Method and apparatus for intelligent response text classification, electronic device, and storage medium

Abstract

The application relates to a method and device for intelligent response text classification, electronic equipment and a storage medium. The method for intelligent response text classification comprises the following steps: S1, data processing, performing an ETL project on data converted through ASR, selecting call data, and then obtaining original corpus data sets required by a model; S2, text composition, performing standardization processing including punctuation symbol processing and stop word removal on the original corpus data sets, using one-hot to encode words, documents and entities, and forming an initial graph; S3, model construction, optimizing a graph convolutional neural network based on a feature projection network Fpnet, improving the graph convolutional network by adopting a multi-head edge node pooling method, forming a multi-head edge pooling graph convolutional network MHSP-GCN architecture, then replacing a common feature learning network C-net and a feature projection network P-net in the Fpnet with the MHSP-GCN, and through a feature projection network OPL, making original features Fp and shared features Fc perform orthogonal projection calculation to obtain purer classification features, so that the accuracy of the intelligent response text classification task is improved.

Description

Technical Field

[0001] This invention relates to the field of communications, and more specifically, to a method, apparatus, electronic device, and storage medium for intelligent response text classification based on a feature projection network optimized graph convolution model. Background Technology

[0002] Driven by the cloud transformation and digital transformation of telecommunications operators, the intelligent response service has developed rapidly, with the number of users reaching new highs. Intelligent response generates a large amount of data every day, making text classification and in-depth processing increasingly important.

[0003] Traditional text classification methods mainly fall into two categories: one is dictionary-based data classification, which compares data with an established dictionary database for classification; the other is machine learning-based data text classification, which often relies on manually designed features and suffers from sparsity and high dimensionality in text representation. Currently, classification models such as Naive Bayes and cosine similarity are used for ordinary text classification based on feature engineering. However, intelligent response text classification tasks differ from ordinary text classification tasks in several ways: First, the data involves a wide range of categories and is all natural language text converted by Automatic Speech Recognition (ASR) technology, mainly consisting of short text data with sparse, ambiguous semantics and a lack of context; second, the text generated by intelligent response exhibits class imbalance, with a large number of texts in a few scenarios, resulting in a long-tail phenomenon.

[0004] Currently, no effective solution has been found to address the aforementioned problems. Therefore, there is an urgent need to develop a method that can effectively improve the accuracy of intelligent response text classification. Summary of the Invention

[0005] The technical problem this invention aims to solve is the low accuracy of existing text classification methods for intelligent response text classification.

[0006] To solve the above-mentioned technical problems, according to one aspect of the present invention, a method for intelligent response text classification is provided, comprising the following steps: S1, data processing: based on intelligent response full-scenario corpus data, the data after ASR (Automatic Speech Recognition) transformation is subjected to ETL (Extraction, Transformation, Loading) engineering, including data cleaning and transformation, and data including call ID, call content, call role, call scenario category, call start time, and call end time are selected to obtain the original corpus dataset required by the model; S2, text graph construction: the original corpus dataset is subjected to standardization processing including punctuation processing and stop word removal, and one-hot encoding is used to encode words, documents, and entities to form an initial graph, wherein word co-occurrence is used to construct edges between words and between words and entities, and word frequency and document frequency of words are used to establish edges between word nodes and document nodes; words are represented as vertices, and word co-occurrence is represented as edges for graph construction, denoted as: G = (V, E), where V represents a node and E represents an edge; S3, Model Construction: Optimize the graph convolutional neural network based on the feature projection network (Fpnet), and improve the graph convolutional network using the multi-head edge pooling method to form the multi-head edge pooling graph convolutional network (MHSP-GCN) architecture. Then, replace the common feature learning network (C-net) and feature projection network (P-net) in Fpnet with MHSP-GCN, and use the feature projection network OPL to make the original features F p and shared features F c Orthogonal projection calculations yield purer classification features in vector space that more clearly indicate the category, thereby improving the accuracy of intelligent response text classification tasks.

[0007] According to an embodiment of the present invention, step S2 may include the following steps: S21, acquiring the original corpus data, and using an affine method to remove punctuation marks and stop words while maintaining data integrity, so as to make it exhibit better randomness; S22, using one-hot encoding to encode words, documents and entities to form an initial graph; S23, using word co-occurrence to construct edges between words and between words and entities; S24, using word frequency and document frequency of words to establish edges between word nodes and document nodes; S25, representing words as vertices and word co-occurrence as edges to construct the graph.

