Large model data intelligent annotation method and system

Abstract

The present application relates to the field of data annotation technology, and specifically discloses a large-model data intelligent annotation method and system, which uses natural language processing technology based on deep learning to perform confidence assessment on each unlabeled text data in an unlabeled text data set, selects text data corresponding to the minimum confidence as representative sample data, and performs corpus expansion on the representative sample data, and then, by performing semantic feature extraction and compensatory interactive fusion on the representative sample data and the representative sample data after corpus expansion, the common information and unique information between the two are fully utilized, thereby achieving comprehensive semantic understanding and intelligent annotation of the representative sample data. In this way, the efficiency and accuracy of data annotation can be significantly improved, while the need for manual intervention is greatly reduced, and the annotation cost is reduced.

Description

Technical Field

[0001] The present application relates to the technical field of data annotation, and more specifically, to a method and system for intelligent annotation of large model data. Background Art

[0002] With the rapid development of the Internet and information technology, the amount of data has exploded. In the era of big data, the quality and value of data are increasingly valued, and high-quality data annotation is an indispensable part of machine learning, especially deep learning. Specifically, data annotation refers to the process of adding labels to raw data so that it can be used to train supervised learning algorithms, enabling them to learn from labeled data and apply them to new, unseen data.

[0003] However, although the traditional manual annotation method can ensure the annotation quality, it is inefficient and costly when facing large-scale data sets, and it is difficult to meet the needs of practical applications. Most automated annotation methods mainly rely on predefined rules or simple statistical models, lacking in-depth semantic understanding of the data set, resulting in insufficient accuracy and reliability of the annotation results.

[0004] Based on this, we look forward to an optimized large-model data intelligent labeling method and system. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a large model data intelligent annotation method and system, which uses natural language processing technology based on deep learning to perform confidence assessment on each unlabeled text data in the unlabeled text data set, selects the text data corresponding to the minimum confidence as the representative sample data, and expands the corpus of the representative sample data, and then, through the semantic feature extraction and compensatory interactive fusion of the representative sample data and the representative sample data after the corpus expansion, the common information and unique information between the two are fully utilized, so as to achieve comprehensive semantic understanding and intelligent annotation of the representative sample data. In this way, the efficiency and accuracy of data annotation can be significantly improved, while the need for manual intervention is greatly reduced, and the annotation cost is reduced.

[0006] According to one aspect of the present application, a large model data intelligent labeling method is provided, which includes:

[0007] Get unlabeled text dataset;

[0008] Performing confidence evaluation on each unlabeled text data in the unlabeled text data set to determine representative sample data;

[0009] Performing corpus expansion on the representative sample data to obtain an expanded description of the representative sample data;

[0010] Semantically encoding the representative sample data and the extended description of the representative sample data respectively to obtain a semantically encoded feature vector of the representative sample data and a semantically encoded feature vector of the extended description of the representative sample data;

[0011] Performing a compensatory interactive fusion on the semantic encoding feature vector representing the sample data and the semantic encoding feature vector representing the extended description of the sample data to obtain a semantic interactive compensated fusion representation vector representing the sample data;

[0012] Based on the semantic interaction compensation fusion representation vector of the representative sample data, a labeling result is generated.

[0013] In the above-mentioned large model data intelligent labeling method, the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description are subjected to compensatory interactive fusion, including: extracting the common features between the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description to obtain a common feature representation vector between the representative sample data features; using the common feature representation vector between the representative sample data features as a mask, guiding the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description to perform feature complementarity enhancement interactive fusion to obtain the semantic interactive compensatory fusion representation vector of the representative sample data.

[0014] In the above-mentioned large model data intelligent labeling method, confidence evaluation is performed on each unlabeled text data in the unlabeled text data set to determine representative sample data, including: semantic embedding coding is performed on each unlabeled text data in the unlabeled text data set to obtain a set of unlabeled data semantic embedding coding vectors; each unlabeled data semantic embedding coding vector in the set of unlabeled data semantic embedding coding vectors is input into a classifier-based confidence evaluator to obtain a set of unlabeled sample confidences; and the unlabeled text data corresponding to the minimum confidence in the set of unlabeled sample confidences is selected as the representative sample data.

[0015] In the above-mentioned large model data intelligent annotation method, the representative sample data is corpus expanded to obtain the representative sample data expanded description, including: inputting the representative sample data into a corpus expansion module based on a large language model to obtain the representative sample data expanded description.

[0016] In the above-mentioned large model data intelligent labeling method, the common features between the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description are extracted to obtain the common feature representation vector between the representative sample data features, including: inputting the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description into a common feature extraction network to obtain the common feature representation vector between the representative sample data features, wherein the common feature extraction network performs point-wise fusion on the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description, and then uses a neural network layer based on the tanh function to extract common features to obtain the common feature representation vector between the representative sample data features.

