features to obtain a fused feature. In this way, by fusing three independently extracted features, the fused feature retains not only the complete contextual features of the encoded representation, but also the independent intent features of the intent encoding, and the features of their interaction. Because the intent features are fully preserved through independent feature extraction channels, the intent information is not diluted in the fused feature.

[0015] Furthermore, this invention identifies key information based on fusion features. Thus, key information is identified based on fusion features. Since the intent information in the fusion features is not diluted, it provides sufficient guidance for the extraction of key information, thereby avoiding omissions or mismatches of key information items and improving the accuracy of key information extraction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a schematic diagram of an application environment for an information recognition method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an information identification method according to an embodiment of the present invention.

[0018] Figure 3 yes Figure 2 A flowchart of step S150.

[0019] Figure 4 This is another flowchart illustrating the information identification method in one embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of the structure of an information recognition device in one embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention.

[0022] Figure 7 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The information identification method provided in this embodiment of the invention can be applied to, for example... Figure 1 In the application environment, Figure 1This is a schematic diagram of an application environment for an information recognition method according to an embodiment of the present invention; wherein, the client communicates with the server via a network. The server can obtain input information; encode the input information to generate an encoded representation of the input information; generate an intent encoding of the input information based on the encoded representation; combine the encoded representation and the intent encoding to obtain a combined result; extract features from the encoded representation, the intent encoding, and the combined result respectively to obtain a first feature, a second feature, and a third feature; fuse the first feature, the second feature, and the third feature to obtain a fused feature; and identify key information in the input information related to the intent based on the fused feature. In this embodiment of the present invention, the input information is encoded to generate an encoded representation, thereby converting the input information into an encoded representation and providing basic features for subsequent processing. Further, this embodiment of the present invention generates an intent encoding based on the encoded representation, thereby explicitly extracting features related to the intent from the encoded representation to form an independent intent encoding. In this way, by specifically generating intent encoding, the intent features have independent feature expression in subsequent processing, rather than just being part of the information mixed in the shared encoding layer. This directly addresses the problem of diluted intent features by explicitly separating them from the encoded representation. Furthermore, embodiments of the invention combine the encoded representation and intent encoding to obtain a combined result, thereby generating a combined result that simultaneously contains both types of information. Further, embodiments of the invention extract features from the encoded representation, intent encoding, and combined result separately to obtain a first feature, a second feature, and a third feature. In this way, independent feature extraction is performed on the encoded representation, intent encoding, and combined result, ensuring that the complete contextual features in the encoded representation, the independent intent features in the intent encoding, and the interactive features resulting from their combination are processed through independent feature extraction channels without interference. This directly addresses the problem of mixed intent features and key information features, giving intent features an independent feature extraction channel and ensuring that intent features are not diluted throughout the processing. Further, embodiments of the invention fuse the first feature, the second feature, and the third feature to obtain a fused feature. In this way, the three independently extracted features are fused, so that the fused feature retains the complete contextual features of the encoded representation, the independent intent features of the intent encoding, and the interactive features between the two. Because the intent features are fully preserved after passing through independent feature extraction channels, the intent information in the fused features is not diluted. Furthermore, embodiments of the present invention identify key information based on the fused features. Thus, key information is identified based on the fused features. Since the intent information in the fused features is not diluted, it provides sufficient guidance for the extraction of key information, thereby avoiding omissions or mismatches of key information items and improving the accuracy of key information extraction.The present invention will now be described in detail through specific embodiments.

[0025] Figure 2 A flowchart of the information recognition method provided in the embodiments of the present invention, such as Figure 2 As shown, the information identification method provided in this embodiment of the invention includes the following steps.

[0026] Step S110: Obtain input information; In some embodiments, the method of obtaining input information may include any one or more combinations of the following: receiving text information input by a user through the input interface of a terminal device; receiving user voice data collected by a voice acquisition device, performing speech-to-text processing on the voice data, and using the converted text information as input information; receiving image data collected by an image acquisition device, performing Optical Character Recognition (OCR) processing on the image data, and using the recognized text content as input information. Input information may be a sentence, a paragraph, a question, or a query request entered by the user. Input information may also be user request data forwarded from other applications or servers.

[0027] In one specific implementation, the terminal device can be a mobile phone, computer, intelligent voice assistant, self-service terminal, etc. When the user inputs text or voice through the input interface of the above-mentioned device, the system can obtain the input information.

[0028] In a simple example in the property insurance sector, a user enters "I was in a car accident last week, how do I apply for compensation now" into the online customer service of an insurance company's app, and the system retrieves that text as input information.

[0029] In a simple example in the healthcare field, a user enters "I've had a cough for several days and would like to get some medicine" into an online hospital's consultation platform, and the system retrieves that text as input information.

[0030] Step S120: Encode the input information to generate an encoded representation of the input information; In some embodiments, encoding the input information may include: performing word segmentation or character segmentation on the input information to obtain each information unit of the input information; mapping each information unit to a vector representation; and then inputting the vector representation corresponding to each information unit into an encoder for encoding processing to generate an encoded representation of the input information.

[0031] Specifically, the encoder can employ a bidirectional long short-term memory (BiLSTM) network to encode the input sequence in both forward and backward directions, obtaining a hidden state vector corresponding to each information unit. The hidden state vector contains the semantic information of the information unit itself and its dependencies in the context. The hidden state vectors corresponding to each information unit constitute the encoded representation. Alternatively, the encoder can employ a self-attention mechanism-based encoder (such as the Transformer encoder), which generates the encoded representation by modeling the dependencies between positions in the sequence.

[0032] In one specific implementation, the input sequence can be encoded using BiLSTM first, and then the output of BiLSTM can be globally semantically modeled using a self-attention mechanism. The hidden state sequence output by BiLSTM is then concatenated with the feature sequence output by the self-attention mechanism to obtain the encoded representation.

[0033] In a simple example in the property insurance field, the system segments and encodes the sentence "I had a car accident last week, how do I apply for compensation now?" to generate an encoded representation of the sentence, which includes the contextual dependencies between the various lexical units.

[0034] In a simple example in the healthcare field, the system segments and encodes the sentence "I've been coughing for several days and want to get some medicine," generating an encoded representation of the sentence.

[0035] Step S130: Generate the intent code of the input information based on the encoding representation; In some embodiments, intent encoding is an encoded representation obtained by performing intent-related processing on the encoded representation, used to characterize the intent-related information contained in the input information. The intent encoding may have the same or different dimensions as the encoded representation. Intent encoding can be a global intent vector corresponding to the entire input information, or it can be a sequence representation that incorporates intent information, used for subsequent combination and fusion with the encoded representation to provide intent-level guidance for the identification of key information.

[0036] For example, the encoded representation can be input into a pre-trained or jointly trained intent feature extraction module, whose output is the intent encoding. In one implementation, the module can map the encoded representation into a fixed-dimensional global vector, where each dimension of the vector represents the relevance of the input information to various intents.