[0008] According to an embodiment of the present invention, step S3 may include the following steps: S31, for the graph data constructed in step S2, simultaneously input a Multi-Head Pooling Graph Convolutional Network (MHSP-GCN) P-net network, a Multi-Head Pooling Graph Convolutional Network (MHSP-GCN) C-net network, and a multi-head edge node pooling layer; S32, the P-net network and C-net network of the MHSP-GCN network pass through a GCN convolutional computation layer, then pass through a residual connection layer to calculate the residual, and perform weighted operations on the residual; S33, the P-net network passes through a second GCN convolutional computation layer and outputs the feature projection F. p The C-net network, after passing through a second GCN convolutional layer, outputs the common feature projection F. c S34. Merge feature projections F using feature projection method. p and common feature projection F c S35, Output the classification category after passing through the fully connected layer.

[0009] According to an embodiment of the present invention, in step S2, word co-occurrence in documents (document-word edges), word co-occurrence in the entire corpus (word-word edges), and word co-occurrence in the knowledge base (word-entity edges / document-entity edges) can be used to construct edges between nodes. The BM25 algorithm is used to calculate the word-document relevance for the weights of document-word and document-entity edges, and the calculation formula is as follows:

[0010] (1)

[0011] (2)

[0012] in, Total number of documents Indicates inclusion The number of documents, , The coordination factors are set to 2 and 0.75 respectively. For words Number of times it appears in the document; The length of the current document, i.e., the number of words; This represents the average length of the indexed documents.

[0013] According to an embodiment of the present invention, in step S2, in order to utilize global word co-occurrence information, the Pointwise Mutual Information (PMI) algorithm is used to calculate the weights between word-word nodes and entity-word nodes, and to calculate the word association metric. PMI is a word association metric that calculates the probability of two words appearing simultaneously in a text; the higher the probability, the closer their correlation and the higher their association degree. The PMI calculation formula is as follows:

[0014] (3)

[0015] (4)

[0016] (5)

[0017] in, Indicates the total number of sliding windows. Indicates words that contain both Harmony The number of sliding windows, This indicates the number of sliding windows containing only the word w.

[0018] According to an embodiment of the present invention, in step S2, when node and nodes When representing words or knowledge entities, the edge weights can be calculated using the PMI algorithm; when calculating the weights of edges formed by document-word or document-knowledge entity nodes, the BM25 algorithm can be used; when nodes form a self-loop, the weight can be set to 1.0; otherwise, the weight can be 0; thus, the edge weight expression is as follows:

[0019] (6)

[0020] The constructed graph is input into a 2-layer GCN network, and then... The classifier makes predictions using the cross-entropy function, expressed as follows:

[0021] (7)

[0022] (8)

[0023] (9)

[0024] in, Using 0 and 1 respectively yields the node representations for the two layers of GCN. Indicates the first The weight matrix of the layer, Indicates the first The weight matrix of the layer, Indicates the first Layer bias terms, Represents the adjacency matrix. for The degree matrix, Represents the eigenvector matrix, For the first Each node's real label This is the loss function.

[0025] According to an embodiment of the present invention, step S3 can construct a model by fusing a multi-headed edge pooling graph convolutional network and Fpnet, replacing the common feature learning network (C-net) and feature projection network (P-net) in Fpnet with MHSP-GCN, and using the feature projection network OPL to make the original features F p and shared features F c Orthogonal projection calculations yield purer classification features. ,feature The category orientation is more explicit in the vector space; the model network consists of two networks: a multi-sided pooling graph convolutional network feature projection network P-net (MHSP-GCN P-net) and a multi-sided pooling graph convolutional network common feature extraction network C-net (MHSP-GCN C-net); the two networks are structurally identical but do not share parameters; the output of MHSP-GCN C-net after adding a GRL inversion layer is the same as that of MHSP-GCN P-net, as shown in Equations 10 and 11. The output layers of the models both use the Softmax normalized activation function, as shown in Equations 12 and 13. The cross-entropy loss function is used for calculation in both networks.

[0026] (10)

[0027] (11)

[0028] (12)

[0029] (13)

[0030] During backpropagation, the network parameters of MHSP-GCN P-net and MHSP-GCN C-net are not shared. Backpropagation only updates the MHSP-GCN C-net network parameters. Backpropagation only updates the parameters of the MHSP-GCN P-net network; although the MHSP-GCN C-net also uses softmax and cross-entropy loss functions, the gradients in the GRL layer of the MHSP-GCN C-net module are reversed during backpropagation, therefore... The value will gradually decrease; Computation and backpropagation are only used to allow the neural network to obtain common features; in the MHSP-GCN P-net module To predict the total classification loss value for the entire model at the end, The value is the final prediction output of the MHSP-GCN network optimized by the entire feature projection.

