A non-iterative semi-supervised learning system for the automatic labeling of unlabeled text data

Abstract

A system for non-iterative, classifier-independent, similarity-based, semi-supervised text classification, consisting of: a data input unit configured to store and provide the input data set, consisting of labeled and unlabeled text data; a text preprocessing module connected to the data input unit, configured to receive the input data set, which includes both tagged and untagged text documents, and to process the text documents by removing URLs, superfluous spaces, HTML tags, email addresses, extending contractions, converting text to lowercase, performing tokenization, removing special characters and stop words, and performing lemmatization; a text vectorization module connected to the text preprocessing module, configured to convert the preprocessed text documents into numerical vector representations using at least one vectorization technique from the group consisting of CountVectorizer, Term Frequency-Inverse Document Frequency (TF-IDF) and Doc2Vec; a similarity calculation module connected to the text vectorization module and configured to: calculate class mean vectors for each class in the labeled text documents; calculate the cosine similarity between each unlabeled document vector and each class mean vector; and assign class labels to the unlabeled text documents based on the highest cosine similarity values; a classification module associated with the similarity calculation module, configured to train at least one classifier on a combined dataset comprising the originally tagged text documents and the newly tagged text documents, with the classification prediction for new text data being generated using the trained classifier; and an output module that is connected to the classification module and configured to provide labeled data as output.

Description

AREA OF INVENTION

[0001] The present disclosure relates to a non-iterative, semi-supervised learning system for the automatic labeling of unlabeled text data. More precisely, the invention relates to a non-iterative, classifier-independent, similarity-based system for semi-supervised text classification that automatically labels unlabeled text data by utilizing the cosine similarity between document vectors and class mean vectors. BACKGROUND OF THE INVENTION

[0002] The rapid growth of digital content and information systems has led to an unprecedented accumulation of unstructured text data. Effective text classification is crucial for organizing, retrieving, and analyzing this information in various applications, including spam detection, sentiment analysis, and message categorization. Traditional supervised learning models for text classification require large amounts of manually annotated data, the acquisition of which is expensive, time-consuming, and impractical in practice.

[0003] To address this challenge, semi-supervised learning (SSL) methods have been developed. These methods use a small set of labeled data together with a large pool of unlabeled data to train effective classifiers. However, existing SSL methods have significant limitations. Traditional approaches such as self-training, co-training, and EM-based algorithms involve complex iterative training processes, classifier-specific dependencies, and high computational costs. Self-training methods can amplify early errors if initial predictions are incorrect. Co-training requires assumptions of conditional independence between feature views, which are rarely true in real-world text data. EM-based approaches are sensitive to model initialization and often converge to unfavorable local optima.

[0004] Recent advances in SSL, including MixMatch, FixMatch, and UDA, have improved performance through data expansion, consistency regularization, and pseudo-labeling. However, these methods require significant computational resources, extensive hyperparameter optimization, and are often designed for large datasets. Their complex architectures increase implementation hurdles and reduce interpretability, making them less suitable for resource-constrained or resource-poor applications.

[0005] In the arts, there is a need for a simple, scalable, and classifier-independent approach that can automatically label unlabeled text documents without requiring iterative retraining or complex computational pipelines, while maintaining competitive classification performance and significantly reducing manual labeling effort. SUMMARY OF THE INVENTION

[0006] The present disclosure relates to a non-iterative, semi-supervised learning system for the automatic labeling of unlabeled text data. More specifically, the invention relates to a non-iterative, semi-supervised system that uses class-based mean vectors and cosine similarity for the automatic labeling of unlabeled documents. The proposed system utilizes various vectorization techniques, including CountVectorizer, TF-IDF, and Doc2Vec, and is not tied to any specific classifier. It employs classifiers such as logistic regression, feedforward neural networks (FNNs), and convolutional neural networks (CNNs), which were evaluated using three benchmark datasets. The results show that the proposed system reduces the manual labeling effort by over 99% while maintaining high classification performance across different models and datasets.