[0017] In the above-mentioned large model data intelligent labeling method, the common feature representation vector between the representative sample data features is used as a mask to guide the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector to perform feature complementarity enhancement interactive fusion to obtain the representative sample data semantic interactive compensation fusion representation vector, including: based on the specific features of the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector relative to the common feature representation vector between the representative sample data features, feature compensation modulation is performed on the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector to obtain the representative sample data semantic feature compensation vector and the representative sample data extended description semantic feature compensation vector; the representative sample data semantic feature compensation vector, the representative sample data extended description semantic feature compensation vector and the common feature representation vector between the representative sample data features are cascaded to obtain the common feature representation vector between the representative sample data features.

[0018] In the above-mentioned large model data intelligent labeling method, based on the specific features of the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector relative to the common feature representation vector between the representative sample data features, the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector are feature compensated and modulated to obtain the representative sample data semantic feature compensation vector and the representative sample data extended description semantic feature compensation vector, including: inputting the representative sample data semantic coding feature vector, the representative sample data extended description semantic coding feature vector and the representative sample data extended description common feature representation vector into a vector probabilization unit based on the Sigmoid function to obtain the probabilistic representative sample data semantic coding feature vector, the probabilistic representative sample data extended description semantic coding feature vector and the probabilistic representative sample data extended description common feature representation vector Common feature representation vector; calculate the specific features of the probabilistic representative sample data semantic coding feature vector relative to the probabilistic representative sample data feature common feature representation vector to obtain the representative sample data specific feature compensation representation vector; calculate the specific features of the probabilistic representative sample data extended description semantic coding feature vector relative to the probabilistic representative sample data feature common feature representation vector to obtain the representative sample data extended description specific feature compensation representation vector; input the representative sample data specific feature compensation representation vector and the representative sample data semantic coding feature vector into the fine-grained compensation module to obtain the representative sample data semantic feature compensation vector; input the representative sample data extended description specific feature compensation representation vector and the representative sample data extended description semantic coding feature vector into the fine-grained compensation module to obtain the representative sample data extended description semantic feature compensation vector.

[0019] In the above-mentioned large model data intelligent labeling method, the specific features of the probabilistic semantic coding feature vector representing the sample data relative to the probabilistic common feature representation vector representing the sample data features are calculated to obtain a compensation representation vector representing the specific features of the sample data, including: calculating the dot division vector between the probabilistic semantic coding feature vector representing the sample data and the probabilistic common feature representation vector representing the sample data features, and calculating the base-two logarithmic value of the absolute value of each eigenvalue in the dot division vector to obtain the specific feature vector representing the sample data; calculating the dot product vector between the specific feature vector representing the sample data and the probabilistic semantic coding feature vector representing the sample data, and calculating the exponential function value with base e and each eigenvalue of the dot product vector as the exponent to obtain an exponential representation vector representing the specific features of the sample data; inputting the exponential representation vector representing the specific features of the sample data into the softmax function for normalization to obtain the compensation representation vector representing the specific features of the sample data.

[0020] In the above-mentioned large model data intelligent labeling method, the specific feature compensation representation vector representing the sample data and the semantic encoding feature vector representing the sample data are input into a fine-grained compensation module to obtain the semantic feature compensation vector representing the sample data, including: calculating the Hadamard product between the specific feature compensation representation vector representing the sample data and the semantic encoding feature vector representing the sample data to obtain the semantic feature compensation vector representing the sample data.

[0021] In the above-mentioned large model data intelligent labeling method, a labeling result is generated based on the semantic interaction compensation fusion representation vector of the representative sample data, including: inputting the semantic interaction compensation fusion representation vector of the representative sample data into a classifier-based intelligent labeling module to obtain the labeling result.

[0022] According to another aspect of the present application, a large model data intelligent annotation system is provided, which includes:

[0023] A data set acquisition module is used to acquire unlabeled text data sets;

[0024] A confidence evaluation module, used for performing confidence evaluation on each unlabeled text data in the unlabeled text data set to determine representative sample data;

[0025] A corpus expansion module, used for expanding the corpus of the representative sample data to obtain an expanded description of the representative sample data;

[0026] A semantic coding module, used for performing semantic coding on the representative sample data and the extended description of the representative sample data to obtain a semantic coding feature vector of the representative sample data and a semantic coding feature vector of the extended description of the representative sample data;

[0027] A feature interaction fusion module, used for performing a compensatory interaction fusion on the semantic encoding feature vector representing the sample data and the semantic encoding feature vector representing the extended description of the sample data to obtain a semantic interaction compensation fusion representation vector representing the sample data;

[0028] The annotation result generation module is used to generate the annotation result based on the semantic interaction compensation fusion representation vector of the representative sample data.

[0029] Compared with the prior art, the large-model data intelligent annotation method and system provided by the present application adopts natural language processing technology based on deep learning to perform confidence assessment on each unlabeled text data in the unlabeled text data set, selects the text data corresponding to the minimum confidence as the representative sample data, and expands the corpus of the representative sample data, and then, by extracting semantic features and compensatory interactive fusion of the representative sample data and the representative sample data after corpus expansion, the common information and unique information between the two are fully utilized, thereby achieving comprehensive semantic understanding and intelligent annotation of the representative sample data. In this way, the efficiency and accuracy of data annotation can be significantly improved, while the need for manual intervention is greatly reduced, and the annotation cost is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0031] Figure 1 The present invention is a flowchart of a method for intelligently labeling large model data according to an embodiment of the present application.