[0037] In a simple example in the property and casualty insurance field, the system processes the encoded representation generated in step S120 through the intent feature extraction module to generate an intent code that represents the user's intent. This intent code contains information related to the intent of "claims application".

[0038] In a simple example in the healthcare field, the system, based on the encoded representation generated in step S120, processes the encoded representation through the intent feature extraction module to generate an intent code that represents the user's intent. This intent code contains information related to the intent to "prescribe medication".

[0039] It should be noted that the specific implementation of the intent feature extraction module described above will be further described in subsequent embodiments.

[0040] In some embodiments of the present invention, step S130 includes the following steps.

[0041] Step S131: Perform sequence encoding on the encoded representation to obtain intermediate encoded features; In some embodiments, sequence encoding of the encoded representation refers to inputting the encoded representation into a sequence encoding network for processing. The sequence encoding network performs contextual modeling on the features at each position in the encoded representation and outputs intermediate encoded features. Intermediate encoded features are encoded features further extracted from the encoded representation and incorporating contextual information. The sequence encoding network can employ a neural network structure capable of modeling sequential data, such as a sequence encoding network based on Long Short-Term Memory (LSTM) or a sequence encoding network based on gated recurrent units (GRUs). The sequence encoding network can perform contextual modeling on the encoded representation from both forward and backward directions to capture the dependencies between positions within the surrounding context. The intermediate encoded features obtained after processing by the sequence encoding network contain richer contextual dependency information compared to the original encoded representation.

[0042] In a simple example within the property and casualty insurance sector, the system inputs the encoded representation generated in step S120 into a sequence coding network for sequence coding processing to obtain intermediate encoded features. These intermediate encoded features are context-enhanced features further extracted from the original encoded representation and are used for subsequent fusion with the original encoded representation.

[0043] In a simple example in the field of healthcare, the system inputs the encoded representation generated in step S120 into a sequence coding network for sequence coding processing to obtain the intermediate encoded features corresponding to the medical consultation text.

[0044] Step S132: The intermediate coding features are fused with the coding representation to obtain the fused coding features; In some embodiments, fusing intermediate coding features with coding representations means fusing the intermediate coding features obtained in step S131 with the original coding representation obtained in step S120 according to a preset fusion method to obtain fused coding features. By fusing intermediate coding features with the original coding representations, contextual information enhanced by sequence coding can be incorporated while retaining the original coding information, thereby preventing information loss during deep network processing. Fusion can be implemented in any of the following ways: element-wise addition, that is, adding the element values ​​at the same position in the intermediate coding features and the coding representation respectively; concatenation, that is, concatenating the intermediate coding features and the coding representation along the feature dimension; weighted summation, that is, assigning weights to the intermediate coding features and the coding representation respectively and then adding them together.

[0045] In a simple example in the property and casualty insurance field, the system fuses the intermediate coding features obtained in step S131 with the original coding representation obtained in step S120 (e.g., element-wise addition) to obtain fused coding features. These fused coding features retain the original coding information while incorporating contextual information enhanced by sequence coding.

[0046] In a simple example in the field of healthcare, the system fuses the intermediate encoding features obtained in step S131 with the original encoding representation obtained in step S120 to obtain the fused encoding features corresponding to the medical consultation text.

[0047] In some embodiments, after sequence encoding the encoded representation to obtain intermediate encoded features, the intermediate encoded features are fused with the encoded representation to obtain fused encoded features. The fusion ratio can be controlled by preset hyperparameters. Specifically, the hyperparameters may include a first hyperparameter and a second hyperparameter. The first hyperparameter controls the weight of the intermediate encoded features, and the second hyperparameter controls the weight of the encoded representation. By adjusting the values ​​of the first and second hyperparameters, the relative preservation degree of the intermediate encoded features and the encoded representation during the fusion process can be controlled. In some implementations, after fusing the intermediate encoded features and the encoded representation, the fusion result can be linearly transformed using trainable parameters, then nonlinearly mapped using a nonlinear activation function (e.g., the LeakyReLU activation function), and then normalized using the Sigmoid activation function to generate intent encoding. Through these methods, the degree of enhancement of intent information can be flexibly controlled in different application scenarios, improving the robustness of the model under low-resource samples.

[0048] Step S133: Perform a nonlinear transformation on the fused coding features to generate the intent code.

[0049] In some embodiments, performing a nonlinear transformation on the fused encoded features refers to inputting the fused encoded features obtained in step S132 into a nonlinear transformation layer for processing, mapping the fused encoded features through a nonlinear activation function to generate intent codes. The nonlinear transformation layer can employ any of the nonlinear activation functions such as the Sigmoid activation function, the Tanh activation function, or the ReLU activation function. Through nonlinear transformation, the fused encoded features can be mapped to a preset value range, while introducing nonlinear factors, resulting in intent codes with better expressive power and generalization performance. The intent code, i.e., the intent code in step S130, is used to combine with the encoded representation in subsequent steps to provide intent-level information guidance for the identification of key information.

[0050] In a simple example within the property insurance field, the system performs a nonlinear transformation on the fused encoded features obtained in step S132 using a nonlinear activation function (such as the sigmoid function) to generate the final intent code representing the user's intent. This intent code has the same dimension as the original encoded representation or a preset correspondence, and can be used for subsequent combination and feature extraction.

[0051] In a simple example in the field of healthcare, the system performs a nonlinear transformation on the fused coding features obtained in step S132 through a nonlinear activation function to generate the final intent code used to represent the user's intent.

[0052] This embodiment further defines a specific implementation method for generating intent codes based on the encoded representation, building upon step S130. First, the encoded representation is sequence encoded to obtain intermediate encoded features. Then, these intermediate encoded features are fused with the original encoded representation. Finally, a nonlinear transformation is performed on the fusion result to generate the intent code. This method enhances the contextual modeling capability of the encoded representation through sequence encoding, preserves the original encoded information through the fusion operation, effectively preventing information forgetting during deep network processing, and improves the expressive power of the intent code through nonlinear transformation. This allows the generated intent code to more accurately and comprehensively represent the intent information contained in the input information, providing higher-quality intent information input for subsequent three-way feature extraction and fusion, thereby improving the accuracy and robustness of the entire information recognition method.

[0053] Step S140: Combine the encoded representation and the intent encoding to obtain the combined result; In some embodiments, combining the encoded representation and the intent encoding can be achieved by combining them according to a preset combination method to obtain a combined result that includes information from both the encoded representation and the intent encoding. For example, the encoded representation and the intent encoding can be concatenated to obtain a combined result. Concatenation can be performed at the sequence dimension, such as concatenating the intent encoding as one or more global vectors to the beginning or end of the sequence of the encoded representation; it can also be performed at the feature dimension, such as concatenating the vectors at each position in the encoded representation with the global intent vector. Alternatively, the encoded representation and the intent encoding can be weighted and summed or multiplied bitwise to obtain a combined result.