[0007] The present disclosure relates to a system for non-iterative, classifier-independent, similarity-based, semi-supervised text classification. The system comprises: a data input unit for storing and providing the input data set containing labeled and unlabeled text data; a text preprocessing module associated with the data input unit, which receives the input data set containing labeled and unlabeled text documents and processes the text documents by removing URLs, superfluous whitespace, HTML tags, email addresses, resolving contractions, converting to lowercase, tokenizing, removing special characters and stop words, and lemmatizing;A text vectorization module connected to the text preprocessing module, which converts the preprocessed text documents into numerical vector representations using at least one vectorization technique from the CountVectorizer, Term Frequency-Inverse Document Frequency (TF-IDF), and Doc2Vec groups; a similarity calculation module connected to the text vectorization module, which calculates class mean vectors for each class in the labeled text documents; calculates the cosine similarity between each vector of unlabeled documents and each class mean vector; and assigns class labels to the unlabeled text documents based on the highest cosine similarity values;a classification module connected to the similarity calculation module and configured to train at least one classifier on a combined dataset comprising the originally labeled text documents and the newly labeled text documents, with the classification prediction for new text data being generated using the trained classifier; and an output module connected to the classification module and configured to provide labeled data as output.

[0008] One objective of the present disclosure is to provide a non-iterative, semi-supervised learning system for the automatic labeling of unlabeled text data.

[0009] Another objective of the present disclosure is to provide a system configured to identify the unmarked text documents based on the cosine similarity between their vector representations and the mean vectors of the marked document classes.

[0010] Another objective of the present disclosure is to minimize the effort required for labeling while ensuring a high level of classification accuracy.

[0011] However, another objective of the present disclosure is to provide a semi-supervised learning-based system that offers a practical solution for bridging the gap between labeled and unlabeled data.

[0012] To further clarify the advantages and features of the present disclosure, the invention is described in more detail with reference to specific embodiments illustrated in the accompanying drawings. It is understood that these drawings merely show typical embodiments of the invention and are therefore not to be understood as limiting its scope of protection. The invention is described and explained in more detail and with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE IMAGES

[0013] These and other features, aspects and advantages of the present disclosure will be better understood when the following detailed description is read with reference to the accompanying drawings, in which identical symbols represent identical parts, wherein: Fig. Figure 1 illustrates a block diagram of a system for non-iterative, classifier-independent, similarity-based, semi-supervised text classification according to an embodiment of the present disclosure; and Fig. Figure 2 shows a block diagram of the workflow of the proposed similarity-based, semi-supervised text classification system according to an embodiment of the present disclosure.

[0014] Furthermore, those skilled in the art will recognize that the elements in the drawings are simplified and not necessarily drawn to scale. For example, the flowcharts illustrate the process by highlighting the main steps to facilitate understanding of this disclosure. With regard to the construction of the device, one or more components may be represented in the drawings by conventional symbols. The drawings may show only those specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawings with details that are already apparent to those skilled in the art from the description contained herein. DETAILED DESCRIPTION:

[0015] To facilitate understanding of the principles of the invention, reference is made below to the embodiment illustrated in the drawings, which is described using specific terms. It is understood, however, that this does not limit the scope of protection of the invention. Rather, modifications and further developments of the illustrated system, as well as further applications of the inventive principles depicted therein, are conceivable, insofar as they would typically occur to a person skilled in the art in the field of the invention.

[0016] It will be clear to those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not to be understood as a limitation of it.

[0017] References to “an aspect”, “another aspect”, or similar phrases in this description mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, phrases such as “in one embodiment”, “in another embodiment”, and similar expressions in this description may, but do not necessarily, all refer to the same embodiment.

[0018] The terms "includes," "comprehensive," or similar expressions denote non-exclusive inclusion. Thus, a procedure or method containing a list of steps does not only include those steps but may also include further steps not explicitly listed or inherent in the procedure or method. Likewise, the statement "includes..." for one or more devices, subsystems, elements, structures, or components, without further limitations, does not preclude the existence of other devices, subsystems, elements, structures, or components.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meanings generally known to those skilled in the art in the field to which this invention belongs. The systems, methods, and examples described herein serve only for illustration and are not to be understood as limiting.