[0031] According to a second aspect of the present invention, an apparatus for intelligent response text classification is provided, comprising: a data processing module, which is used to perform an ETL process, including data cleaning and transformation, on the data after ASR conversion based on intelligent response full-scenario corpus data, selecting data including call ID, call content, call role, call scenario category, call start time, and call end time, thereby obtaining the original corpus dataset required by the model; and a text graphing module, which performs standardization processing on the original corpus dataset, including punctuation processing and stop word removal, and uses one-hot encoding to encode words, documents, and entities to form an initial graph. The model constructs edges between words and between words and entities using word co-occurrence, and establishes edges between word nodes and document nodes using word frequency and document frequency. Words are represented as vertices, and word co-occurrence as edges for graph construction. The model construction module optimizes the graph convolutional neural network based on the Feature Projection Network (Fpnet), improves the graph convolutional network using multi-head edge pooling, forming a Multi-Head Edge Pooling Graph Convolutional Network (MHSP-GCN) architecture. Then, MHSP-GCN replaces the common feature learning network (C-net) and feature projection network (P-net) in Fpnet, and the original feature Fpnet is processed by the feature projection network OPL. p and shared features F c Orthogonal projection calculations yield purer classification features in vector space that more clearly indicate the category, thereby improving the accuracy of intelligent response text classification tasks.

[0032] According to a third aspect of the present invention, an electronic device is provided, comprising: a memory, a processor, and an intelligent response text classification program stored in the memory and executable on the processor, wherein the intelligent response text classification program, when executed by the processor, implements the steps of the intelligent response text classification method described above.

[0033] According to a fourth aspect of the present invention, a computer storage medium is provided, wherein the computer storage medium stores an intelligent response text classification program, which, when executed by a processor, implements the steps of the intelligent response text classification method described above.

[0034] Compared with the prior art, the technical solution provided by the embodiments of the present invention can achieve at least the following beneficial effects:

[0035] This invention employs a text classification method based on a feature projection network-optimized graph convolutional model. Targeting the features of intelligent response text corpora, this method first improves the graph convolutional neural network by using multi-head edge pooling to enhance important nodes, thereby extracting global and important textual information features. Then, it leverages a feature projection network (FP-net) with text enhancement capabilities to improve representation learning, effectively enhancing text classification results and performance. This significantly improves the accuracy of intelligent response text classification, enhances service quality, and unlocks the potential value of data. It possesses broad application scenarios and commercial value.

[0036] This invention enhances representation learning by not discarding unimportant nodes while selecting important nodes, defining the MHSP-GCN structure. Rich links effectively associate nodes (including sparse ones), thereby enabling efficient representation of these nodes and obtaining global information.

[0037] The MHSP-GCN network proposed in this invention introduces multi-head edge pooling to enhance the representation learning of important nodes. These selected and enhanced nodes contain more distinctive features, which can lead to more accurate classification.

[0038] The MHSP-GCN proposed in this invention is an innovative network structure based on GCN, which has a stronger ability to cover all data. When applied to short text classification, it can provide a certain degree of attention to long-tail (sparse) words. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0040] Figure 1 This is a flowchart illustrating an intelligent response service modeling method according to an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram illustrating the text structure of an intelligent response corpus according to an embodiment of the present invention.

[0042] Figure 3 This is a flowchart illustrating the MHSP-GCN based on feature projection according to an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram illustrating the feature projection Fpnet network structure according to an embodiment of the present invention.

[0044] Figure 5This is a schematic diagram illustrating the structure of the Multi-Headed Edge Pooling Graph Convolutional Network (MHSP-GCN) according to an embodiment of the present invention.

[0045] Figure 6 This is a schematic diagram illustrating the multi-head edge node pooling vector representation according to an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 described 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.

[0047] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a limitation of quantity, but rather indicate the presence of at least one.

[0048] Figure 1 This is a flowchart illustrating an intelligent response service modeling method according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the text structure of an intelligent response corpus according to an embodiment of the present invention.

