Short text clustering analysis method based on inductive learning

Abstract

The invention discloses a short text clustering analysis method based on inductive learning. The method comprises the following steps: preprocessing a short text, and mapping the short text to a graph network based on word co-occurrence in a sliding window; reserving global information of vertexes in the graph network, and obtaining vertex-community distribution as attribute features of the vertexes by executing feature extraction; and for the words in the newly generated text, neighbor information is aggregated by using the graph neural network, new vertex representation is learned, and the neighbor information of the words in the newly generated text is aggregated through recursion. According to the invention, the accuracy and robustness of short text representation are improved.

Description

Technical Field

[0001] This invention relates to the field of data mining and information retrieval technology, and more specifically, to a short text clustering analysis method based on inductive learning. Background Technology

[0002] Short text clustering has garnered significant attention due to the explosive growth of short texts in real-world applications such as tweets and Google News. Aiming to group semantically similar documents without supervision or manual label assignment, short text clustering has proven beneficial in various applications, including topic discovery and document summarization. One of the main challenges of short text clustering compared to texts with rich context is the sparsity problem of short documents covering only a few words. To address this, some works have resorted to external knowledge to enrich short text representations. For example, leveraging Wikipedia to expand short texts or incorporating ontology to enrich the text. Since these methods are based on supervised learning, their performance is affected by the quality of the external resources.

[0003] Currently, short text representation learning performed in unsupervised environments can be categorized into graph probabilistic models and neural network embeddings. Graph probabilistic models (e.g., LDA) have attracted widespread attention due to their excellent interpretability. For example, GSDMM, specifically designed for short text clustering, employs a Dirichlet distribution and follows the assumption that all words in a document are related to the same topic. BTM proposes an explicit (bi-term) approach to model word co-occurrence patterns in short texts. Subsequently, variants of probabilistic models (such as GSDPMM and DP-BMM) have been proposed for different applications of short texts. However, these models are all limited by sparse word co-occurrence patterns and are not suitable for inductive short text clustering where few words have been previously accessed during training. Furthermore, these models are statistical models and lack consideration for word order information, thus affecting the accuracy of representation learning.

[0004] Recently, with the rise of neural network technology, Word2vec initially combined a two-layer neural network to learn word embeddings. Subsequent variants include sentence-based embeddings (such as SIF and STV) and paragraph embeddings (such as PV and PV2). While these embedding methods are context-aware and perform well in mapping training data to a low-dimensional space, they all fail to support transductive models for document clustering tasks. Furthermore, most models are not designed specifically for short texts, making it difficult to distinguish clusters of short documents with few words in the training set. Simultaneously, these models ignore word information propagation guided by text structure and are computationally inefficient because there is no parameter sharing between words during training.

[0005] Analysis reveals that while many other deep clustering methods have demonstrated feasibility in graph mining, their applicability to inductive short text clustering remains challenging. Generally, a major limitation of existing short text clustering methods lies in their inability to handle the sparsity of words in newly generated documents. To date, numerous short text representation learning methods have been proposed, primarily focusing on learning compact low-dimensional embeddings to alleviate the sparse representation problem. For example, probabilistic graphical models such as GSDMM and BTM aim to learn document topic and word topic representations (or distributions) by leveraging the latent topic structure of short texts. More recently, embedding learning methods, including Word2vec and SIF, have focused on maximizing the probability of word co-occurrence within a sliding window to learn word representations. However, both probabilistic graphical models designed primarily for specific clustering tasks and embedding learning models employing classic clustering methods like K-means for document clustering require a generalized transduction model present in all documents of the corpus during training. While these models can map accessed data (i.e., training data) well to a low-dimensional space, they cannot generalize to newly generated documents, thus hindering their ability to cluster them.

[0006] Essentially, the main limitation of existing STC techniques is their inability to handle the sparsity of words in newly generated documents well. Short texts containing a small number of words, compared to long documents, make traditional document representations sparse, such as Term Frequency-Inverse Document Frequency (TF-IDF) or bag-of-words. Furthermore, these sparse representations can lead to poor clustering performance, which inherently relies on calculating the distance between representations. However, current models either focus solely on learning representations from the local co-occurrence of words within a sliding window, such as Skip-Gram, or simply rely on statistical models, thus ignoring the propagation of word information guided by text structure, which could potentially improve clustering performance by generating new document representations. Moreover, existing models often suffer from computational inefficiency because words in the encoder do not share parameters; each word is presented as a unique vector during training.