[0020] Embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0021] The functional units described in this specification are referred to as devices. A device may be implemented in programmable hardware such as processors, digital signal processors, central processing units, FPGAs, PALs, PLDs, cloud processing systems, or similar. Devices may also be implemented in software for execution by various processor types. An identified device may contain executable code and, for example, comprise one or more physical or logical blocks of computer instructions, which may be organized as an object, procedure, function, or other construct. However, the executable files of an identified device need not be physically related; they may consist of different instructions stored in different locations that, when logically combined, constitute the device and fulfill its purpose.

[0022] The executable code of a device or module can consist of a single instruction or multiple instructions and can even extend across different code sections, applications, and storage media. Similarly, operational data within the device can be identified and represented, and can exist in any suitable form and be organized in any data structure. The operational data can be captured as a single data record or distributed across various storage media and may exist, at least partially, as electronic signals within a system or network.

[0023] References to “a selected embodiment”, “an embodiment”, or “an embodiment” in this description mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the disclosed subject matter. Therefore, the phrases “a selected embodiment”, “in an embodiment”, or “in an embodiment” appearing at different points in this description do not necessarily refer to the same embodiment.

[0024] Furthermore, the described features, structures, or properties can be combined in one or more embodiments in any suitable manner. The following description contains numerous specific details to enable a comprehensive understanding of the embodiments of the disclosed subject matter. However, a person skilled in the art will recognize that the disclosed subject matter can also be realized without one or more of the specific details or with other methods, components, materials, etc. In other cases, known structures, materials, or processes are not presented or described in detail so as not to obscure aspects of the disclosed subject matter.

[0025] According to the exemplary embodiments, the disclosed computer programs or modules can be executed in a variety of ways, for example, as an application running in the memory of a device or as a hosted application running on a server and communicating with the device application or browser via various standard protocols such as TCP / IP, HTTP, XML, SOAP, REST, JSON, and other suitable protocols. The disclosed computer programs can be written in programming languages ​​that run either in the device's memory or on a hosted server, such as BASIC, COBOL, C, C++, Java, Pascal, or scripting languages ​​such as JavaScript, Python, Ruby, PHP, Perl, or other suitable programming languages.

[0026] Some of the described embodiments involve data transmission over a network, such as the transmission of various inputs or files. The network may include, for example, the internet, wide area networks (WANs), local area networks (LANs), analog or digital wired and wireless telephone networks (e.g., PSTN, ISDN, cellular networks, and xDSL), radio, television, cable, satellite, and / or other transmission or tunneling mechanisms for data. It may include multiple networks or subnetworks, each of which may, for example, have a wired or wireless data path. The network may include a circuit-switched voice network, a packet-switched data network, or another network for transmitting electronic data. For example, it may be based on the Internet Protocol (IP) or Asynchronous Transfer Mode (ATM) and support voice communication using VoIP, Voice over ATM, or similar protocols.In one embodiment, the network comprises a mobile network configured for the exchange of text or SMS messages.

[0027] Examples of networks include Personal Area Networks (PAN), Storage Area Networks (SAN), Home Area Networks (HAN), Campus Area Networks (CAN), Local Area Networks (LAN), Wide Area Networks (WAN), Metropolitan Area Networks (MAN), Virtual Private Networks (VPN), Enterprise Private Networks (EPN), the Internet, Global Area Networks (GAN), and so on.

[0028] Fig. Figure 1 shows a block diagram of a system (100) for a non-iterative, classifier-independent, similarity-based, semi-supervised text classification according to an embodiment of the present disclosure.