[0049] like Figure 1 As shown, the method for intelligent response text classification includes the following steps:

[0050] S1. Data Processing: Based on the full-scenario corpus data of intelligent response, the data after ASR (Automatic Speech Recognition) transformation is processed by ETL (Extraction, Transformation, Loading) engineering, which includes data cleaning and transformation. Data including call ID, call content, call role, call scenario category, call start time, and call end time are selected to obtain the original corpus dataset required by the model.

[0051] S2. Text Graph Construction: The original corpus dataset undergoes standardization processing, including punctuation removal and stop word removal. One-hot encoding is used to encode words, documents, and entities, forming an initial graph. Word co-occurrence is used to construct edges between words and between words and entities, while word frequency and document frequency are used to establish edges between word nodes and document nodes. Words are represented as vertices, and word co-occurrence as edges for graph construction, denoted as: G = (V, E), where V represents a node and E represents an edge, as shown below. Figure 2 As shown in the figure, W represents word nodes, T represents document nodes, and E represents entity edge nodes.

[0052] S3. Model Construction: The graph convolutional neural network (PCNN) is optimized based on the Feature Projection Network (Fpnet). A multi-edge pooling method is used to improve the PCNN, forming a Multi-Edge Pooled Graph Convolutional Network (MHSP-GCN) architecture. Then, MHSP-GCN replaces the common feature learning network (C-net) and the feature projection network (P-net) in Fpnet. The original feature Fpnet is then processed by the feature projection network (OPL). p and shared features F c Orthogonal projection calculations yield purer classification features. ,feature The category is more clearly defined in the vector space, thereby improving the accuracy of intelligent response text classification tasks.

[0053] This invention employs a text classification method based on a feature projection network-optimized graph convolutional model. Targeting the features of intelligent response text corpora, this method first improves the graph convolutional neural network by using multi-head edge pooling to enhance important nodes, thereby extracting global and important textual information features. Then, it leverages a feature projection network (FP-net) with text enhancement capabilities to improve representation learning, effectively enhancing text classification results and performance. This significantly improves the accuracy of intelligent response text classification, enhances service quality, and unlocks the potential value of data. It possesses broad application scenarios and commercial value.

[0054] According to one or more embodiments of the present invention, step S2 includes the following steps: S21, acquiring the original corpus data, and using an affine method to remove punctuation marks and stop words while maintaining data integrity, so as to make it exhibit better randomness; S22, using one-hot encoding to encode words, documents and entities to form an initial graph; S23, using word co-occurrence to construct edges between words and between words and entities; S24, using word frequency and document frequency of words to establish edges between word nodes and document nodes; S25, representing words as vertices and word co-occurrence as edges to construct the graph.

[0055] Figure 3 This is a flowchart illustrating the MHSP-GCN based on feature projection according to an embodiment of the present invention.

[0056] like Figure 3 As shown, step S3 includes the following steps: S31, for the graph data constructed in step S2, simultaneously input a Multi-Head Pooling Graph Convolutional Network (MHSP-GCN) P-net network, a Multi-Head Pooling Graph Convolutional Network (MHSP-GCN) C-net network, and a multi-head edge node pooling layer; S32, the P-net and C-net networks of the MHSP-GCN network pass through a GCN convolutional computation layer, then through a residual connection layer to calculate the residuals, and perform weighted operations on the residuals; S33, the P-net network passes through a second GCN convolutional computation layer and outputs the feature projection F. p The C-net network, after passing through a second GCN convolutional layer, outputs the common feature projection F. c S34. Merge feature projections F using feature projection method. p and common feature projection F c S35, Output the classification category after passing through the fully connected layer.

[0057] The MHSP-GCN network proposed in this invention introduces multi-head edge pooling to enhance the representation learning of important nodes. These selected and enhanced nodes contain more distinctive features, which can lead to more accurate classification.

[0058] According to one or more embodiments of the present invention, in step S2, word co-occurrence in documents (document-word edges), word co-occurrence in the entire corpus (word-word edges), and word co-occurrence in the knowledge base (word-entity edges / document-entity edges) are used to construct edges between nodes. The BM25 algorithm is used to calculate the word-document relevance for the weights of document-word and document-entity edges, and the calculation formula is as follows:

[0059] (1)

[0060] (2)

[0061] in, Total number of documents Indicates inclusion The number of documents, , The coordination factors are set to 2 and 0.75 respectively. For words Number of times it appears in the document; The length of the current document, i.e., the number of words; This represents the average length of the indexed documents.