[0032] Figure 2 Schematic diagram of data flow of a large model data intelligent annotation method according to an embodiment of the present application.

[0033] Figure 3 This is a flowchart of sub-step S2 of the large model data intelligent labeling method according to an embodiment of the present application.

[0034] Figure 4 This is a flowchart of sub-step S5 of the large model data intelligent labeling method according to an embodiment of the present application.

[0035] Figure 5 This is a flowchart of sub-step S52 of the large model data intelligent labeling method according to an embodiment of the present application.

[0036] Figure 6 This is a flowchart of sub-step S521 of the large model data intelligent annotation method according to an embodiment of the present application.

[0037] Figure 7 It is a block diagram of a large model data intelligent annotation system according to an embodiment of the present application. DETAILED DESCRIPTION

[0038] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0039] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0040] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0041] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0042] In response to the technical problems described in the above background technology, this application proposes an optimized large-model data intelligent annotation method, which uses natural language processing technology based on deep learning to perform confidence assessment on each unlabeled text data in the unlabeled text data set, selects the text data corresponding to the minimum confidence as the representative sample data, and expands the corpus of the representative sample data, and then extracts semantic features and performs compensatory interactive fusion on the representative sample data and the representative sample data after corpus expansion to make full use of the common information and unique information between the two, thereby achieving comprehensive semantic understanding and intelligent annotation of the representative sample data. In this way, the efficiency and accuracy of data annotation can be significantly improved, while the need for manual intervention is greatly reduced, reducing the cost of annotation.

[0043] Figure 1 The present invention is a flowchart of a method for intelligently labeling large model data according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the large model data intelligent labeling method according to the embodiment of the present application. Figure 1 and Figure 2As shown, the large model data intelligent annotation method includes the following steps: S1, obtaining an unlabeled text data set; S2, performing confidence assessment on each unlabeled text data in the unlabeled text data set to determine representative sample data; S3, performing corpus expansion on the representative sample data to obtain an expanded description of the representative sample data; S4, performing semantic encoding on the representative sample data and the expanded description of the representative sample data to obtain a semantic encoding feature vector of the representative sample data and a semantic encoding feature vector of the expanded description of the representative sample data respectively; S5, performing compensatory interactive fusion on the semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the expanded description of the representative sample data to obtain a semantic interactive compensated fusion representation vector of the representative sample data; S6, generating an annotation result based on the semantic interactive compensated fusion representation vector of the representative sample data.

[0044] In the above-mentioned large model data intelligent labeling method, the step S1 is to obtain an unlabeled text dataset. In the embodiment of the present application, the unlabeled text dataset can be any type of text dataset, such as news articles, social media posts, academic papers, product reviews, etc. In addition, in the present application, all actions of obtaining data are carried out under the premise of complying with the corresponding data protection laws and policies of the country of residence and obtaining authorization from the owner of the corresponding device.

[0045] In the above-mentioned large model data intelligent labeling method, the step S2 is to perform confidence evaluation on each unlabeled text data in the unlabeled text data set to determine representative sample data. Figure 3 FIG. 1 is a flowchart of sub-step S2 of the large model data intelligent labeling method according to an embodiment of the present application. Figure 3 As shown, the step S2 includes the steps of: S21, performing semantic embedding coding on each unlabeled text data in the unlabeled text data set to obtain a set of unlabeled data semantic embedding coding vectors; S22, inputting each unlabeled data semantic embedding coding vector in the set of unlabeled data semantic embedding coding vectors into a classifier-based confidence evaluator to obtain a set of unlabeled sample confidences; S23, selecting the unlabeled text data corresponding to the minimum confidence in the set of unlabeled sample confidences as the representative sample data.

[0046] Specifically, the step S21, semantic embedding coding is performed on each unlabeled text data in the unlabeled text data set to obtain a set of semantic embedding coding vectors of unlabeled data. It should be understood that in order to achieve semantic understanding and unified representation of unlabeled text data, so as to facilitate subsequent data annotation analysis, the present application further performs semantic embedding coding on each unlabeled text data in the unlabeled text data set, so as to map the text data to a high-dimensional semantic feature space, so as to better understand the intrinsic meaning of each unlabeled text data, and convert it into a continuous vector representation in the semantic feature space. In this way, the contextual meaning of unlabeled text data can be understood, rather than being limited to the statistical analysis of vocabulary on the surface, thereby effectively improving the accuracy of data annotation. In an embodiment of the present application, the word2vec model is adopted to realize the semantic embedding coding of the unlabeled text data set.

[0047] Specifically, in step S22, each unlabeled data semantic embedding coding vector in the set of unlabeled data semantic embedding coding vectors is input into a confidence evaluator based on a classifier to obtain a set of unlabeled sample confidences. It should be understood that, considering that in a large-scale unlabeled text data set, labeling all text data is not only inefficient but also costly, the present application proposes a representative sample selection method based on confidence assessment. Specifically, the present application calculates the confidence score of each unlabeled data semantic embedding coding vector through a confidence evaluator based on a classifier to reveal the semantic information richness of each unlabeled text data and the necessity of data labeling, so as to facilitate the targeted selection of representative samples for labeling.