[0054] In a simple example in the property insurance field, the system concatenates the encoded representation obtained in step S120 and the intent encoding obtained in step S130 along the feature dimension to obtain a combined result that integrates the original semantic information and intent information.

[0055] In a simple example in the field of healthcare, the system concatenates the encoded representation obtained in step S120 and the intent encoding obtained in step S130 along the feature dimension to obtain the combined result of the input information.

[0056] Step S150: Extract features from the encoded representation, intent encoding, and combination result respectively to obtain the first feature, the second feature, and the third feature; In some embodiments, feature extraction is performed separately on the encoded representation, intent encoding, and combined result. This means performing feature extraction operations on the encoded representation, intent encoding, and combined result separately to extract corresponding features from each type of information. Specifically, feature extraction on the encoded representation yields a first feature, feature extraction on the intent encoding yields a second feature, and feature extraction on the combined result yields a third feature. The three feature extraction operations process their respective inputs independently and do not affect each other. By performing feature extraction on the three inputs separately, multiple feature representations from different information sources can be obtained.

[0057] Each feature extraction operation can be implemented using a feature extraction network, such as an encoder that encodes each input and outputs the corresponding feature vector or feature sequence. The feature extraction network can be any one of a Bidirectional Long Short-Term Memory (BiLSTM) network, a Gated Recurrent Unit (GRU), a Convolutional Neural Network (CNN), or a Transformer encoder. Each feature extraction operation can use the same type of network structure or different types of network structures; this embodiment does not limit this. The specific implementation of each feature extraction operation will be further described in subsequent embodiments.

[0058] In a simple example in the property insurance field, the system extracts features from the encoded representation obtained in step S120, the intent encoding obtained in step S130, and the combined result obtained in step S140: the first feature is obtained by extracting features from the encoded representation, the second feature is obtained by extracting features from the intent encoding, and the third feature is obtained by extracting features from the combined result.

[0059] In a simple example in the healthcare field, the system extracts features from the encoded representation, intent encoding, and combined result of the medical consultation text, resulting in three different features.

[0060] In some embodiments of the present invention, such as Figure 3 As shown, step S150 includes the following steps.

[0061] Step S151: Through parallel bidirectional long short-term memory networks, feature extraction is performed on the encoded representation, intent encoding, and combination results to obtain the first feature, the second feature, and the third feature.

[0062] In some embodiments, feature extraction of the encoded representation, intent encoding, and combined results using parallel bidirectional long short-term memory networks (BiLSTMs) means employing three independent BiLSTM networks as feature extraction networks to process three different inputs, thereby obtaining features from three different sources. Specifically, the three BiLSTM networks may use the same network structure (e.g., all having the same number of hidden layers and hidden units), but have their own independent network parameters. Each BiLSTM network receives one type of input and performs forward computation independently. Each BiLSTM network can model the input sequence from both forward and backward directions, capturing the dependencies of each sequence position in the context and outputting the corresponding feature sequence.

[0063] In some embodiments, to capture feature information from different inputs from different perspectives, the three BiLSTM networks can be different BiLSTM networks. Here, "different" means that each BiLSTM network has independent network parameters. The BiLSTM networks do not share parameters and independently extract features from their respective inputs. By employing parallel different BiLSTM networks, it is possible to ensure that the three inputs obtain independent feature extraction while avoiding mutual interference between features from different inputs during the extraction stage. The feature vectors extracted by each BiLSTM network constitute the first feature, the second feature, and the third feature.

[0064] In a simple example within the property insurance field, the system employs three parallel BiLSTM networks with independent parameters to extract features from the encoded representation, intent encoding, and combined result obtained in the preceding steps: the first BiLSTM network extracts features from the encoded representation, yielding the first feature; the second BiLSTM network extracts features from the intent encoding, yielding the second feature; and the third BiLSTM network extracts features from the combined result, yielding the third feature. The three BiLSTM networks operate in parallel without interfering with each other. The resulting three features characterize the input information from different perspectives.

[0065] In a simple example in the healthcare field, the system uses the same approach to extract features from the encoded representation, intent encoding, and combined result of the medical consultation text through three parallel BiLSTM networks, resulting in three different features.

[0066] In some embodiments, when feature extraction is performed on the encoded representation, intent encoding, and combined result, the three parallel bidirectional long short-term memory networks can each have different task orientations. Specifically, the bidirectional long short-term memory network for feature extraction on the encoded representation is used to capture slot-related features in the input information, enhancing the representational ability of the slot extraction task; the bidirectional long short-term memory network for feature extraction on the intent encoding is used to capture intent-related features in the input information, enhancing the representational ability of the intent recognition task; and the bidirectional long short-term memory network for feature extraction on the combined result is used to capture the interaction features between the intent and the slot. By having the three bidirectional long short-term memory networks extract features from their respective inputs from different task perspectives, each network can learn feature representations specific to a particular task, thereby improving the joint effect of multi-intent recognition and key information extraction.

[0067] This embodiment further defines a specific implementation method for feature extraction of the encoded representation, intent encoding, and combination result, based on step S150. It uses parallel bidirectional long short-term memory networks to independently extract features from the three inputs. This method leverages the ability of BiLSTM networks to model sequences from both the preceding and following directions, effectively capturing the contextual dependencies in each input sequence. Simultaneously, by employing parallel BiLSTM networks, the three inputs are processed through independent feature extraction channels. This not only avoids mutual interference between features from different inputs during the extraction stage but also allows each channel to learn the optimal feature representation suitable for its respective input. This results in the extracted first, second, and third features more accurately and comprehensively reflecting the effective information in their respective inputs, providing higher-quality feature input for subsequent feature fusion, thereby improving the accuracy and robustness of the entire information recognition method.

[0068] Step S160: The first feature, the second feature, and the third feature are fused to obtain the fused feature; In some embodiments, fusing the first, second, and third features refers to integrating the features output from the three feature extraction channels to obtain a fused feature representation. Specifically, the fusion can be implemented in any of the following ways: element-wise addition, which adds the element values ​​at the same position in the three feature vectors respectively; concatenation, which concatenates the three feature vectors along the feature dimension; weighted summation, which assigns weights to the three features respectively and then performs a weighted summation; and gated fusion, which adaptively learns the fusion weights of each feature through a gating mechanism and then performs weighted fusion.

[0069] In a simple example in the property insurance field, the system adds the first feature, the second feature and the third feature obtained in step S150 element by element to obtain a fused feature, which simultaneously contains the original encoded information, the intent enhancement information and the interaction information between the two.

[0070] In a simple example in the field of healthcare, the system also adds the three features obtained in step S150 element by element to obtain the fused features corresponding to the medical consultation text.

[0071] In some embodiments of the present invention, step S160 includes the following steps.

[0072] The first feature vector, the second feature vector, and the third feature vector are added element by element to obtain the fused feature.