[0029] According to Fig. 1 comprises: a data input unit (102) configured to store and provide the input data set containing labeled and unlabeled text data; a text preprocessing module (104) connected to the data input unit (102) and configured to receive the input data set containing labeled and unlabeled text documents and process the text documents by removing URLs, superfluous spaces, HTML tags, email addresses, resolving contractions, converting text to lowercase, performing tokenization, removing special characters and stop words, and performing lemmatization; a text vectorization module (106) connected to the text preprocessing module (104) and configured to convert the preprocessed text documents into numerical vector representations using at least one vectorization technique from the group CountVectorizer, Term Frequency-Inverse Document Frequency (TF-IDF), and Doc2Vec;A similarity calculation module (108) connected to the text vectorization module (106) is configured to: calculate class mean vectors for each class in the labeled text documents; calculate the cosine similarity between each unlabeled document vector and each class mean vector; and assign class labels to the unlabeled text documents based on the highest cosine similarity values; a classification module (110) connected to the similarity calculation module (108) is configured to train at least one classifier on a combined dataset comprising the originally labeled text documents and the newly labeled text documents, generating the classification prediction for new text data using the trained classifier; and an output module (112) connected to the classification module (110) is configured to provide labeled data as output.

[0030] In one embodiment, the text preprocessing module (104) comprises: a contraction extension component configured to extend contractions using a predefined list of at least 100 common contractions; a lemmatization component configured to reduce words to their base forms; and a stop word removal component configured to remove common words that do not contribute to classification accuracy.

[0031] In one embodiment, the text vectorization module (106) is configured to support multiple vectorization techniques simultaneously, and the system is configured to compare the performance of different vectorization methods to achieve optimal classification accuracy.

[0032] In one embodiment, the similarity calculation module (108) is configured to calculate the cosine similarity and select the class with the highest cosine similarity value for label assignment.

[0033] In one embodiment, the classification module (110) comprises several classification algorithms selected from the group consisting of the following algorithms: Multinomial Naïve Bayes; Stochastic Gradient Descent Classifier (SGDClassifier); Logistic Regression; Feedforward Neural Networks (FNN); and Convolutional Neural Networks (CNN).

[0034] In one embodiment, the feedforward neural networks (FNNs) comprise: three layers, including a first layer with 128 neurons with ReLU activation, a second layer with 64 neurons with ReLU activation, and an output layer with softmax activation; a dropout rate of 0.5; an Adam optimizer with a learning rate of 0.001; and a categorical cross-entropy loss function.

[0035] In one embodiment, the Convolutional Neural Networks (CNN) comprise: an ID convolution layer with 128 neurons and a kernel size of 5 with ReLU activation; a Max pooling layer with a pool size of 2; a flattening layer configured to convert 3D inputs to 1D outputs; a fully connected layer with 64 neurons and ReLU activation; and a dense output layer with Softmax activation.

[0036] In one embodiment, the data input unit (102), the preprocessing module (104), the text vectorization module (106), the similarity calculation module (108), the classification module (110) and the output module (112) can be implemented in programmable hardware devices such as processors, digital signal processors, central processing units, field-programmable gate arrays, programmable array logic, programmable logic devices, cloud processing systems or the like.

[0037] The present invention relates to a non-iterative, semi-supervised learning system that automatically labels unlabeled text data using the cosine similarity between document vectors and class mean vectors. The proposed method supports various vectorization techniques, including CountVectorizer, TF-IDF, and Doc2Vec, and is classifier-independent. This means it can be used with both traditional and deep learning models such as Multinomial Naive Bayes, SGDClassifier, logistic regression, feedforward neural networks (FNNs), and convolutional neural networks (CNNs).

[0038] Fig. Figure 2 shows a block diagram of the workflow of the proposed similarity-based, semi-supervised text classification system according to an embodiment of the present disclosure.

[0039] The proposed system, configured for semi-supervised learning, is in Fig.Figure 2 illustrates this. It serves for text classification and processes both labeled and unlabeled documents as input. The system uses semi-supervised learning to label unlabeled documents. A classifier is trained using labeled documents and then used to predict or generate the labels of the unlabeled documents. The labeled and unlabeled datasets are represented as vectors with the same number of features (same dimensions) to label the unlabeled documents. The mean vectors of the classes in the labeled dataset are calculated. If the similarity between the mean vector of a class and the document vector of a document in the unlabeled dataset is greatest, the corresponding class label from the labeled dataset is assigned to the unlabeled document.