[0048] Specifically, the step S23 selects the unlabeled text data corresponding to the minimum confidence in the set of unlabeled sample confidences as the representative sample data. It should be understood that text data with a low confidence score indicates that its semantics are relatively vague, usually located at the boundary of different categories, with high uncertainty, and it is difficult for the model to make a clear classification decision. By selecting the unlabeled text data corresponding to the minimum confidence in the set of unlabeled sample confidences as the representative sample data, the effect of annotation can be maximized under limited annotation resources, the efficiency and quality of annotation can be improved, and unnecessary annotation of samples that the model has already determined can be avoided, thereby saving annotation costs and time.

[0049] In the above-mentioned large model data intelligent annotation method, the step S3 performs corpus expansion on the representative sample data to obtain an expanded description of the representative sample data. In a specific example of the present application, the step S3 further includes: inputting the representative sample data into a corpus expansion module based on a large language model to obtain an expanded description of the representative sample data. It should be understood that since the representative sample data may contain less information, it is difficult to directly classify and annotate. Therefore, in order to further enrich its content, the present application uses a corpus expansion module based on a large language model to expand the corpus of the representative sample data to generate an expanded description of the representative sample data that is semantically related to the representative sample data and contains more contextual information and detailed descriptions, thereby increasing the semantic information volume and diversity of the sample data. In this way, additional contextual information can be provided for the data annotation of the representative sample data, enhancing the model's semantic understanding of the representative sample data, thereby improving the accuracy of the annotation. In an embodiment of the present application, an advanced large language model such as a GPT (Generative Pre-trained Transformer) model can be used to perform corpus expansion tasks.

[0050] In the above-mentioned large model data intelligent annotation method, the step S4, respectively, semantically encodes the representative sample data and the extended description of the representative sample data to obtain the semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the extended description of the representative sample data. It should be understood that in order to extract the semantic features of the representative sample data and the extended description of the representative sample data, the present application further uses a pre-trained language model to semantically encode the two respectively, so as to capture the deep contextual semantic information in the text data, and convert the two into continuous vector representations, thereby obtaining the semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the extended description of the representative sample data. In an embodiment of the present application, the BERT model is used to semantically encode the representative sample data and the extended description of the representative sample data respectively. Those of ordinary skill in the art should know that the BERT model is based on a bidirectional Transformer architecture design, and can simultaneously consider the left and right context information of each vocabulary unit in the representative sample data and the extended description of the representative sample data, so as to more effectively understand the text structure and semantic information, and provide a more accurate semantic representation for subsequent classification and annotation tasks.

[0051] In the above-mentioned large model data intelligent annotation method, the step S5 performs compensatory interactive fusion on the semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the extended description of the representative sample data to obtain a semantic interactive compensated fusion representation vector of the representative sample data. It should be understood that, considering that the extended description of the representative sample data has semantic relevance to the representative sample data, there may be information differences between the two in some aspects. Therefore, in order to better utilize the complementary information between the two in the feature fusion stage and enhance the comprehensive understanding of the representative sample data, the present application proposes a compensatory interactive fusion method, which mines the common features between the semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the extended description of the representative sample data, and uses the common features as a mask to perform feature compensatory interactive fusion on the two, so as to make full use of the common information and unique information between the representative sample data and its extended description, while reducing the redundancy between the two, so as to achieve a comprehensive semantic understanding of the representative sample data.

[0052] Figure 4 FIG. 5 is a flowchart of sub-step S5 of the large model data intelligent labeling method according to an embodiment of the present application. Figure 4 As shown, the step S5 includes the steps of: S51, extracting the common features between the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data to obtain a common feature representation vector between the features of the representative sample data; S52, using the common feature representation vector between the features of the representative sample data as a mask, guiding the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data to perform feature complementarity enhancement interactive fusion to obtain the semantic interactive compensation fusion representation vector of the representative sample data.

[0053] Specifically, the step S51 further includes: inputting the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data into a common feature extraction network to obtain a common feature representation vector between the features of the representative sample data, wherein the common feature extraction network performs point addition fusion on the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data, and then uses a neural network layer based on a tanh function to extract common features to obtain a common feature representation vector between the features of the representative sample data. That is, the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data are input into the common feature extraction network, and a neural network architecture is used to learn the common pattern between the two, and to mine the internal connection between the two, so as to generate a common feature representation vector between the features of the representative sample data, thereby providing a reference benchmark for the subsequent feature complementarity enhancement and interactive fusion.

[0054] Figure 5 FIG. 5 is a flowchart of sub-step S52 of the large model data intelligent labeling method according to an embodiment of the present application. Figure 5 As shown, the step S52 includes the steps of: S521, based on the specific features of the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description relative to the common feature representation vector between the representative sample data features, performing feature compensation modulation on the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the representative sample data extended description to obtain the semantic feature compensation vector of the representative sample data and the semantic feature compensation vector of the representative sample data extended description; S522, cascading the semantic feature compensation vector of the representative sample data, the semantic feature compensation vector of the representative sample data extended description and the common feature representation vector between the representative sample data features to obtain the common feature representation vector between the representative sample data features.