[0007] In recent years, Graph Neural Networks (GNNs), such as GCN and GraphSAGE, have shown groundbreaking performance on many graph mining tasks, while relatively little work has been done on document clustering. One advantage of GNNs is that the vertices of a graph can share the learned parameters in the neural network, thus allowing for easy aggregation of neighbor information for each vertex. However, a key difference between graphs and text lies in their different data structures, particularly the lack of explicit edges between words in a document. Furthermore, incorporating GNNs into document representation learning for short text clustering remains challenging. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a short text clustering analysis method based on inductive learning. This method includes the following steps: The short text is preprocessed and mapped to a graph network based on word co-occurrence within a sliding window; The global information of vertices in the graph network is preserved, and the vertex-community distribution is obtained as the attribute feature of the vertex by performing feature extraction. For words in the newly generated document, a graph neural network is used to aggregate neighbor information and learn new vertex representations, where the neighbor information of words in the newly generated text is aggregated recursively.

[0009] Compared with existing technologies, the advantages of this invention are that it designs a short text clustering analysis model (IDRL) ​​based on inductive learning. By mapping the short text structure to a graph network and recursively aggregating the neighbor information of words in the newly generated text, it embeds a limited number of previously learned words to reconstruct the representation of the newly generated short text. This solves the problems of sparsity and poor clustering performance in word co-occurrence patterns presented by traditional methods in short texts, as well as the problem that learning compact low-dimensional embeddings cannot well handle the representation of unseen documents with little learned information in short text representation learning methods.

[0010] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0012] Figure 1 This is a flowchart of a short text clustering analysis method based on inductive learning according to an embodiment of the present invention; Figure 2 This is an overview diagram of a short text clustering analysis model based on inductive learning according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the clustering accuracy of clustering results according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the normalized mutual information performance of clustering results according to an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the impact of dimensional settings on clustering performance, Micro-F1, and Macro-F1, respectively, according to an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the clustering performance of the number of clusters on four datasets according to an embodiment of the present invention. Detailed Implementation

[0013] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0014] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0015] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0016] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0017] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0018] This invention starts with the original text and naturally constructs a graph based on word co-occurrence within a sliding window, such as establishing edges between words. These edges can then be fed into a neural network for further processing. This invention follows the GraphSAGE aggregation method, generating word representations by aggregating their neighbor information. In this way, the correlation between local connectivity learned from word co-occurrence and the global structure obtained from multi-layer aggregation can be uniformly preserved, and the sparsity problem in short text representation learning can be alleviated. Finally, for the output layer, a negative sampling objective function is used to encourage target words to be close to their neighbors while simultaneously moving away from their negative samples in the embedding space.

[0019] Overall, since the goal of this invention is to learn the representation of documents composed of words, the corpus is represented as... ,in Represents a set of documents. A vocabulary representing the document. For each word... I hope to learn low-dimensional embedding. This preserves the relevance of words, among which, This indicates the embedded dimension of the exhibition space.

[0020] Next, to generate a representation of the newly generated document, it is necessary to aggregate the neighbor information of previously unlearned words. Therefore, the short text is mapped onto a graph network to easily mine the adjacency structure of words. For example, the graph network can be represented as... ,in It is a set of vertices. The same notation is used here. To represent vertices and words, because they reference the same object, also use... This represents the edge set generated based on the co-occurrence of words within a sliding window.

[0021] See Figure 1 As shown, the short text clustering analysis method based on inductive learning provided includes the following steps: Step S1: Preprocess the short text and generate a unit graph based on word co-occurrence within the sliding window.

[0022] Combination Figure 2 As shown in the upper part, the input short text is preprocessed before constructing the graph. A paper title can be used as an example, and stop words and stems are removed using the standard NLTK1 tool. Then, all words are presented as vertices, and edges between them are constructed within a sliding window. For example, for the target word 'decis', when the window size is 2, the edges between 'decis' and 'gener,markov,process,dynam' can be obtained accordingly. It is worth noting that because the inductive learning-based short text clustering analysis model (IDRL) ​​designed in this invention can learn information propagation guided by the graph structure, it is not necessary to use a large window size and long distances to capture word relationships.

[0023] Step S2: Obtain the vertex community distribution to complete vertex attributes.