[0073] In some embodiments, element-wise addition of the first, second, and third feature vectors means that, assuming the first, second, and third feature vectors have the same dimension, the element values ​​at the same position in each of the three feature vectors are added together to obtain a new vector with the same dimension as the three original feature vectors. This new vector is the fused feature. Specifically, assuming the first, second, and third feature vectors are all n-dimensional vectors, represented as A = [a1, a2, ..., a...], ... n B = [b1, b2, ..., b] n ],C = [c1, c2,..., c n If the element-wise summation yields the fusion feature F = [a1+b1+c1, a2+b2+c2, ..., a], then the fusion feature obtained is F = [a1+b1+c1, a2+b2+c2, ..., a]. n +b n +c n By adding elements together, three feature vectors can be fused in the same feature dimension space without introducing additional trainable parameters.

[0074] In some embodiments, before element-wise addition, the first, second, and third feature vectors can be normalized or scaled separately to adjust their magnitudes before addition. When adding element-wise, the first, second, and third feature vectors need to have the same dimension. If the three feature vectors have different dimensions, they can be mapped to the same dimension using a linear mapping layer before element-wise addition.

[0075] In an example from the property insurance field, suppose the system extracts a first feature vector A = [0.8, 0.2, 0.5], a second feature vector B = [0.3, 0.7, 0.1], and a third feature vector C = [0.4, 0.5, 0.6], all with the same dimension. The system adds them element-wise to obtain a fused feature F = [0.8+0.3+0.4, 0.2+0.7+0.5, 0.5+0.1+0.6] = [1.5, 1.4, 1.2]. This fused feature integrates feature information from three different sources: encoded representation, intent encoding, and a combination of both.

[0076] In a healthcare example, the system extracts three feature vectors, the first feature vector, and the second feature vector, which also have the same dimension. The system adds them element-wise to obtain a fused feature. This fused feature integrates the feature information of the medical consultation text across three different dimensions, and is used for subsequent key information identification.

[0077] This embodiment further defines a specific implementation method for fusing the first feature vector, the second feature vector, and the third feature vector based on step S160: element-wise addition. This method does not require the introduction of additional trainable parameters, is computationally simple and efficient, and can achieve the fusion of three features without increasing the number of model parameters. Furthermore, element-wise addition can preserve the independent information of each dimension of features in the same feature space, avoiding mutual interference between features. This allows the fused features to simultaneously retain information from three different dimensions: the encoded representation, the intent encoding, and the combination of the two, providing richer feature representations for subsequent key information recognition and intent recognition.

[0078] Step S170: Based on the fusion features, identify key information in the input information that is associated with the intent.

[0079] In some embodiments, identifying key information associated with the intent in the input information based on fusion features refers to decoding or classifying the fusion features to extract one or more key information items associated with the intent from the input information. Key information is the specific information required to complete the intent. For example, when the intent is "to purchase insurance," key information may include the product name, the insured person, and the insurance period; when the intent is "to make an appointment," key information may include the department name, symptom description, and expected appointment time.

[0080] In one specific implementation, sub-features corresponding to each information unit in the input information can be obtained from the fused features. These sub-features are then classified according to multiple preset categories to obtain category labels for each information unit. The category labels are then combined according to the sequence position of each information unit in the input information to obtain the key information. For example, category labels can correspond to slot labels (such as "B-product", "I-product", "O", etc.). In another implementation, a sequence-to-sequence generation method can be used to directly generate a sequence representation of the key information based on the fused features.

[0081] In a simple example in the property insurance field, for the question "I was in a car accident last week, how do I apply for compensation now?", the system performs slot classification based on fusion features, identifying that "last week" corresponds to "accident time", "car accident" corresponds to "accident type", and "compensation" corresponds to "business type", and after combination, obtains the key information {accident time: last week, accident type: car accident, business type: compensation}.

[0082] In a simple example in the healthcare field, for the question "I've been coughing for several days and want to get some medicine," the system performs slot classification based on fusion features, identifying "cough" as "symptom," "several days" as "duration," and "get some medicine" as "demand type." After combining these, the system obtains the key information {symptom: cough, duration: several days, demand type: medicine}.

[0083] It is understood that embodiments of the present invention can obtain input information; encode the input information to generate an encoded representation of the input information; generate an intent encoding of the input information based on the encoded representation; combine the encoded representation and the intent encoding to obtain a combined result; extract features from the encoded representation, the intent encoding, and the combined result respectively to obtain a first feature, a second feature, and a third feature; fuse the first feature, the second feature, and the third feature to obtain a fused feature; and identify key information in the input information associated with the intent based on the fused feature. In embodiments of the present invention, the input information is encoded to generate an encoded representation, thereby converting the input information into an encoded representation and providing basic features for subsequent processing. Further, embodiments of the present invention generate an intent encoding based on the encoded representation, thereby explicitly extracting intent-related features from the encoded representation to form an independent intent encoding. In this way, by specifically generating intent encoding, the intent features have independent feature expression in subsequent processing, rather than just being part of the information mixed in the shared encoding layer. This directly addresses the problem of intent feature dilution by explicitly separating the intent features from the encoded representation. Furthermore, this embodiment of the invention combines the encoded representation and intent encoding to obtain a combined result, thereby generating a combined result that simultaneously contains both types of information. Further, this embodiment of the invention extracts features from the encoded representation, intent encoding, and combined result separately to obtain a first feature, a second feature, and a third feature. In this way, the encoded representation, intent encoding, and combined result are each subjected to independent feature extraction. This ensures that the complete contextual features in the encoded representation, the independent intent features in the intent encoding, and the interactive features resulting from their combination are processed through independent feature extraction channels without interference. This directly addresses the problem of mixed intent features and key information features, providing the intent features with an independent feature extraction channel and ensuring that the intent features are not diluted during the entire processing. Further, this embodiment of the invention fuses the first feature, the second feature, and the third feature to obtain a fused feature. This fuses the three independently extracted features, ensuring that the fused feature retains the complete contextual features of the encoded representation, the independent intent features of the intent encoding, and the interactive features between the two. Because the intent features are fully preserved through independent feature extraction channels, the intent information in the fused feature is not diluted. Furthermore, embodiments of the present invention identify key information based on fusion features. Thus, key information is identified based on fusion features. Since the intent information in the fusion features is not diluted, it provides sufficient guidance for the extraction of key information, thereby avoiding omissions or mismatches of key information items and improving the accuracy of key information extraction.

[0084] In some embodiments of the present invention, step S170 includes the following steps.

[0085] Step S171: Obtain the sub-features corresponding to each information unit of the input information in the fused features; In some embodiments, the input information is encoded as an coded representation containing multiple sequence positions. The fusion feature is obtained by extracting and fusing features through three paths based on the coded representation. Therefore, the fusion feature retains feature information corresponding to each sequence position of the input information; that is, the fusion feature contains sub-features corresponding to each information unit. Obtaining the sub-features corresponding to each information unit of the input information in the fusion feature means extracting the feature vectors corresponding to each information unit of the input information from the fusion feature. Each information unit of the input information refers to the basic unit obtained after word segmentation or character segmentation of the input information. Each information unit has its own corresponding sequence position in the input information. Since the coded representation maintains a one-to-one correspondence with each position in the input information sequence, the fusion feature generated based on this coded representation also maintains this correspondence. Therefore, sub-features corresponding to each information unit can be extracted from the fusion feature according to their sequence positions.