[0055] Figure 6 FIG. 5 is a flowchart of sub-step S521 of the large model data intelligent labeling method according to an embodiment of the present application. Figure 6 As shown, the step S521 includes the steps of: S5211, inputting the semantic coding feature vector representing the sample data, the semantic coding feature vector representing the extended description of the sample data, and the common feature representation vector representing the common features between the features of the sample data into a vector probabilization unit based on the Sigmoid function to obtain the probabilistic semantic coding feature vector representing the sample data, the probabilistic semantic coding feature vector representing the extended description of the sample data, and the probabilistic common feature representation vector representing the common features between the features of the sample data; S5212, calculating the specific features of the probabilistic semantic coding feature vector representing the sample data relative to the probabilistic common feature representation vector representing the common features between the features of the sample data to obtain the specific feature compensation representation vector representing the sample data ; S5213, calculate the specific features of the semantic coding feature vector of the probabilistic representative sample data extended description relative to the common feature representation vector between the probabilistic representative sample data features to obtain the specific feature compensation representation vector of the representative sample data extended description; S5214, input the specific feature compensation representation vector of the representative sample data and the semantic coding feature vector of the representative sample data into the fine-grained compensation module to obtain the semantic feature compensation vector of the representative sample data; S5215, input the specific feature compensation representation vector of the representative sample data extended description and the semantic coding feature vector of the representative sample data extended description into the fine-grained compensation module to obtain the semantic feature compensation vector of the representative sample data extended description.

[0056] In a specific example of the present application, the specific features of the probabilistic semantic coding feature vector representing the sample data relative to the probabilistic common feature representation vector representing the sample data features are calculated to obtain a compensation representation vector representing the specific features of the sample data, including: calculating the dot division vector between the probabilistic semantic coding feature vector representing the sample data and the probabilistic common feature representation vector representing the sample data features, and calculating the base-two logarithmic value of the absolute value of each eigenvalue in the dot division vector to obtain the specific feature vector representing the sample data; calculating the dot product vector between the specific feature vector representing the sample data and the probabilistic semantic coding feature vector representing the sample data, and calculating the exponential function value with base e as the exponent and each eigenvalue of the dot product vector as the exponent to obtain an exponential representation vector representing the specific features of the sample data; inputting the exponential representation vector representing the specific features of the sample data into the softmax function for normalization to obtain the compensation representation vector representing the specific features of the sample data.

[0057] In a specific example of the present application, the representative sample data specific feature compensation representation vector and the representative sample data semantic encoding feature vector are input into a fine-grained compensation module to obtain the representative sample data semantic feature compensation vector, including: calculating the Hadamard product between the representative sample data specific feature compensation representation vector and the representative sample data semantic encoding feature vector to obtain the representative sample data semantic feature compensation vector.

[0058] That is, first, the representative sample data semantic coding feature vector, the representative sample data extended description semantic coding feature vector, and the generated representative sample data feature common feature representation vector are input into the vector probabilization unit based on the Sigmoid function for normalization processing to ensure that the value range of the feature vector is between 0 and 1, so as to better measure the importance of the feature. Then, the specific features of the probabilistic representative sample data semantic coding feature vector and the probabilistic representative sample data extended description semantic coding feature vector relative to the probabilistic representative sample data feature common feature representation vector are calculated respectively, and the feature differences between the original features and the common features are learned to mine the unique semantic information of the representative sample data and the representative sample data extended description in addition to the common features, so as to generate the representative sample data specific feature compensation representation vector and the representative sample data extended description specific feature compensation representation vector. Subsequently, the specific feature compensation representation vector of the representative sample data and the specific feature compensation representation vector of the representative sample data extended description are used as weights, and the original semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the representative sample data extended description are compensated and strengthened by positional point multiplication, so as to ensure that during the fusion process, as much unique information of the representative sample data and the representative sample data extended description is retained as much as possible.

[0059] More specifically, in step S522, the representative sample data semantic feature compensation vector, the representative sample data extended description semantic feature compensation vector and the representative sample data feature common feature representation vector are cascaded to obtain the representative sample data feature common feature representation vector. That is, the representative sample data semantic encoding feature vector and the representative sample data extended description semantic encoding feature vector that have been compensated and enhanced are fused with the representative sample data feature common feature representation vector to comprehensively consider the common information of the representative sample data and the representative sample data extended description and the unique information in their respective feature sources, and generate the representative sample data semantic interaction compensation fusion representation vector, thereby realizing the comprehensive capture of the representative sample data semantic information.