[0024] To preserve global information about the vertices in the graph, nonnegative matrix factorization (NMF) is performed to obtain vertex community distributions as vertex attribute features. Let As a symmetric adjacency matrix of a network, NMF can be solved: (1) Where X represents the vertex attribute characteristics (i.e., the attribute characteristics of the vertex). The vertex community distribution represents the encoding of a global understanding of the network state, and the vertex count represents the number of vertices. Indicates the number of potential communities. It is the Frobenius norm of the matrix. It is a harmonic factor that balances the two components. Generally speaking, compared to simply using X as the final vertex representation, this vertex attribute completion method of the present invention can preserve vertex communities from the global network structure and vertex-vertex correlations from local connectivity in the embedding space, thereby helping to alleviate the sparsity problem in short text representation learning.

[0025] It should be noted that the IDRL framework is highly versatile, and other feature extraction methods can be used to replace NMF for vertex attribute feature learning. This invention prefers to use NMF because of its universality and fast training (time complexity is linearly related to vertex size).

[0026] Step S3: For words in the newly generated text, aggregate neighbor information through representation learning.

[0027] For inductive learning, the designed model needs to allow efficient generation of embeddings (such as words representing the same object) for newly generated vertices. GCN first introduces an efficient variant of a convolutional neural network that can operate directly on graphs. Improved versions of GCN can be used for inductive learning on large-scale networks. In this embodiment, a local spectral convolutional aggregator is used for representation learning, and its formula is defined as follows: (2) Where MEAN represents the average of the elements of the vector. It is the weight matrix of the layer. , It is the embedding vector in the layer. , It is the sigmoid function, and Represents the neighbors of a vertex , It is a context vertex. It should be noted that it is possible to let... , where x represents the vertex community distribution learned from Equation (1). In summary, it can be seen from Equation (2) that the MEAN aggregator can aggregate the neighbor representations of vertices, including newly generated vertices for inductive learning. Then, to learn vertex representations in an unsupervised setting, negative sampling (NS) is used to the output representation, which favors the target vertex being close to its neighbors while being far from its negative samples in the embedding space. For example, the objective function is set as: (3) in, It is the target vertex. It is its context vertex. It is the sigmoid function, that is , Indicates a negative sampling distribution. It is from The negative samples. K represents the number of negative samples used for estimation, and , , It is a representation aggregated from features involving its local neighbors. The objective is to promote nearby vertices having similar embeddings while being different from their negative vertices. Furthermore, the distribution of negative samples... The formula is as follows: (4) in, In the graph, the vertex The degree, It is an empirical power, usually set to 3 / 4. Generally speaking, Navier-Stokes (NS) is the most widely used method for optimizing unsupervised representation learning objective functions because of its low computational complexity. To further understand this invention, Algorithm 1 below demonstrates the training process of IDRL. Specifically, the original short text is first preprocessed (line 2). Then, a graph is constructed to represent the text data (line 4). Next, NMF is used to obtain vertex attribute features (line 6). Then, forward propagation is performed to obtain aggregate embeddings (lines 8-14). Finally, NS is used as the objective function and stochastic gradient descent is used for optimization (line 15). The last two steps are repeated until convergence is achieved.

[0028]

[0029] In summary, the short text clustering analysis method based on inductive learning in this invention first maps the original text to a graph based on the co-occurrence of words within a sliding window. Then, a graph neural network is constructed, which can generate word representations by aggregating their neighbor information. Finally, an unsupervised representation learning method is used.

[0030] To further verify the effectiveness of this invention, the performance of document representation on real datasets was evaluated through clustering and classification tasks. Furthermore, the parameter sensitivity of IDRL to these tasks was also verified.

[0031] 1. Dataset Experiments were conducted on four widely used short text datasets, and the statistical data are shown in Table 1.

[0032] Table 1: Statistics of the dataset

[0033] TweetSet consists of tweets from the 2011 and 2012 microblog tracks published at the Text Retrieval Conference. It contains 2,472 tweets across 89 topics. TSet, SSet, and TSSet are three variants of Google News datasets, representing headlines, snippets, and combinations thereof, respectively. Generally, they contain 11,109 news items across 152 topics.

[0034] 2. Baseline Model Several state-of-the-art methods designed for short texts were used as baselines, including embedding learning models and probabilistic graphical models. The baselines are described below.

[0035] SIF proposes a relatively simple yet effective sentence embedding model that significantly outperforms baselines, including some complex supervision methods such as RNN and LSTM models. Word2vec is one of the most popular models. It first combines a two-layer neural network to learn word embeddings. The designed network structure is shallow and computationally efficient on large text corpora.