[0040] In one embodiment, the system's preprocessing module is configured to clean and standardize text data to prepare it for analysis. Preprocessing operations include removing URLs, redundant whitespace, HTML tags, email addresses, contractions, converting text to lowercase, tokenization, removing special characters and stop words, and lemmatization. The lemmatization component of the preprocessing module reduces words to their base forms; for example, "running" is converted to "run." This minimizes dimensionality in vector representations such as CountVectorizer and TF-IDF while improving token alignment, thus increasing the semantic clarity of embeddings in Doc2Vec. The contraction extension component of the preprocessing module is configured to use a predefined list of one hundred contractions for extension.Extending contractions standardizes the language for improved token matching and optimized semantic representation. For example, the expression "I'm not sure" is extended to "I am not sure," allowing the system to learn more precise word and sentence embeddings.

[0041] The system's text vectorization module is configured to represent preprocessed text data in vector form for natural language processing (NLP) tasks. The system supports various methods, including CountVectorizer, Term Frequency-Inverse Document Frequency (TF-IDF), and Doc2Vec, each of which converts text into numerical formats that can be processed by machine learning algorithms.

[0042] The CountVectorizer component of the text vectorization module converts a corpus of text documents into a matrix of token frequencies. It generates a vocabulary of unique words and represents each document as a vector indicating the frequency of each word. For a corpus D = { d 1, d 2,..... dn}, CountVectorizer creates a document-term matrix X ∈ Rm × n, where m is the vocabulary size.

[0043] The TF-IDF component of the text vectorization module is configured to enhance CountVectorizer by considering the relative importance of words across the entire corpus. It combines the term frequency (TF), which represents word frequency within a document, with the inverse document frequency (IDF), which indicates word importance across the entire corpus. The TF-IDF value for a term i in document j is defined as follows: TF−IDFi,j=TFi,j×log(N / ni)

[0044] In the equation above, N denotes the total number of documents and n i The number of documents containing the term i. TF-IDF reduces the influence of frequent words and allows for better class separation because class-specific keywords are given more weight. This representation improves performance in similarity-based tagging because it highlights discriminative terms. However, it still provides sparse representations and has low semantic depth.

[0045] The Doc2Vec component of the text vectorization module is configured to generate dense vector representations of entire documents, capturing both syntactic and semantic features. Based on neural network models, Doc2Vec builds upon Word2Vec and exists in two variants: Distributed Memory (PV-DM), which predicts a target word using contextual words and a document vector, and Distributed Bag of Words (PV-DBOW), which predicts words within a document using the document vector. Doc2Vec learns compact document representations that reflect semantic meaning, making it particularly suitable for tasks such as sentiment analysis and document classification. Unlike CountVectorizer and TF-IDF, Doc2Vec generates dense vectors capable of capturing both context and meaning.With smaller datasets, however, Doc2Vec can produce overlapping representations across different classes, as documents on different topics may exhibit semantic similarities. This limitation is particularly noticeable with short-text datasets like Inshorts, where the limited semantic context reduces its effectiveness. Doc2Vec generally requires larger and longer documents to make clear class distinctions and is less effective with simple, cosine-similarity-based tagging, as it tends to blur class boundaries.

[0046] In one embodiment, the integration of CountVectorizer, TF-IDF, and Doc2Vec configures the system's text vectorization module to be robust against varying text properties and display requirements. CountVectorizer enables frequency-based structure classification, TF-IDF highlights the relevance of terms for differentiated labeling, and Doc2Vec captures context-related semantics.

[0047] In one embodiment, the system's similarity calculation module is configured to assign labels to unlabeled documents based on the cosine similarity between their vector representation and the mean vector of the labeled documents of each class. The cosine similarity metric used by the system is defined as follows: sim(du→,μc→)=(du→.μc→) / (|du→|×|μc→|)

[0048] This similarity measure captures angular proximity in high-dimensional space and is therefore particularly effective in text analysis, where document vectors are often sparse. Cosine similarity is well-suited for processing high-dimensional and sparse text data by normalizing vector magnitudes and emphasizing directional alignment. This effectively captures thematic similarity regardless of document length or term frequency.