[0060] Accordingly, the step S5 includes: processing the representative sample data semantic encoding feature vector and the representative sample data extended description semantic encoding feature vector using the following feature compensation interaction formula to obtain a representative sample data semantic interaction compensation fusion representation vector, wherein the feature compensation interaction formula is:

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] in, represents the semantic encoding feature vector representing the sample data, represents the semantic encoding feature vector of the extended description of the representative sample data, represents the common feature representation vector between sample data features, and Represent the weight parameter matrix and bias term respectively, represents the hyperbolic tangent function, represents the sigmoid function, Indicates adding by position point. represents the probabilistic semantic encoding feature vector of sample data, It represents the probabilistic representation of the sample data expansion description semantic encoding feature vector, represents the common feature representation vector between the features of sample data in a probabilistic way. , and They represent the first The probabilistic representation of the sample data expansion describes the semantic coding feature vector The eigenvalue and the probabilistic representation of the common feature representation vector between the sample data features eigenvalues, represents the logarithmic function with base 2, represents the natural exponential function, represents the normalized exponential function, and They represent the compensation representation vector of specific features of sample data and the compensation representation vector of specific features of extended description of sample data respectively. represents dot product, and They represent the semantic feature compensation vector representing the sample data and the semantic feature compensation vector representing the extended description of the sample data, respectively. Indicates cascade operation, Represents the semantic interaction compensation fusion representation vector of the representative sample data.

[0071] In the above-mentioned large model data intelligent labeling method, the step S6 generates a labeling result based on the semantic interaction compensation fusion representation vector of the representative sample data. In a specific example of the present application, the step S6 further includes: inputting the semantic interaction compensation fusion representation vector of the representative sample data into a classifier-based intelligent labeling module to obtain the labeling result. Specifically, the classifier learns the deep semantic information of the representative sample data contained therein by performing multi-layer perception processing on the semantic interaction compensation fusion representation vector of the representative sample data, and maps the semantic interaction compensation fusion representation vector of the representative sample data to the corresponding category label in combination with the feature classification mapping rules learned during the training process, thereby achieving accurate semantic labeling of the representative sample data.

[0072] Here, in the case where the semantic coding feature vector representing the sample data and the semantic coding feature vector representing the extended description of the sample data respectively represent the text semantic coding features representing the sample data and the text semantic coding features representing the extended description of the sample data, when performing compensatory interaction between features with common features as masks, the common feature masked interactive compensation representation of the semantic interactive compensation fusion representation vector representing the sample data will also have a lack of interactive feature instance decision-making relative to the respective text content distribution field selective state space of the source modal data, thereby affecting the accuracy of the annotation results obtained through the classifier-based intelligent annotation module.

[0073] Preferably, inputting the semantic interaction compensation fusion representation vector representing the sample data into a classifier-based intelligent annotation module to obtain an annotation result includes:

[0074] Calculate the distance between each pair of eigenvalues ​​of the semantic interaction compensation fusion representation vector representing the sample data, such as the L2 distance, and take the square root of the distance to obtain the distance representation matrix representing the semantic interaction compensation fusion of the sample data, that is,

[0075]

[0076] in, represents the semantic interaction compensation fusion representation vector of the representative sample data, and The first representation vector representing the semantic interaction compensation fusion representation vector of the representative sample data is represented by and Eigenvalues, The first representation vector representing the semantic interaction compensation fusion representation vector of the representative sample data is represented by and The distance between eigenvalues, Represents the sample data semantic interaction compensation fusion distance representation matrix The eigenvalue of the position;

[0077] The representative sample data semantic interaction compensation fusion autocorrelation matrix of the representative sample data semantic interaction compensation fusion representation vector as a row vector is obtained, that is, ,in, represents the semantic interaction compensation fusion representation vector of the representative sample data, represents the transpose of a vector, represents matrix multiplication, It represents the semantic interaction compensation fusion autocorrelation matrix of sample data;

[0078] The representative sample data semantic interaction compensation fusion representation vector is matrix-multiplied with the representative sample data semantic interaction compensation fusion distance representation matrix to obtain the representative sample data semantic interaction compensation fusion first-level mapping vector, that is, ,in, represents the semantic interaction compensation fusion representation vector of the representative sample data, It represents the semantic interaction compensation fusion distance representation matrix of sample data, represents matrix multiplication, It represents the first-level mapping vector representing the semantic interaction compensation fusion of sample data;

[0079] The first-level mapping vector of the semantic interaction compensation fusion of the representative sample data is matrix-multiplied by the matrix product of the distance representation matrix of the semantic interaction compensation fusion of the representative sample data and the self-correlation matrix of the semantic interaction compensation fusion of the representative sample data to obtain a multi-level mapping vector of the semantic interaction compensation fusion of the representative sample data ,in, represents the first-level mapping vector representing the semantic interaction compensation fusion of sample data, It represents the semantic interaction compensation fusion distance representation matrix of sample data, represents matrix multiplication, represents the semantic interaction compensation fusion autocorrelation matrix of sample data, It represents the multi-level mapping vector of semantic interaction compensation fusion of sample data;

[0080] The representative sample data semantic interaction compensation fusion multi-level mapping vector is interpolated with the representative sample data semantic interaction compensation fusion associated eigenvector composed of the eigenvalues ​​of the representative sample data semantic interaction compensation fusion auto-association matrix to obtain an optimized representative sample data semantic interaction compensation fusion representation vector, wherein interpolation or zero padding is performed when the eigenvalue is insufficient;

[0081] The optimized representative sample data semantic interaction compensation fusion representation vector is input into a classifier-based intelligent labeling module to obtain a labeling result.