[0036] NMF utilizes nonnegative matrix factorization to obtain the vertex community distribution in a global understanding of the network. NMF is chosen as one of the baseline models because it can be applied to obtain vertex attribute features of the model.

[0037] GSDMM is one of the advanced probabilistic graphical models designed for short text clustering. It employs a hybrid of Dirichlet and multinomial distributions to learn document and word representations.

[0038] BTM is a classic probabilistic model that uses aggregation patterns (called bi-terms) to alleviate the problem of sparse word co-occurrence at the document level in short texts. 3. Parameter settings and evaluation indicators For models requiring predefined topic / community IDs (i.e., NMF, GSDMM, BTM, and proposed IDRL), provide the actual settings shown in the dataset. Since SIF requires pre-trained embeddings to initialize parameters, GloVe word embeddings trained on several large social network corpora are used. For Word2vec and IDRL, which are optimized using negative sampling, the number of negative samples can be uniformly set to 5. For other model parameters, follow the preferred settings in their respective papers. Furthermore, since IDRL utilizes a convergent neural network, the number of network layers can be adjusted accordingly. Hidden Dimensions And the neighborhood sample sizes of the layers are respectively and Furthermore, stochastic gradient descent is performed using the Adam optimizer and an initial learning rate of 1e-3.

[0039] To evaluate document clustering performance, two widely used metrics can be employed: Clustering Accuracy (ACC) and Normalized Mutual Information (NMI). Both ACC and NMI will be 1 when the clustering results perfectly match the true topics. Even when the clustering results are randomly generated, the values ​​will be close to zero. Furthermore, to measure inductive classification performance, the Liblinear library and default settings are used to build a classifier, employing Micro-F1 and Macro-F1 metrics.

[0040] 4. Inductive cluster evaluation In the experiments, the performance of the proposed IDRL was verified by inductive document clustering. Specifically, a portion of short texts (90%) was randomly selected as training data for representation learning. Then, combined with the learned word embeddings, each document embedding was obtained by recursively aggregating the neighbor information of its words. K-means was then run on the remaining documents (10%), for clustering words that had been visited previously. To ensure reliability, 10 runs were performed and the average of the experimental results was reported. Figure 3 and Figure 4 The clustering performance of various models on four short text datasets is demonstrated. More specifically, the following observations can be obtained: (1) From Figure 3 It can be observed that IDRL outperforms other baseline models in ACC, which verifies the effectiveness of the present invention. Generally, the order of ACC performance, excluding IDRL, is GSDMM > BTM > SIF > Word2vec > NMF. One possible reason is that GSDMM and BTM are specifically designed for clustering tasks (such as performing transductive clustering), while SIF, Word2vec, and NMF are more general representation learning methods.

[0041] (2) From Figure 4 As can be seen, IDRL can also achieve consistent improvements over state-of-the-art models, which complements IDRL's superiority in handling newly generated document clustering. Furthermore, IDRL outperforms NMF, indicating that IDRL's advantage surpasses NMF's prior knowledge (which this invention applies to obtain vertex community distributions as vertex attribute features before inductive representation learning).

[0042] 5. Summarize and classify evaluation To verify whether the model proposed in this invention can learn more discriminative representations, an inductive classification task was performed on four real-world datasets. More specifically, 30%, 50%, 70%, and 90% of the documents were randomly selected as the training set, and the remainder were used as the test set. Tables 2, 3, 4, and 5 show the document classification results, with the highest scores highlighted in bold.

[0043] Table 2: Document classification results of TweetSet (%)

[0044] Table 3: Document classification results of TSet (%)

[0045] Table 4: Document Classification Results of SSet (%)

[0046] Table 5: Document Classification Results of TSSet (%)

[0047] The following results can be observed from the table above: (1) The model of the present invention (IDRL) ​​significantly outperforms other models on all datasets with different training ratios, which demonstrates the effectiveness of IDRL in aggregating neighbor information of previously unlearned words to generate representations of newly generated documents.

[0048] (2) Specifically, the overall performance of the baseline follows this order: SIF, Word2vec > NMF > BTM > GSDMM. Although BTM and GSDMM perform well on clustering tasks, the representations they learn are less discriminative than those learned by SIF, Word2vec, and NMF. Overall, IDRL improves the average training rate of the second-best model by 11.5%, 8.1%, 3.7%, and 4.9% on TWeetSet, TSet, SSet, and TSSet, respectively. Experimental results show that the IDRL proposed in this invention can achieve more benefits on shorter texts (see Table 1 for dataset statistics), which verifies the ability of IDRL in learning short text representations.