[0086] In a simple example in the property insurance field, a user inputs "I was in a car accident last week, how do I apply for compensation now?" This input is segmented into eight information units: "I," "last week," "car accident," "in," "now," "how," "apply," and "compensation." The system extracts eight sub-features corresponding to these eight information units from the fused features. Each sub-feature is associated with the semantic and contextual information of the information unit, and each sub-feature is used for subsequent classification of the information unit. In one specific implementation, if the fused features are a hidden state sequence h = [h1, h2, ..., h8], then h1 is extracted as the sub-feature corresponding to "I," h2 is extracted as the sub-feature corresponding to "last week," and so on, to obtain the sub-features corresponding to each information unit.

[0087] In a simple example in the field of healthcare, a user inputs "I've been coughing for several days and want to get some medicine." This input is segmented into eight information units: "I," "cough," "good," "several days," "has," "want," "get some," and "medicine." The system then extracts eight sub-features corresponding to these eight information units from the fused features. Each sub-feature is used to classify the information unit in the subsequent process.

[0088] Step S172: According to the preset multiple slot categories, classify each sub-feature separately to obtain the slot category corresponding to each information unit; In some embodiments, the preset multiple slot categories refer to a predefined set of labels used to mark the category to which each information unit in the input information belongs. Slot categories characterize the information role of an information unit in the input information. For example, for user questions in the insurance field, slot categories may include product type, time information, location information, amount information, service type, etc.; for user questions in the healthcare field, slot categories may include symptom name, department name, drug name, time information, body part, etc. Classifying each sub-feature separately means inputting each sub-feature obtained in step S171 into a classifier, which outputs the probability distribution or score of the sub-feature belonging to each preset slot category, and determining the slot category to which the information unit corresponding to the sub-feature belongs based on the probability distribution or score.

[0089] In one specific implementation, each sub-feature is processed by a linear mapping layer (such as a fully connected layer) and a softmax function, outputting a probability distribution for each slot category. The slot category with the highest probability is then taken as the slot category of the information unit. In another implementation, a Conditional Random Field (CRF) can be used to jointly model the dependencies between slot categories of adjacent information units, making the output slot category sequence for each information unit more coherent and reasonable.

[0090] In a simple example within the property insurance sector, the predefined slot categories include "Product Type," "Time Information," "Location Information," "Accident Type," and "Business Type." For the sub-feature corresponding to the information unit "Car Collision," the classifier categorizes it as "Accident Type"; for the sub-feature corresponding to "Last Week," the classifier categorizes it as "Time Information"; for the sub-feature corresponding to "Claim," the classifier categorizes it as "Business Type"; and for other information units (such as "I," "I have," "Now," "How," "Apply"), the classifier can categorize them as "Other" or non-critical information categories.

[0091] In a simple example in the healthcare field, the preset slot categories include "symptom name", "duration", "body part", and "need type". For the sub-feature corresponding to the information unit "cough", the classifier classifies it as "symptom name"; for the sub-feature corresponding to "for several days", the classifier classifies it as "duration"; and for the sub-feature corresponding to "prescribe some medicine", the classifier classifies it as "need type".

[0092] In some embodiments, after obtaining the slot category corresponding to each information unit, the slot features can be enhanced based on a Graph Attention Network (GAT). Specifically, a graph structure can be constructed, where nodes include slot nodes corresponding to each information unit in the input information and intent nodes of the input information. The graph structure can include a local slot graph and a global intent-slot graph. In the local slot graph, any slot node is only connected to other slot nodes within its context window in the input information, used to establish local dependencies between slots and alleviate the problem of slot inconsistency. In the global intent-slot graph, each slot node is connected to each intent node, used to realize global interaction between sentence-level intents and slots. Based on the graph structure, information is aggregated from the nodes in the graph structure through a graph attention network. Each node adaptively aggregates the feature information of its neighboring nodes through an attention mechanism to obtain enhanced slot features, and the final slot category of each information unit is determined based on the enhanced slot features. Graph attention networks can employ a multi-head attention mechanism, where multiple attention heads calculate attention coefficients and aggregate information. They can also include a trainable weight matrix for linear transformation of node features. Through this global-local interactive processing based on graph attention networks, slot recognition results can be more consistent and accurate in multi-intent scenarios.

[0093] Step S173: Combine the slot categories corresponding to each information unit according to the sequence position of each information unit in the input information to obtain key information.

[0094] In some embodiments, combining the slot categories corresponding to each information unit according to the sequence position of each information unit in the input information means arranging and combining the slot categories of each information unit obtained in step S172 according to the original order (i.e., sequence position) of each information unit in the input information to form a complete slot category sequence, which serves as key information. For example, if the sequence of each information unit in the input information is [w1, w2, w3, w4, w5], and the corresponding slot categories are [c1, c2, c3, c4, c5], then the key information obtained after combining according to the sequence position is a structured representation of [c1, c2, c3, c4, c5] or [c1, c2, c3, c4, c5]. Key information is structured information extracted from the input information, containing various attribute values ​​related to user intent, such as product name, time information, symptom description, etc., which can be directly used for downstream business processing.

[0095] In a simple example in the property insurance field, the slot categories for each information unit—"Me," "Last Week," "Collision," "Done," "Now," "How," "Apply," and "Claim"—are "Other," "Time Information," "Accident Type," "Other," "Other," "Other," "Other," and "Business Type," respectively. Combining these information units according to their sequential position in the input information yields the key information: {Time Information: Last Week, Accident Type: Collision, Business Type: Claim}. This structured key information can be pushed to the claims processing system to automatically fill out claims application forms or assist agents in quickly understanding user needs.

[0096] In a simple example in the healthcare field, the information units "I", "Cough", "Good", "A few days", "Finished", "Want", "Prescription", and "Medication" correspond to the slot categories "Other", "Symptom Name", "Other", "Duration", "Other", "Other", "Other", and "Need Type", respectively. Combining these in sequence yields the key information: {Symptom Name: Cough, Duration: Several days, Need Type: Prescription}. This structured key information can be pushed to an online consultation system to help doctors quickly understand the patient's chief symptoms and needs.

[0097] This embodiment further defines a specific implementation method for identifying key information based on fused features. It obtains sub-features corresponding to each information unit from the fused features, classifies each sub-feature to obtain the slot category corresponding to each information unit, and then combines the slot categories according to their sequence positions to obtain structured key information. This method decomposes the extraction of key information into fine-grained classification of each information unit, enabling more precise location of intent-related information fragments in the input information, thus improving the accuracy and interpretability of key information extraction. By combining the slot categories according to their sequence positions to form structured key information, the extracted key information has clear field meanings and can be easily used in downstream business systems for information filling, query routing, or agent assistance.