[0062] According to one or more embodiments of the present invention, in step S2, in order to utilize global word co-occurrence information, the weights between word-word nodes and entity-word nodes are calculated using the Pointwise Mutual Information (PMI) algorithm to calculate the word association metric. PMI is a word association metric that calculates the probability of two words appearing simultaneously in a text; the higher the probability, the closer their correlation and the higher their association degree. The PMI calculation formula is as follows:

[0063] (3)

[0064] (4)

[0065] (5)

[0066] in, Indicates the total number of sliding windows. Indicates words that contain both Harmony The number of sliding windows, This indicates the number of sliding windows containing only the word w.

[0067] According to one or more embodiments of the present invention, in step S2, when node and nodes When representing words or knowledge entities, the edge weights are calculated using the PMI algorithm; when calculating the weights of edges formed by document-word or document-knowledge entity nodes, the BM25 algorithm is used; when nodes form a self-loop, the weight is set to 1.0; otherwise, the weight is 0; thus, the edge weight expression is as follows:

[0068] (6)

[0069] The constructed graph is input into a 2-layer GCN network, and then... The classifier makes predictions using the cross-entropy function, expressed as follows:

[0070] (7)

[0071] (8)

[0072] (9)

[0073] in, Using 0 and 1 respectively yields the node representations for the two layers of GCN. Indicates the first The weight matrix of the layer, Indicates the first The weight matrix of the layer, Indicates the first Layer bias terms, Represents the adjacency matrix. for The degree matrix, Represents the eigenvector matrix, For the first Each node's real label This is the loss function.

[0074] Figure 4 This is a schematic diagram illustrating the feature projection (Fpnet) network structure according to an embodiment of the present invention.

[0075] like Figure 4 As shown, the Feature Projection Network (Fpnet) is a text augmentation neural network structure that embeds a Gradient Reversal Layer (GRL) to classify common features across multiple text categories. By leveraging feature projection to improve representation learning, it effectively enhances text classification performance. Fpnet consists of two sub-networks: the Common Feature Learning Network (C-net) on the right and the Projection Network (P-net) on the left. Its main focus is on using both the P-net and C-net networks to train different tasks. Although the two networks have the same structure, they do not share parameters. The C-net incorporates a Gradient Reversal Layer (GRL), while the P-net incorporates a Projection Layer (OPL). By projecting the learned information vectors of the input text into a more discriminative semantic space, the influence of common features is eliminated. The Fpnet network results are shown below. Figure 4 As shown.

[0076] Figure 5 This is a schematic diagram illustrating the structure of a Multi-Headed Edge Pooling Graph Convolutional Network (MHSP-GCN) according to an embodiment of the present invention.

[0077] like Figure 5 As shown, pooling reduces the parameter size to generate smaller representations through node selection (similar to downsampling methods). However, we propose to simultaneously select important nodes to enhance representation learning without discarding unimportant nodes, defining a Multi-Headed Pooling Graph Convolutional Network (MHSP-GCN) structure, as follows: Figure 5As shown, the MHSP-GCN structure employs GCN convolutional computation layers to extract features. From top to bottom, the first GCN convolutional computation layer aggregates information from nodes directly adjacent to the center node (order 1). The second GCN convolutional computation layer aggregates information from nodes adjacent to the center node (order 2). Multi-edge pooling is our proposed pooling layer used to evaluate and select important nodes. Information from unselected nodes may be lost during pooling; residual connection layers are used to recover their information, and weighting operations are used to weight nodes based on the attention scores calculated by multi-edge pooling.

[0078] This invention enhances representation learning by not discarding unimportant nodes while selecting important nodes, defining the MHSP-GCN structure. Rich links effectively associate nodes (including sparse ones), thereby enabling efficient representation of these nodes and obtaining global information.

[0079] Figure 6 This is a schematic diagram illustrating the multi-head edge node pooling vector representation according to an embodiment of the present invention.

[0080] The purpose of multi-edge node pooling is to correctly select important nodes and reduce the influence of unimportant nodes. A self-attention mechanism is introduced to score nodes. Node scores are calculated using learnable projection vectors, represented as follows: Figure 6 As shown.

[0081] According to one or more embodiments of the present invention, step S3 constructs a model by fusing a multi-headed edge pooling graph convolutional network and Fpnet, replacing the common feature learning network (C-net) and feature projection network (P-net) in Fpnet with MHSP-GCN, and using the feature projection network OPL to make the original features F p and shared features F c Orthogonal projection calculations yield purer classification features. ,feature The category orientation is more explicit in the vector space; the model network consists of two networks: a multi-sided pooling graph convolutional network feature projection network P-net (MHSP-GCN P-net) and a multi-sided pooling graph convolutional network common feature extraction network C-net (MHSP-GCN C-net); the two networks are structurally identical but do not share parameters; the output of MHSP-GCN C-net after adding a GRL inversion layer is the same as that of MHSP-GCN P-net, as shown in Equations 10 and 11. The output layers of the models both use the Softmax normalized activation function, as shown in Equations 12 and 13. The cross-entropy loss function is used for calculation in both networks.