[0049] In one embodiment, the system's classification module is configured to train classification algorithms using the dataset that includes both the originally labeled documents and those relabeled through similarity-based matching. The classification module supports both traditional and deep learning machine learning algorithms to evaluate the effectiveness of the proposed labeling method. Traditional machine learning algorithms supported by the classification module include Multinomial Naive Bayes (MultinomialNB), Stochastic Gradient Descent Classifier (SGDClassifier), and Logistic Regression. The classification module also incorporates deep learning algorithms such as Feedforward Neural Networks (FNNs) and Convolutional Neural Networks (CNNs) to enhance classification performance.Through this integration of traditional and deep learning algorithms, the system's classification module ensures flexibility and a comprehensive evaluation of the labeled datasets, thus supporting the effectiveness of the proposed non-iterative, classifier-independent, similarity-based, semi-supervised text classification framework.

[0050] The system's performance evaluation module is configured to assess the model's effectiveness using standard performance indicators, including accuracy, precision, hit rate, F1 score, and ROC-AUC score (Olson & Delen, 2008).

[0051] In one embodiment, a performance evaluation is conducted to determine the performance of the proposed system based on accuracy, precision, hit rate, F1 score, and ROC-AUC curve. Accuracy indicates the system's success in fulfilling its task by determining the percentage of correct predictions or classifications. Precision measures system performance based on the quality of positive predictions. Hit rate measures the system's ability to correctly identify positive instances in a dataset. The F1 score evaluates system performance by combining precision and hit rate into a single value. The ROC-AUC curve measures the system's classification performance in distinguishing between positive and negative classes.

[0052] In one embodiment, a complexity analysis of the system is performed, evaluating the time complexity of the labeling and classification operations. The analysis involves splitting the dataset into a training and a test dataset. The total computational cost depends essentially on three operations: calculating the mean vector, performing similarity calculations, and vectorization costs. Calculating the class-wise mean vectors, which averages the vectors of all labeled documents, requires O(p × m × n), where m is the number of documents in a class. Similarity calculations, which compare each unlabeled document to p mean vectors using cosine similarity, require O(j × p × n). The computational cost of vectorization depends on the representation method used.Vectorization with CountVectorizer or TF-IDF requires O(q × n), while the Doc2Vec representation method requires O(q × k × d), where k is the window size and d is the embedding dimension. Therefore, the total computational complexity of the system for class assignment can be expressed as follows: T(n)=O(p×m×n)+O(j×p×n)+O(vectorization cost)

[0053] While the labeling algorithm is linear in terms of the number of unlabeled documents during implementation, the overall runtime depends significantly on the similarity calculation and vectorization costs, which must be taken into account when implementing the system on large datasets or real-time streaming data.

[0054] The system's non-iterative labeling framework eliminates the need for iterative retraining, as is common with traditional semi-supervised methods such as self-training or EM-based models. By using cosine similarity in a single-pass labeling mechanism, the system avoids dependence on initial conditions. The system's classifier independence ensures that the labeling algorithm remains model-agnostic and can therefore be integrated into any downstream classifier—both in traditional machine learning models and deep learning architectures. This contrasts with existing approaches that are tightly coupled to specific classifiers such as Naive Bayes or Support Vector Machines (SVMs).The system's flexibility with regard to multi-embedding enables adaptability to various text representation techniques by supporting and evaluating three different methods: CountVectorizer, TF-IDF, and Doc2Vec. Previous approaches, in contrast, typically rely on a single representation. Furthermore, the system's simplicity and scalability offer easy implementation and high computational efficiency, as no training is required during labeling. The algorithm's low complexity ensures scalability to large datasets, making it particularly suitable for practical industrial applications where interpretability and low computational costs are critical requirements.