[0082] Correspondingly, through the linear target mapping representation based on the similarity distance representation matrix of the semantic interaction compensation fusion representation vector representing the sample data, the self-correlation of the complete similarity instantiation of the semantic interaction compensation fusion representation vector representing the sample data is instantiated by a secondary target mapping representation based on a multi-level distribution hierarchy, and the association mismatch negative influence factor is compensated by the association fusion kernel bias to improve the degree of instance determination of the eigenvalue of the semantic interaction compensation fusion representation vector representing the sample data under the similarity constraint, that is, the degree of significance of the eigenvalue as an instance for classification and regression determination, thereby improving the accuracy of the annotation results obtained by the classifier-based intelligent annotation module of the semantic interaction compensation fusion representation vector representing the sample data.

[0083] In summary, a large model data intelligent annotation method based on an embodiment of the present application is explained, which uses natural language processing technology based on deep learning to perform confidence assessment on each unlabeled text data in an unlabeled text data set, selects text data corresponding to the minimum confidence as representative sample data, and performs corpus expansion on the representative sample data, and then, by performing semantic feature extraction and compensatory interactive fusion on the representative sample data and the representative sample data after corpus expansion, the common information and unique information between the two are fully utilized, thereby achieving comprehensive semantic understanding and intelligent annotation of the representative sample data. In this way, the efficiency and accuracy of data annotation can be significantly improved, while the need for manual intervention is greatly reduced, reducing the cost of annotation.

[0084] Furthermore, a large model data intelligent labeling system is also provided.

[0085] Figure 7 FIG. 1 is a block diagram of a large model data intelligent annotation system according to an embodiment of the present application. Figure 7 As shown, according to the large model data intelligent annotation system 100 of the embodiment of the present application, it includes: a data set acquisition module 110, which is used to acquire an unlabeled text data set; a confidence assessment module 120, which is used to perform confidence assessment on each unlabeled text data in the unlabeled text data set to determine the representative sample data; a corpus expansion module 130, which is used to perform corpus expansion on the representative sample data to obtain an expanded description of the representative sample data; a semantic encoding module 140, which is used to perform semantic encoding on the representative sample data and the expanded description of the representative sample data to obtain a semantic encoding feature vector of the representative sample data and a semantic encoding feature vector of the expanded description of the representative sample data; a feature interaction fusion module 150, which is used to perform compensatory interactive fusion on the semantic encoding feature vector of the representative sample data and the semantic encoding feature vector of the expanded description of the representative sample data to obtain a semantic interaction compensation fusion representation vector of the representative sample data; and a labeling result generation module 160, which is used to generate a labeling result based on the semantic interaction compensation fusion representation vector of the representative sample data.

[0086] Here, those skilled in the art can understand that the specific operations of each module in the above-mentioned large model data intelligent annotation system have been referred to above. Figures 1 to 6 The large model data intelligent labeling method has been introduced in detail, and therefore, its repeated description will be omitted.

[0087] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to being implemented by adopting the above specific details.

[0088] In the above embodiments, the description of each embodiment has its own emphasis. For the parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiment described above is only schematic. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0089] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference to a figure in a claim should not be considered as limiting the claim to which it relates.

[0090] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0091] Finally, it should be noted that the above description has been given for the purpose of illustration and description. In addition, the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A large model data intelligent labeling method, characterized in that: include: Get unlabeled text dataset; Performing confidence evaluation on each unlabeled text data in the unlabeled text data set to determine representative sample data, and selecting the unlabeled text data corresponding to the minimum confidence in the set of unlabeled sample confidences as the representative sample data; Performing corpus expansion on the representative sample data to obtain an expanded description of the representative sample data; Semantically encoding the representative sample data and the extended description of the representative sample data respectively to obtain a semantically encoded feature vector of the representative sample data and a semantically encoded feature vector of the extended description of the representative sample data; Performing a compensatory interactive fusion on the semantic encoding feature vector representing the sample data and the semantic encoding feature vector representing the extended description of the sample data to obtain a semantic interactive compensated fusion representation vector representing the sample data; Generate a labeling result based on the semantic interaction compensation fusion representation vector of the representative sample data; Among them, the compensatory interactive fusion of the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data includes: extracting the common features between the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data to obtain a common feature representation vector between the features of the representative sample data; using the common feature representation vector between the features of the representative sample data as a mask, guiding the semantic coding feature vector of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data to perform feature complementarity enhancement interactive fusion to obtain the semantic interactive compensation fusion representation vector of the representative sample data; The common feature representation vector between the representative sample data features is used as a mask to guide the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector to perform feature complementarity enhancement interactive fusion to obtain the representative sample data semantic interactive compensation fusion representation vector, including: Based on the specific features of the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector relative to the common feature representation vector between the representative sample data features, feature compensation modulation is performed on the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector to obtain a representative sample data semantic feature compensation vector and a representative sample data extended description semantic feature compensation vector; The representative sample data semantic feature compensation vector, the representative sample data extended description semantic feature compensation vector and the representative sample data feature common feature representation vector are cascaded to obtain the representative sample data feature common feature representation vector.