[0098] In some embodiments of the present invention, such as Figure 4 As shown, after step S160, the following steps are also included.

[0099] Step S180: Based on the fusion features, the intent label corresponding to the input information is identified.

[0100] In some embodiments, identifying the intent label corresponding to the input information based on the fusion features means performing classification processing based on the fusion features obtained in step S160 and outputting one or more intent labels corresponding to the input information, that is, determining the intent category expressed by the user input information (e.g., query intent, purchase intent, complaint intent, appointment intent, etc.). The intent label can be one or more category identifiers in a preset intent category set.

[0101] Specifically, the fusion features can be processed to map the degree of matching of the input information to each preset intent category, and the intent label of the input information can be determined based on the degree of matching of each preset intent category. For example, based on the fusion features, a classifier can output the degree of matching of the input information to each intent category, and select one or more intent categories with the highest degree of matching as the intent label corresponding to the input information. The aforementioned classifier can be trained together with the encoding and fusion parts of the entire recognition model during training, or it can be obtained by fine-tuning an existing pre-trained classifier.

[0102] In some embodiments, the identified intent label can be a single label (e.g., "claims consultation") or multiple labels (e.g., "purchase consultation" and "claims consultation" appearing simultaneously), depending on whether the input information contains multiple intents. Through the above method, the fusion features extracted in the preceding steps can be further transformed into highly readable intent labels that can be directly used for downstream business decisions, thereby providing more direct semantic understanding results for application scenarios such as intelligent customer service, agent assistance, and self-service triage.

[0103] In a simple example in the property insurance field, based on the fusion features obtained from the aforementioned steps, the system identifies the intent tag corresponding to the user's input information "I was in a car accident last week, how do I apply for compensation now" as "claims consultation". Based on this, the system can accurately route the user to the claims service channel or push the intent tag to the agent's end for reference.

[0104] In a simple example in the healthcare field, based on the fusion features obtained from the aforementioned steps, the system identifies the intent tag corresponding to the user's input information "I've had a cough for several days and would like to get some medicine" as "online consultation." Based on this, the system can accurately direct the user to the online consultation service channel or further guide the user to fill out a consultation form.

[0105] In a simple example in the property and casualty insurance field, where multiple intentions are involved, if a user enters "I want to buy an accident insurance policy and also ask about the progress of my previous car insurance claim," the system can simultaneously identify the two intention tags, "purchase inquiry" and "claims inquiry," thereby triggering the corresponding multi-intention joint processing flow.

[0106] It should be noted that the intent label recognition described above and the key information recognition (slot extraction) in step S170 of Embodiment 1 are two parallel branch tasks. They share the same fusion feature generated in step S160, but the recognition targets are different. Step S170 outputs key information items associated with the intent (i.e., the specific attribute values ​​required to complete the intent), while step S180 outputs the intent label corresponding to the input information (i.e., what the user wants to do). The combination of the two can provide comprehensive semantic understanding results for downstream services.

[0107] It is understood that, through this embodiment, the intent label corresponding to the input information can be further identified based on the fused features. Since the intent label is obtained based on the fused feature recognition generated by fusing three features—encoded representation, intent encoding, and the combination of the two—the intent recognition result can make full use of the interaction information between intent encoding and encoding representation, effectively improving the accuracy of intent recognition in multi-intent scenarios, and providing a reliable intent judgment basis for subsequent business processing such as intelligent routing, agent allocation, and automatic response.

[0108] In some embodiments of the present invention, step S180 includes the following steps.

[0109] Step S181: Perform a linear transformation on the fused features to obtain the feature vector corresponding to the fused features; In this feature vector, each element is used to represent a preset intent category, and the value of each element represents the degree of matching with the intent category corresponding to the element. In some embodiments, linear transformation of the fused features refers to inputting the fused features (usually a high-dimensional vector) obtained in the previous steps into a linear mapping layer for processing. The fused features are transformed from the original dimensional space to the intent category space through linear mapping, and a feature vector with the same dimension as the preset number of intent categories is output.

[0110] Specifically, the linear transformation can be implemented using a fully connected layer. This fully connected layer has a trainable weight matrix and bias terms. The fused features are linearly computed by the fully connected layer, outputting a feature vector. Assuming the dimension of the fused features is d and the number of preset intent categories is n, this linear transformation maps the d-dimensional vector to an n-dimensional vector. The dimension of the feature vector is the same as the total number of preset intent categories. Each element in the feature vector corresponds one-to-one with a preset intent category, that is, the i-th element corresponds to the i-th preset intent category. The value of each element represents the degree of matching between the input information and the intent category corresponding to that element, which can also be understood as the relevance or score of the input information to that intent category. The values ​​of each element can be real numbers; the larger the value, the higher the probability that the input information belongs to the corresponding intent category; the smaller the value, the lower the probability that the input information belongs to the corresponding intent category.

[0111] For example, if the preset intent categories include four categories: "purchase consultation", "claims consultation", "complaints and suggestions" and "casual chat", then after linear transformation, a four-dimensional feature vector is obtained. The first element value of the four-dimensional feature vector represents the degree of matching of the input information to "purchase consultation", the second element value represents the degree of matching of the input information to "claims consultation", the third element value represents the degree of matching of the input information to "complaints and suggestions", and the fourth element value represents the degree of matching of the input information to "casual chat".

[0112] In a simple example within the property insurance field, the preset intent categories include "claims consultation," "purchase consultation," "policy inquiry," and "complaints and suggestions." The system performs a linear transformation on the fused features obtained in the preceding steps, resulting in a four-dimensional feature vector, for example, [0.85, 0.12, 0.08, 0.03]. The first element, 0.85, represents the degree of matching for the input information "I was in a car accident last week, how do I apply for compensation?" to "claims consultation"; the second element, 0.12, represents the degree of matching to "purchase consultation"; the third element, 0.08, represents the degree of matching to "policy inquiry"; and the fourth element, 0.03, represents the degree of matching to "complaints and suggestions." Without normalization, each element value is a real number; the above values ​​are merely illustrative and used to represent the relative magnitudes of the elements.

[0113] In a simple example in the healthcare field, the preset intent categories include "online consultation," "appointment booking," "report inquiry," and "medication purchase." The system performs a linear transformation on the fused features obtained in the preceding steps to obtain a four-dimensional feature vector, for example, [0.72, 0.15, 0.08, 0.05], where the first element value of 0.72 represents the degree to which the input information "I have been coughing for several days and would like to get some medicine" matches "online consultation."

[0114] Step S182: Determine the target intent category corresponding to the element with the largest value in the feature vector, and use the target intent category as the intent label of the input information.