[0055] In one embodiment, the system's data input module is configured to use three different corpora of varying sizes to evaluate performance and generalizability. These corpora include the BBC corpus, the Inshorts news dataset, and the 20 Newsgroups dataset. The BBC corpus contains 2,225 annotated documents distributed across five subject categories: business, entertainment, politics, sports, and technology. The Inshorts dataset consists of 12,120 short news articles with three columns (headline, article, and category) and covers seven categories: technology, sports, politics, entertainment, world affairs, automotive, and science. The 20 Newsgroups dataset comprises nearly 20,000 documents with 20 different classifications and serves as a comprehensive benchmark. The system is configured to implement both the traditional 70:30 split between training and test data and a 30:70 configuration.The 30-70 configuration reflects the conditions of semi-supervised learning, where only 30% of the documents are labeled and the remaining 70% are unlabeled. With 30% labeled data, class means can be estimated effectively, and cosine similarity is applied to label the unlabeled portion. This configuration proves robust in high-dimensional, sparse data spaces, such as those generated by TF-IDF or Doc2Vec, and corresponds to real-world conditions where labeled data is scarce and unlabeled data is abundant. The system is also configured to handle datasets with varying domain characteristics, document lengths, and vocabulary complexities. BBC news documents are longer and more formal, which favors TF-IDF and Doc2Vec embeddings.Inshorts documents are shorter and more concise, often under 100 words, which, while posing challenges for semantic embedding, allows for the use of accuracy as the primary metric. The "20 Newsgroups" dataset contains diverse and informal content with high vocabulary variability, and TF-IDF delivers stable results. Despite these differences, the system demonstrates robust classification accuracy and adaptability across all three datasets.

[0056] In one embodiment, a hyperparameter tuning component of the system is configured to optimize both traditional classifiers and deep learning architectures. For traditional classifiers such as logistic regression, the SGDClassifier, and multinomial Naive Bayes, grid-based cross-validation is applied to refine the key parameters. For deep learning classifiers, including feedforward neural networks (FNNs) and convolutional neural networks (CNNs), validation-based tuning is used to configure the network architectures. The FNN consists of three layers: two hidden layers with 128 and 64 neurons, respectively, and a ReLU activation function, followed by a softmax output layer, a dropout rate of 0.5, and Adam optimization with a learning rate of 0.001.The CNN consists of a one-dimensional convolutional layer with 128 neurons and a kernel size of 5, using the ReLU activation function, followed by max pooling, flattening, a fully connected layer with 64 neurons using the ReLU activation function, and a softmax output layer. This output layer was also trained with Adam optimization and a dropout rate of 0.5. The system is configured to ensure reliable evaluation across datasets of varying sizes and domains, while maintaining efficiency and accuracy through optimized hyperparameter selection.

[0057] The system was validated in an implementation through extensive experimentation with three benchmark datasets: BBC, Inshorts, and 20-Newsgroups. It achieved accuracy of 96.88% on BBC, 93.59% on Inshorts, and 92.49% on 20-Newsgroups, with only 30% of the data annotated. This configuration reduced manual annotation effort by more than 99%, thus confirming the efficiency of the semi-supervised framework. Analysis revealed that the TF-IDF representation consistently outperformed CountVectorizer and Doc2Vec, achieving accuracy improvements of between 3 and 12 percentage points in most experimental scenarios. Of the classification components, logistic regression and the feedforward neural network (FNN) showed the best overall performance, while the SGDClassifier demonstrated stable reliability with sparse vector representations.It was found that deep learning classifiers were more sensitive to dataset size and embedding choice, but their performance stabilized under TF-IDF representations. The system also demonstrated robustness in low-monitoring scenarios, maintaining high accuracy and F1 scores with only 30% labeled data, thus highlighting its practical utility when labeled data resources are limited. Statistical validations using paired t-tests and ablation studies confirmed the reliability and consistency of the system's performance under various conditions. The results show that the system effectively bridges the gap between labeled and unlabeled text data without iterative retraining.Its design ensures resource efficiency, scalability and interpretability, and is therefore particularly suitable for real-time document classification as well as for bootstrapping labeled datasets in data-poor areas.