[0082] (10)

[0083] (11)

[0084] (12)

[0085] (13)

[0086] During backpropagation, the network parameters of MHSP-GCN P-net and MHSP-GCN C-net are not shared. Backpropagation only updates the MHSP-GCN C-net network parameters. Backpropagation only updates the parameters of the MHSP-GCN P-net network; although the MHSP-GCN C-net also uses softmax and cross-entropy loss functions, the gradients in the GRL layer of the MHSP-GCN C-net module are reversed during backpropagation, therefore... The value will gradually decrease; Computation and backpropagation are only used to allow the neural network to obtain common features; in the MHSP-GCN P-net module To predict the total classification loss value for the entire model at the end, The value is the final prediction output of the MHSP-GCN network optimized by the entire feature projection.

[0087] The MHSP-GCN proposed in this invention is an innovative network structure based on GCN, which has a stronger ability to cover all data. When applied to short text classification, it can provide a certain degree of attention to long-tail (sparse) words.

[0088] According to a second aspect of the present invention, an apparatus for intelligent response text classification is provided, comprising: a data processing module, a text mapping module, and a model building module.

[0089] The data processing module is used to perform ETL engineering on the data after ASR transformation based on the full-scenario intelligent response corpus data. It selects data including call ID, call content, call role, call scenario category, call start time, and call end time to obtain the original corpus dataset required by the model.

[0090] The text graph construction module performs standardization processing on the original corpus dataset, including punctuation processing and stop word removal. It uses one-hot encoding to encode words, documents, and entities to form an initial graph. In this process, word co-occurrence is used to construct edges between words and between words and entities, and word frequency and document frequency of words are used to establish edges between word nodes and document nodes. Words are represented as vertices, and word co-occurrence is represented as edges for graph construction.

[0091] The model building module optimizes the graph convolutional neural network based on the Feature Projection Network (Fpnet), and improves the graph convolutional network using a multi-head edge pooling method to form a Multi-Head Edge Pooling Graph Convolutional Network (MHSP-GCN) architecture. Then, MHSP-GCN replaces the common feature learning network (C-net) and the feature projection network (P-net) in Fpnet. The feature projection network (OPL) then uses MHSP-GCN to improve the performance of the original feature learning network (Fpnet). p and shared features F c Orthogonal projection calculations yield purer classification features. ,feature The category is more clearly defined in the vector space, thereby improving the accuracy of intelligent response text classification tasks.

[0092] The application process of this technical solution in intelligent response is as follows:

[0093] User B applies for a communication assistant service;

[0094] When user A misses a call from user B, the call will be answered by an intelligent answering robot.

[0095] Caller A communicated with the intelligent response robot about the business scenario of picking up a package;

[0096] The dialogue process is provided by key components such as intelligent response ASR (Automatic Speech Recognition), NLP (Natural Language Processing), TTS (Text To Speech), and voice gateway;

[0097] During the call, the MHSP-GCN model optimized by the feature projection of this technical solution performs scene recognition, and the response process points to the corresponding dialogue process;

[0098] The audio recording, response text, and call summary text are pushed to user B via SMS and WeChat message.

[0099] User B handles subsequent call processing via push notifications.

[0100] According to another aspect of the present invention, an apparatus for intelligent response text classification is provided, comprising: a memory, a processor, and an intelligent response text classification program stored in the memory and executable on the processor, wherein the intelligent response text classification program, when executed by the processor, implements the steps of the intelligent response text classification method described above.

[0101] The present invention also provides a computer storage medium.

[0102] The computer storage medium stores an intelligent response text classification program, which, when executed by the processor, implements the steps of the intelligent response text classification method described above.

[0103] The method implemented when the intelligent response text classification program running on the processor is executed can be referred to in various embodiments of the intelligent response text classification method of the present invention, and will not be repeated here.

[0104] The present invention also provides a computer program product.

[0105] The computer program product of the present invention includes an intelligent response text classification program, which, when executed by a processor, implements the steps of the intelligent response text classification method as described above.