[0115] In some embodiments, determining the target intent category corresponding to the element with the largest value in the feature vector means selecting the element with the largest value from all elements of the feature vector obtained in step S181, identifying the position of this element in the feature vector (i.e., the element index), determining the preset intent category corresponding to this element based on the element index, and using this preset intent category as the intent label for the input information. For example, if the feature vector is a four-dimensional vector, with the four dimensions corresponding to "purchase consultation," "claims consultation," "complaints and suggestions," and "casual conversation," and the second dimension has the largest value, then "claims consultation" is determined as the intent label for the input information. By using the above method of "selecting the category corresponding to the maximum value," continuous matching degree values ​​can be converted into discrete intent category outputs, thereby obtaining intent labels that can be directly used for downstream business decisions.

[0116] In one specific implementation, the value of each element in the feature vector can be a normalized probability value (such as a probability distribution processed by the Softmax function), or it can directly use the output value of an unnormalized linear transformation (i.e., logits). Regardless of the form, the category corresponding to the element with the largest value is used as the intent label.

[0117] In a simple example in the property insurance field, the four-dimensional feature vector obtained in step S181 is [0.85, 0.12, 0.08, 0.03]. The system determines that the element with the largest value is the first element (value is 0.85), determines that the preset intent category corresponding to this element is "claims consultation", and uses "claims consultation" as the intent label of the input information.

[0118] In a simple example in the field of healthcare, the four-dimensional feature vector obtained in step S181 is [0.72, 0.15, 0.08, 0.05]. The system determines that the element with the largest value is the first element (with a value of 0.72), determines that the preset intent category corresponding to this element is "online consultation", and uses "online consultation" as the intent label for this input information.

[0119] In some embodiments, identifying the intent label corresponding to the input information can also be achieved using a word-level multi-label classification combined with a voting mechanism. Specifically, each information unit in the input information is classified using multi-label intent classification to obtain the classification result (e.g., positive or negative prediction) for each information unit in each preset intent category. Then, for each preset intent category, the number of positive predictions for that category among all information units is counted. When this number exceeds a preset threshold, the intent category is determined as one of the intent labels of the input information. The preset threshold can be a pre-set value, such as half the number of all information units. Through the above voting mechanism, the intent recognition result does not depend on the classification result of a single information unit, but rather makes a decision based on the prediction results of all information units, thereby capturing the distribution characteristics of each information unit in the input information across different intents and improving the accuracy and robustness of intent recognition in multi-intent scenarios.

[0120] This embodiment further defines a specific implementation method for identifying intent labels based on fused features. A linear transformation maps the fused features to the intent category space, obtaining matching degree values ​​for each preset intent category. The category with the highest matching degree value is then selected as the intent label. This method is simple to implement, computationally efficient, and can quickly and accurately convert fused features into intent labels. Furthermore, the linear transformation layer can be jointly trained end-to-end with the entire recognition model, ensuring consistency between the feature representation of intent classification and the feature representation of the entire recognition framework, further improving the accuracy of intent recognition in multi-intent scenarios. The intent labels identified through this embodiment can be used for intelligent agent list filtering. When a user's intent is a genuine need such as "claims consultation," they are identified as a high-value candidate and pushed to agents; when a user's intent is uncertain or falls into categories with no clear intent such as "casual conversation," they are identified as a low-value candidate and filtered. It can also be used for intelligent triage in the healthcare field. When a user's intent is "online consultation," they are directed to the doctor's consultation channel; when the user's intent is "report query," they are directed to the report query channel.

[0121] In some embodiments, the above-described information recognition method may further include a step of jointly training the model. Specifically, during training, a first loss value for intent recognition and a second loss value for key information recognition are determined. The first loss value measures the difference between the intent label predicted by the model and the actual intent label, and can be calculated using the cross-entropy loss function; the second loss value measures the difference between the key information (i.e., slot sequence) predicted by the model and the actual key information, and can also be calculated using the cross-entropy loss function. The first and second loss values ​​are weighted and summed to obtain a weighted sum used to update the model parameters, and the model parameters are updated based on the weighted sum. The weight coefficients used in the weighted sum can be adjusted according to the actual application scenario. When the application scenario emphasizes the accuracy of intent recognition, the weight of the first loss value can be increased; when the application scenario emphasizes the accuracy of key information extraction, the weight of the second loss value can be increased. Through joint training, the two sub-tasks of intent recognition and key information extraction can share the underlying encoding representation and fusion features during training, enabling the two tasks to promote each other and optimize together, thereby improving the overall performance of the model.

[0122] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0123] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0124] In one embodiment, an information recognition device is provided, which corresponds one-to-one with the information recognition methods described in the above embodiments. For example... Figure 5 As shown, the information recognition device includes an information acquisition module 510, a first encoding module 520, a second encoding module 530, an encoding combination module 540, a feature extraction module 550, a feature fusion module 560, and an information recognition module 570. Detailed descriptions of each functional module are as follows: Information acquisition module 510 is used to acquire input information; The first encoding module 520 is used to encode the input information and generate an encoded representation of the input information; The second encoding module 530 is used to generate the intent encoding of the input information based on the encoding representation; The encoding combination module 540 is used to combine the encoded representation and the intent encoding to obtain the combined result; The feature extraction module 550 is used to extract features from the encoded representation, intent encoding and the combination result respectively to obtain the first feature, the second feature and the third feature; The feature fusion module 560 is used to fuse the first feature, the second feature and the third feature to obtain the fused feature; The information recognition module 570 is used to identify key information in the input information that is associated with the intent based on the fusion features.

[0125] In one embodiment, the information recognition module 570 is further configured to: Based on the fusion features, the intent label corresponding to the input information is identified.

[0126] In one embodiment, the information recognition module 570 is further configured to: A linear transformation is performed on the fused features to obtain the feature vector corresponding to the fused features. Each element of the feature vector is used to represent a preset intent category, and the value of each element represents the degree of matching with the intent category corresponding to the element. Determine the target intent category corresponding to the element with the largest value in the feature vector, and use the target intent category as the intent label of the input information.

[0127] In one embodiment, the information recognition module 570 is specifically used for: Obtain the sub-features corresponding to each information unit of the input information in the fused features; According to the preset multiple slot categories, each sub-feature is classified and processed separately to obtain the slot category corresponding to each information unit; The key information is obtained by combining the slot categories corresponding to each information unit according to the sequence position of each information unit in the input information.

[0128] In one embodiment, the second encoding module 530 is specifically used for: Sequence encoding is performed on the encoded representation to obtain intermediate encoded features; The intermediate coding features are fused with the coding representation to obtain the fused coding features; The fused coding features are subjected to a nonlinear transformation to generate the intent code.

[0129] In one embodiment, the feature extraction module 550 is specifically used for: By using different parallel bidirectional long short-term memory networks, feature extraction is performed on the encoded representation, intent encoding, and combined results to obtain the first feature, the second feature, and the third feature.

[0130] In one embodiment, the feature fusion module 560 is specifically used for: The first feature vector, the second feature vector, and the third feature vector are added element by element to obtain the fused feature.