[0049] 5. Parameter sensitivity In the experiment, the sensitivity of IDRL to parameters during clustering and classification tasks was also analyzed. First, the impact of dimension settings on clustering performance, Micro-F1, and Macro-F1 was verified, with dimensions set to {100, 128, 160, 190}, respectively. The experimental results were reported on TweetSet, as follows: Figure 5 As shown, "TR" is used to represent the training ratio. Specifically, from Figure 5 (a) As can be seen, NMI performance increases slowly with increasing dimensionality, while ACC performance remains stable. Generally, IDRL's clustering performance is robust to dimensionality settings. Furthermore, as... Figure 5 (b) and Figure 5As shown in (c), the classification performance curves, including Micro-F1 and Macro-F1, show an upward trend with slight fluctuations as the number of dimensions increases. For ease of demonstration, only the performance of TweetSet is shown, with training ratios of 30%, 50%, 70%, and 90%, respectively. In summary, the IDRL of this invention achieves relatively stable performance across different dimensions.

[0050] Furthermore, the impact of the number of clusters on clustering performance was further estimated by running K-means clustering on the test documents. The number of clusters was set to {70, 90, 110, 130, 150, 170, 190}, and NMI results were reported on four datasets, as shown below. Figure 6 As shown, ACC performance with similar trends is ignored. From Figure 6 As can be seen, IDRL achieves optimal performance on TweetSet when the cluster count is set to 90. Furthermore, on other datasets, IDRL initially increases with cluster size, then its performance decreases and stabilizes around 150. For larger values, IDRL begins to gradually decline. These detected cluster counts match the fundamental facts of the overall topics in the datasets, suggesting that IDRL can dynamically detect cluster counts within short texts.

[0051] In summary, this invention proposes an Inductive Learning-Based Short Text Clustering Analysis (IDRL) ​​model for short text clustering, which can effectively generate embeddings for newly generated documents that have been accessed with few words during training. Specifically, firstly, short texts are mapped to a graph network based on word co-occurrence within a sliding window. Then, global information of vertices in the graph is preserved (each word is represented as a vertex), and nonnegative matrix factorization is performed to obtain vertex-community distributions as vertex attribute features. Finally, for inductive learning, a graph neural network is constructed to aggregate neighbor information of words in newly generated documents. Overall, the proposed model is practical, and experimental results on real-world word datasets validate the effectiveness of IDRL in inductive clustering and classification tasks.

[0052] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0053] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0054] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0055] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0056] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0057] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0058] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0059] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A short text clustering analysis method based on inductive learning, comprising the following steps: The short text is preprocessed and mapped to a graph network based on word co-occurrence within a sliding window; The global information of vertices in the graph network is preserved, and the vertex-community distribution is obtained as the attribute feature of the vertex by performing feature extraction. For words in the newly generated text, a graph neural network is used to aggregate neighbor information and learn new vertex representations, where the neighbor information of words in the newly generated text is aggregated recursively.

2. The method according to claim 1, characterized in that, The attribute characteristics of the vertex are obtained according to the following formula: in, It is the distribution of vertex communities that encodes a global understanding. Indicates the number of potential communities. It is the Frobenius norm of the matrix. It is the harmonic factor. X is the set of vertices in a graph network, where X represents the vertex attribute features.

3. The method according to claim 1, characterized in that, The aggregated neighbor information is learned using a local spectral convolutional aggregator, and is represented as follows: Where MEAN represents the average of the elements of the vector. It is a layer The weight matrix, It is a layer Embedded vectors in It is the sigmoid function, and Represents the neighbors of a vertex , It is the context vertex.

4. The method according to claim 1, characterized in that, The objective function for learning the new vertex representation is set as follows: in, It is the target vertex. It is the context vertex. It is the sigmoid function. Indicates the negative sample distribution. It is from The negative samples, K represents the number of negative samples used for estimation, and , , It is a representation aggregated from features involving its local neighbors.

5. The method according to claim 4, characterized in that, The distribution of the negative samples Represented as: in, In the graph, the vertex The degree, It is an empirical power.

6. The method according to claim 1, characterized in that, The preprocessing of the short text includes removing stop words and stemming.

7. The method according to claim 1, characterized in that, The sliding window is set to 2.

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

9. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images