[0106] The method implemented when the intelligent response text classification program running on the processor is executed can be referred to in various embodiments of the intelligent response text classification method of the present invention, and will not be repeated here.

[0107] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0108] The above description is merely an exemplary embodiment of the present invention and is not intended to limit the scope of protection of the present invention, which is determined by the appended claims.

Claims

1. A method for intelligent response text classification, comprising the following steps: S1. Data processing: Based on the full-scenario corpus data of intelligent response, the data after ASR conversion is subjected to ETL engineering including data cleaning and transformation. Data including call ID, call content, call role, call scenario category, call start time, and call end time are selected to obtain the original corpus dataset required by the model. S2. Text graph construction: The original corpus dataset undergoes standardization processing, including punctuation removal and stop word removal. One-hot encoding is used to encode words, documents, and entities to form an initial graph. in, We construct edges between words and between words and entities using word co-occurrence, and establish edges between word nodes and document nodes using word frequency and word document frequency; words are represented as vertices and word co-occurrence as edges to construct the graph. S3. Model Construction: Based on the feature projection network Fpnet, the graph convolutional neural network is optimized. A multi-head edge pooling method is used to improve the graph convolutional network, forming the MHSP-GCN architecture. Then, MHSP-GCN replaces the common feature learning network C-net and the feature projection network P-net in Fpnet. The original feature Fpnet is then processed through the feature projection network OPL. p and shared features F c Orthogonal projection calculations yield purer classification features in the vector space that more clearly indicate the category, thereby improving the accuracy of intelligent response text classification tasks. Step S3 includes the following steps: S31. For the graph data constructed in step S2, simultaneously input the multi-head pooling graph convolutional network MHSP-GCNP-net, the multi-head pooling graph convolutional network MHSP-GCNC-net, and the multi-head edge node pooling layer. The S32 and MHSP-GCN networks' P-net and C-net networks pass through a GCN convolutional computation layer, then through a residual connection layer to compute residuals, and finally perform weighted operations on the residuals. The S33 and P-net networks, after passing through a second GCN convolutional layer, output feature projection F. p The C-net network, after passing through a second GCN convolutional layer, outputs the common feature projection F. c ; S34. Merge feature projections F using feature projection method. p and common feature projection F c ; S35, output the classification category after passing through the fully connected layer.

2. The method as described in claim 1, wherein, Step S2 includes the following steps: S21. Obtain the original corpus data, and use the affine method to remove punctuation marks and stop words while maintaining data integrity, so that it can exhibit better random properties; S22. Use one-hot encoding to encode words, documents, and entities to form an initial graph; S23. Construct edges between words and between words and entities using word co-occurrence; S24. Use word frequency and document frequency of words to establish edges between word nodes and document nodes; S25. Construct a graph by representing words as vertices and word co-occurrences as edges.

3. The method as described in claim 1, wherein, Step S2 uses word co-occurrence of document-word edges in documents, word co-occurrence of words in the entire corpus, and word co-occurrence of words-entity edges / document-entity edges in the knowledge base to construct edges between nodes. Specifically, the BM25 algorithm is used to calculate the word-document relevance for the weights of document-word and document-entity edges. The calculation formula is as follows: （1） （2） in, Total number of documents Indicates inclusion The number of documents, , The coordination factors are set to 2 and 0.75 respectively. For words Number of times it appears in the document; The length of the current document, i.e., the number of words; This represents the average length of the indexed documents.

4. The method of claim 3, wherein, In step S2, in order to utilize global word co-occurrence information, the point mutual information algorithm (PMI) is used to calculate the weights between word-word nodes and entity-word nodes, and to calculate the word association metric. PMI is a word association metric that measures the probability of two words appearing together in a text. The higher the probability, the stronger the correlation and the higher the degree of association. The PMI calculation formula is as follows: （3） （4） （5） in, Indicates the total number of sliding windows. Indicates words that contain both Harmony The number of sliding windows, This indicates the number of sliding windows containing only the word w.

5. The method of claim 4, wherein, In step S2, when node and nodes When representing words or knowledge entities, the edge weights are calculated using the PMI algorithm; when calculating the weights of edges formed by document-word or document-knowledge entity nodes, the BM25 algorithm is used; when nodes form a self-loop, the weight is set to 1.0; otherwise, the weight is 0; thus, the edge weight expression is as follows: （6） The constructed graph is input into a 2-layer GCN network, and then... The classifier makes predictions using the cross-entropy function, expressed as follows: （7） （8） （9） in, Using 0 and 1 respectively yields the two-layer node representation of GCN. Indicates the first The weight matrix of the layer, Indicates the first The weight matrix of the layer, Indicates the first Layer bias terms, Represents the adjacency matrix. for The degree matrix, Represents the eigenvector matrix, For the first Each node's real label This is the loss function.