[0131] It is understood that embodiments of the present invention can obtain input information; encode the input information to generate an encoded representation of the input information; generate an intent encoding of the input information based on the encoded representation; combine the encoded representation and the intent encoding to obtain a combined result; extract features from the encoded representation, the intent encoding, and the combined result respectively to obtain a first feature, a second feature, and a third feature; fuse the first feature, the second feature, and the third feature to obtain a fused feature; and identify key information in the input information associated with the intent based on the fused feature. In embodiments of the present invention, the input information is encoded to generate an encoded representation, thereby converting the input information into an encoded representation and providing basic features for subsequent processing. Further, embodiments of the present invention generate an intent encoding based on the encoded representation, thereby explicitly extracting intent-related features from the encoded representation to form an independent intent encoding. In this way, by specifically generating intent encoding, the intent features have independent feature expression in subsequent processing, rather than just being part of the information mixed in the shared encoding layer. This directly addresses the problem of intent feature dilution by explicitly separating the intent features from the encoded representation. Furthermore, this embodiment of the invention combines the encoded representation and intent encoding to obtain a combined result, thereby generating a combined result that simultaneously contains both types of information. Further, this embodiment of the invention extracts features from the encoded representation, intent encoding, and combined result separately to obtain a first feature, a second feature, and a third feature. In this way, the encoded representation, intent encoding, and combined result are each subjected to independent feature extraction. This ensures that the complete contextual features in the encoded representation, the independent intent features in the intent encoding, and the interactive features resulting from their combination are processed through independent feature extraction channels without interference. This directly addresses the problem of mixed intent features and key information features, providing the intent features with an independent feature extraction channel and ensuring that the intent features are not diluted during the entire processing. Further, this embodiment of the invention fuses the first feature, the second feature, and the third feature to obtain a fused feature. This fuses the three independently extracted features, ensuring that the fused feature retains the complete contextual features of the encoded representation, the independent intent features of the intent encoding, and the interactive features between the two. Because the intent features are fully preserved through independent feature extraction channels, the intent information in the fused feature is not diluted. Furthermore, embodiments of the present invention identify key information based on fusion features. Thus, key information is identified based on fusion features. Since the intent information in the fusion features is not diluted, it provides sufficient guidance for the extraction of key information, thereby avoiding omissions or mismatches of key information items and improving the accuracy of key information extraction.

[0132] Identification methods based on the above information, such as Figure 6As shown in the diagram, this embodiment of the invention also provides a structural schematic of an apparatus for an information identification method. The apparatus includes a processor 61 and a memory 62 coupled to the processor 61. The memory 62 stores a computer program, which, when executed by the processor 61, causes the processor 61 to perform the steps of the information identification method described in the above embodiment.

[0133] For further details regarding the implementation of the above technical solution by the processor 61 in the device for the above information identification method steps, please refer to the description of the information identification method provided in the above-described embodiments of the invention, which will not be repeated here.

[0134] The processor 61 can also be called a CPU (Central Processing Unit). The processor 61 may be an integrated circuit chip with signal processing capabilities. The processor 61 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or the processor 61 can be any conventional processor.

[0135] like Figure 7 As shown in the diagram, this embodiment of the invention also provides a schematic diagram of a computer-readable storage medium, on which a readable computer program 71 is stored. The computer program 71 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in various embodiments of the invention. The aforementioned storage medium includes various storage media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks or optical disks, ROM (Read-Only Memory), RAM (Random Access Memory), or terminal devices such as computers, servers, mobile phones, and tablets.

[0136] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0137] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0139] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0140] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available storage medium that a computer can store or a data storage device such as a server or data center that integrates one or more available storage media. The available storage medium may be a magnetic storage medium (e.g., floppy disk, hard disk, magnetic tape, optical storage medium) or a semiconductor storage medium (e.g., SSD (solid state disk)).

[0141] The technical solution provided by the present invention has been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0142] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0143] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0146] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying information, characterized in that, include: Obtain input information; The input information is encoded to generate an encoded representation of the input information; Based on the encoded representation, the intent encoding of the input information is generated; The encoded representation and the intent encoding are combined to obtain a combined result; Feature extraction is performed on the encoded representation, the intent encoding, and the combined result to obtain a first feature, a second feature, and a third feature; The first feature, the second feature, and the third feature are fused to obtain a fused feature; Based on the fusion features, key information related to the intent is identified in the input information.

2. The information identification method according to claim 1, characterized in that, After fusing the first feature, the second feature, and the third feature to obtain the fused feature, the method further includes: Based on the fusion features, the intent label corresponding to the input information is identified.

3. The information identification method according to claim 2, characterized in that, The step of identifying the intent label corresponding to the input information based on the fusion features includes: A linear transformation is performed on the fused features to obtain the feature vector corresponding to the fused features. Each element of the feature vector is used to represent a preset intent category, and the value of each element represents the degree of matching with the intent category corresponding to the element. Determine the target intent category corresponding to the element with the largest value in the feature vector, and use the target intent category as the intent label of the input information.

4. The information identification method according to claim 1, characterized in that, The step of identifying key information related to intent in the input information based on the fusion features includes: Obtain the sub-features in the fusion features that correspond to each information unit of the input information; According to multiple preset slot categories, each of the sub-features is classified and processed to obtain the slot category corresponding to each information unit; The key information is obtained by combining the slot categories corresponding to each information unit according to the sequence position of each information unit in the input information.

5. The information identification method according to claim 1, characterized in that, The step of generating the intent encoding of the input information based on the encoded representation includes: The encoded representation is sequence encoded to obtain intermediate encoded features; The intermediate coding features are fused with the coding representation to obtain the fused coding features; The fused coding features are subjected to a nonlinear transformation to generate the intent code.

6. The information identification method according to claim 1, characterized in that, The step of extracting features from the encoded representation, the intent encoding, and the combined result to obtain a first feature, a second feature, and a third feature includes: By using different parallel bidirectional long short-term memory networks, features are extracted from the encoded representation, the intent encoding, and the combined result to obtain the first feature, the second feature, and the third feature.

7. The information identification method according to claim 1, characterized in that, The process of fusing the first feature, the second feature, and the third feature to obtain a fused feature includes: The first feature vector, the second feature vector, and the third feature vector are added element by element to obtain the fused feature.

8. An information identification device, characterized in that, include: The information acquisition module is used to acquire input information; The first encoding module is used to encode the input information and generate an encoded representation of the input information; The second encoding module is used to generate the intent encoding of the input information based on the encoding representation; The encoding combination module is used to combine the encoded representation and the intent encoding to obtain a combination result; The feature extraction module is used to extract features from the encoded representation, the intent encoding, and the combination result respectively to obtain a first feature, a second feature, and a third feature; The feature fusion module is used to fuse the first feature, the second feature and the third feature to obtain a fused feature; The information recognition module is used to identify key information in the input information that is associated with the intent based on the fusion features.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the information recognition method as described in any one of claims 1 to 7.

10. A computer-readable medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the information recognition method as described in any one of claims 1 to 7.