[0058] The drawings and the preceding description illustrate embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another. For example, the process flows described here can be modified and are not limited to the manner described herein. Furthermore, the actions of a flowchart need not be performed in the sequence shown; nor do all actions necessarily need to be carried out. Actions that do not depend on other actions can be performed in parallel with the other actions. The scope of protection of the embodiments is in no way limited by these specific examples. Numerous variations, whether explicitly stated in the description or not, such as...Differences in structure, dimensions, and materials are possible. The scope of protection of the embodiments is at least as comprehensive as described by the following claims.

[0059] The advantages, other benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and any components that can effect or enhance an advantage, benefit, or solution are not to be construed as critical, necessary, or essential features or components of the claims. REFERENCES 100 A system for non-iterative, classifier-independent, similarity-based, semi-supervised text classification. 102 Data input unit 104 Preprocessing module 106 Text vectorization module 108 Similarity Calculation Module 110 Classification module 112 Output module 202 records with labels 204 Text preprocessing 206 Text presentation 208 Data division into two different types 210 Training (70%) and examination (30%) 212 Training (30%) and examination (70%) 214 Similarity between the mean of the label classes and the mean of the unlabeled dataset (predicted labels) to the test dataset 216 Training of the classifiers 218 tested classifiers with the test dataset (with the predicted labels) QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature

[0000] Olson & Delen, 2008)

[0050]

Claims

[1] A system for non-iterative, classifier-independent, similarity-based, semi-supervised text classification, consisting of: a data input unit configured to store and provide the input data set, consisting of labeled and unlabeled text data; a text preprocessing module connected to the data input unit, configured to receive the input data set, which includes both tagged and untagged text documents, and to process the text documents by removing URLs, superfluous spaces, HTML tags, email addresses, extending contractions, converting text to lowercase, performing tokenization, removing special characters and stop words, and performing lemmatization; a text vectorization module connected to the text preprocessing module, configured to convert the preprocessed text documents into numerical vector representations using at least one vectorization technique from the group consisting of CountVectorizer, Term Frequency-Inverse Document Frequency (TF-IDF) and Doc2Vec; a similarity calculation module connected to the text vectorization module and configured to: calculate class mean vectors for each class in the labeled text documents; calculate the cosine similarity between each unlabeled document vector and each class mean vector; and assign class labels to the unlabeled text documents based on the highest cosine similarity values; a classification module associated with the similarity calculation module, configured to train at least one classifier on a combined dataset comprising the originally tagged text documents and the newly tagged text documents, with the classification prediction for new text data being generated using the trained classifier; and an output module that is connected to the classification module and configured to provide labeled data as output. [2] System according to claim 1, wherein the text preprocessing module comprises: a contraction extension component configured to extend contractions using a predefined list of at least 100 common contractions; a lemmatization component configured to reduce words to their base forms; and a stop word removal component configured to remove common words that do not contribute to classification accuracy. [3] System according to claim 1, wherein the text vectorization module is configured to support multiple vectorization techniques simultaneously, and the system is configured to compare the performance of different vectorization methods to achieve optimal classification accuracy. [4] System according to claim 1, wherein the similarity calculation module is configured to calculate the cosine similarity and select the class with the highest cosine similarity value for label assignment. [5] System according to claim 1, wherein the classification module comprises several classification algorithms selected from the group consisting of the following algorithms: Multinomial Naive Bayes; Stochastic Gradient Descent Classifier (SGDClassifier); Logistic Regression; Feedforward Neural Networks (FNN); and Convolutional Neural Networks (CNN). [6] System according to claim 5, wherein the feedforward neural networks (FNNs) comprise: three layers, consisting of a first layer with 128 neurons with ReLU activation, a second layer with 64 neurons with ReLU activation and an output layer with softmax activation; a dropout rate of 0.5; an Adam optimizer with a learning rate of 0.001; and a categorical cross-entropy loss function. [7] System according to claim 5, wherein the Convolutional Neural Networks (CNN) comprise: a 1D convolution layer with 128 neurons and a kernel size of 5 with ReLU activation; a Max pooling layer with a pool size of 2; a flattening layer configured to convert 3D inputs to 1D outputs; a fully connected layer with 64 neurons and ReLU activation; and a dense output layer with Softmax activation.

Images