2. The large model data intelligent labeling method according to claim 1 is characterized in that: Performing a confidence evaluation on each unlabeled text data in the unlabeled text dataset to determine representative sample data, including: Performing semantic embedding coding on each unlabeled text data in the unlabeled text data set to obtain a set of unlabeled data semantic embedding coding vectors; Each unlabeled data semantic embedding coding vector in the set of unlabeled data semantic embedding coding vectors is input into a classifier-based confidence evaluator to obtain a set of unlabeled sample confidences.

3. The large model data intelligent labeling method according to claim 2 is characterized in that: Expanding the representative sample data to obtain an expanded description of the representative sample data includes: The representative sample data is input into a corpus expansion module based on a large language model to obtain an expanded description of the representative sample data.

4. The large model data intelligent labeling method according to claim 3 is characterized in that: Extracting common features between the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector to obtain a common feature representation vector between representative sample data features, including: The representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector are input into a common feature extraction network to obtain a common feature representation vector between the representative sample data features, wherein the common feature extraction network performs point-wise fusion on the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector, and then uses a tanh function-based neural network layer to extract common features to obtain a common feature representation vector between the representative sample data features.

5. The large model data intelligent labeling method according to claim 4 is characterized in that: Based on the specific features of the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector relative to the common feature representation vector between the representative sample data features, feature compensation modulation is performed on the representative sample data semantic coding feature vector and the representative sample data extended description semantic coding feature vector to obtain a representative sample data semantic feature compensation vector and a representative sample data extended description semantic feature compensation vector, including: Input the representative sample data semantic coding feature vector, the representative sample data extended description semantic coding feature vector and the representative sample data feature common feature representation vector into a vector probabilization unit based on a Sigmoid function to obtain a probabilistic representative sample data semantic coding feature vector, a probabilistic representative sample data extended description semantic coding feature vector and a probabilistic representative sample data feature common feature representation vector; Calculate the specific features of the semantic encoding feature vector representing the sample data with respect to the common feature representation vector between the features of the sample data to obtain a compensation representation vector representing the specific features of the sample data; Calculate the specific features of the semantic encoding feature vector representing the extended description of the probabilistic sample data relative to the common feature representation vector between the probabilistic sample data features to obtain a compensation representation vector representing the specific features of the extended description of the sample data; Inputting the representative sample data specific feature compensation representation vector and the representative sample data semantic encoding feature vector into a fine-grained compensation module to obtain the representative sample data semantic feature compensation vector; The specific feature compensation representation vector of the extended description of the representative sample data and the semantic coding feature vector of the extended description of the representative sample data are input into the fine-grained compensation module to obtain the semantic feature compensation vector of the extended description of the representative sample data.

6. The large model data intelligent labeling method according to claim 5, characterized in that: Calculating the specific features of the semantic encoding feature vector representing the sample data with respect to the common feature representation vector between the features of the sample data representing the sample data to obtain a compensation representation vector representing the specific features of the sample data, including: Calculate the dot-wise division vector between the semantic encoding feature vector representing the probabilistic sample data and the common feature representation vector representing the features of the probabilistic sample data, and calculate the base-two logarithm of the absolute value of each eigenvalue in the dot-wise division vector to obtain a specific feature vector representing the sample data; Calculate the dot product vector between the specific feature vector representing the sample data and the probabilistic semantic encoding feature vector representing the sample data, and calculate an exponential function value with e as the base and each eigenvalue of the dot product vector as the exponent to obtain an exponential specific feature representation vector representing the sample data; The indexed representation vector representing the specific features of the sample data is input into a softmax function for normalization processing to obtain the compensated representation vector representing the specific features of the sample data.

7. The large model data intelligent labeling method according to claim 6 is characterized in that: Inputting the representative sample data specific feature compensation representation vector and the representative sample data semantic encoding feature vector into a fine-grained compensation module to obtain the representative sample data semantic feature compensation vector, including: The Hadamard product between the representative sample data specific feature compensation representation vector and the representative sample data semantic encoding feature vector is calculated to obtain the representative sample data semantic feature compensation vector.

8. The large model data intelligent labeling method according to claim 7 is characterized in that: Based on the semantic interaction compensation fusion representation vector of the representative sample data, a labeling result is generated, including: The semantic interaction compensation fusion representation vector of the representative sample data is input into a classifier-based intelligent labeling module to obtain the labeling result.

9. A large model data intelligent labeling system, used to execute the large model data intelligent labeling method according to claim 1, characterized in that: include: A data set acquisition module is used to acquire unlabeled text data sets; A confidence evaluation module, used for performing confidence evaluation on each unlabeled text data in the unlabeled text data set to determine representative sample data; A corpus expansion module, used for expanding the corpus of the representative sample data to obtain an expanded description of the representative sample data; A semantic coding module, used for performing semantic coding on the representative sample data and the extended description of the representative sample data to obtain a semantic coding feature vector of the representative sample data and a semantic coding feature vector of the extended description of the representative sample data; A feature interaction fusion module, used for performing a compensatory interaction fusion on the semantic encoding feature vector representing the sample data and the semantic encoding feature vector representing the extended description of the sample data to obtain a semantic interaction compensation fusion representation vector representing the sample data; The annotation result generation module is used to generate the annotation result based on the semantic interaction compensation fusion representation vector of the representative sample data.

Images