6. The method of claim 5, wherein, Step S3 constructs a model by fusing a multi-headed edge pooling graph convolutional network and Fpnet. MHSP-GCN replaces the common feature learning network C-net and the feature projection network P-net in Fpnet. The feature projection network OPL then transforms the original features into Fpnet. p and shared features F c Orthogonal projection calculations yield purer classification features. ,feature The category is more clearly defined in the vector space; The model network consists of two networks: MHSP-GCN P-net (Multi-Polygonal Graph Convolutional Network Feature Projection Network) and MHSP-GCN C-net (Multi-Polygonal Graph Convolutional Network Common Feature Extraction Network). These two networks are structurally identical but do not share parameters. Adding a GRL inversion layer to MHSP-GCN C-net produces the same output as MHSP-GCN P-net, as shown in Equations 10 and 11. Both output layers use the Softmax normalized activation function, as shown in Equations 12 and 13. The cross-entropy loss function is used for calculation in both networks. （10） （11） （12） （13） During backpropagation, the network parameters of MHSP-GCN P-net and MHSP-GCN C-net are not shared. Backpropagation only updates the MHSP-GCN C-net network parameters. Backpropagation only updates the parameters of the MHSP-GCN P-net network; although the MHSP-GCN C-net also uses softmax and cross-entropy loss functions, the gradients in the GRL layer of the MHSP-GCN C-net module are reversed during backpropagation, therefore... The value will gradually decrease; Computation and backpropagation are only used to allow the neural network to obtain common features; in the MHSP-GCN P-net module To predict the total classification loss value for the entire model at the end, The value is the final prediction output of the MHSP-GCN network optimized by the entire feature projection.

7. An apparatus for intelligent response text classification, comprising: The data processing module is used to perform ETL engineering, including data cleaning and transformation, on the data after ASR conversion based on the full-scenario intelligent response corpus data. It selects data including call ID, call content, call role, call scenario category, call start time, and call end time to obtain the original corpus dataset required by the model. The text graphing module performs standardization processing on the original corpus dataset, including punctuation processing and stop word removal. It then uses one-hot encoding to encode words, documents, and entities to form an initial graph. Specifically, word co-occurrence is used to construct edges between words and between words and entities, and word frequency and document frequency of words are used to establish edges between word nodes and document nodes; words are represented as vertices and word co-occurrence is represented as edges to construct the graph; The model building module is based on the Feature Projection Network (Fpnet) optimized graph convolutional neural network. It improves the graph convolutional network using a multi-edge pooling method, forming the Multi-Edge Pooling Graph Convolutional Network (MHSP-GCN) architecture. Then, MHSP-GCN replaces the common feature learning network (C-net) and the feature projection network (P-net) in Fpnet. Finally, the feature projection network (OPL) is used to optimize the original feature learning network (Fpnet). p and shared features F c Orthogonal projection calculations yield purer classification features in the vector space that more clearly indicate the category, thereby improving the accuracy of intelligent response text classification tasks. Specifically, for the graph data constructed in the text graph construction module, the model construction module simultaneously inputs a multi-head pooling graph convolutional network (MHSP-GCNP-net), a multi-head pooling graph convolutional network (MHSP-GCNC-net), and a multi-head edge node pooling layer. The P-net and C-net networks of the MHSP-GCN network undergo a GCN convolutional computation layer, followed by a residual connection layer to calculate residuals, and then weighted processing of the residuals. The P-net network then passes through a second GCN convolutional computation layer, outputting the feature projection F. p The C-net network, after passing through a second GCN convolutional layer, outputs the common feature projection F. c The feature projection F is fused using a feature projection method. p and common feature projection F c The fully connected layer outputs the classification category.

8. An electronic device, comprising: The system includes a memory, a processor, and an intelligent response text classification program stored in the memory and executable on the processor, wherein the intelligent response text classification program, when executed by the processor, implements the steps of the intelligent response text classification method as described in any one of claims 1 to 6.

9. A computer storage medium, wherein, The computer storage medium stores an intelligent response text classification program, which, when executed by a processor, implements the steps of the intelligent response text classification method as described in any one of claims 1 to 6.

Images