Description:FIELD OF THE INVENTION
[0001] The invention relates to natural language processing, particularly to automated text coherence analysis employing multiple feature extraction methods and meta-learning frameworks that detect coherence break points and provide improvement suggestions.

DESCRIPTION OF THE RELATED ART
[0002] The following description of the related art is intended to provide background information pertaining to the field of disclosure. This section may include certain aspects of the art that may be related to various features of the present disclosure. However, it should be appreciated that this section is used only to enhance the understanding of the reader with respect to the present disclosure, and not as admissions of the prior art.
[0003] Text coherence plays a critical role in the quality and effectiveness of written communication. Coherent texts demonstrate logical flow, organized structure, and semantic continuity, which significantly impact comprehensibility, information retention, and persuasiveness. Traditional approaches to evaluating text coherence rely primarily on manual assessment by human readers, which is inherently subjective, time-consuming, and difficult to scale.
[0004] Previous computational approaches to coherence assessment have focused on isolated aspects of text structure. Entity-based methods track the grammatical roles of entities across sentences but often fail to capture deeper semantic relationships. Topic modeling approaches can identify thematic elements but struggle with structural coherence. Embedding-based techniques provide semantic representations but may miss discourse-level patterns. Similarity-based methods employing cosine similarity often produce misleading results, where coherent texts show lower similarity scores than incoherent ones, failing to account for grammatical consistency, tense shifts, and discourse structures. Moreover, existing systems typically lack the ability to provide specific, actionable feedback on improving coherence.
[0005] Academic and professional contexts particularly suffer from coherence issues. Students writing academic reports often struggle with maintaining logical flow throughout their documents, leading to reduced clarity and impact. Traditional educational feedback mechanisms are limited by the availability of human evaluators and the subjective nature of manual coherence assessment. Automated systems that rely solely on semantic similarity fail to identify critical coherence issues such as abrupt topic shifts, grammatical inconsistencies, and structural discontinuities.
[0006] Additionally, conventional coherence analysis systems often employ single-feature approaches or simple machine learning models that fail to capture the multifaceted nature of text coherence. These systems typically do not integrate entity transitions, topic distributions, and contextual embeddings in a comprehensive framework, resulting in incomplete coherence evaluation. Furthermore, existing systems lack constraint-based analysis that combines multiple linguistic and semantic features through weighted evaluation to accurately identify coherence break points.
[0007] Therefore, there exists a requirement for an automated text coherence analysis approach that combines multiple feature types through advanced meta-learning frameworks, detects specific coherence break points employing constraint-based analysis with weighted combinations of binary and continuous measures, and provides constraint-specific targeted feedback for improving textual organization and flow.

OBJECTS OF THE PRESENT DISCLOSURE
[0008] Some of the objects of the present disclosure, which at least one embodiment herein satisfies are as listed herein below.
[0009] An object of the present disclosure is to provide an automated text coherence analysis approach which integrates multiple feature types for comprehensive coherence assessment.
[0010] An object of the present disclosure is to implement a multi-feature meta-learning framework which combines entity grid data, topic representations, and contextual embeddings to classify text coherence levels.
[0011] An object of the present disclosure is to detect coherence break points within text employing constraint-based analysis that evaluates binary indicators including entity shift, pronoun shift, tense shift, grammatical errors, and discourse markers, along with continuous measures including part-of-speech similarity, embedding similarity, and topic similarity.
[0012] An object of the present disclosure is to calculate weighted sums of multiple constraints for each sentence pair and compare against predefined thresholds to identify specific locations where coherence degrades.
[0013] An object of the present disclosure is to generate coherence improvement feedback through feature-wise analysis that identifies which specific constraints contributed to each coherence break point and provides constraint-specific recommendations.
[0014] An object of the present disclosure is to provide targeted feedback including sentence splitting recommendations for semantic similarity drops, transition phrases for topic shifts, and grammatical corrections for tense and pronoun inconsistencies.
[0015] An object of the present disclosure is to present coherence classification outputs and improvement feedback through a visually interpretable user interface format.
[0016] An object of the present disclosure is to apply Local Interpretable Model-Agnostic Explanations to identify specific features which contribute to coherence classification.
[0017] An object of the present disclosure is to enable optimization of the coherence analysis models through retraining processes which incorporate feedback on coherence classification outputs.

SUMMARY
[0018] This section is provided to introduce certain objects and aspects of the present disclosure in a simplified form that are further described below in the detailed description. This summary is not intended to identify the key features or the scope of the claimed subject matter.
[0019] The present disclosure generally relates to natural language processing and automated text analysis systems. More particularly, the present disclosure relates to an automated text coherence analysis system that employs multi-feature meta-learning frameworks for accurate coherence classification through integration of entity grid data, topic representations, and contextual embeddings, providing enhanced text assessment capabilities through constraint-based analysis and weighted sum-based decision making for improved writing quality and educational effectiveness.
[0020] An aspect of the present disclosure relates to an automated text coherence analysis system comprising a processor and memory storing instructions for coherence analysis. The system includes modules for: receiving text input with paragraphs and sentences; extracting entity grid data, topic representations using Latent Dirichlet Allocation and graph convolutional networks, and contextual embeddings from transformer-based language models; implementing a multi-feature meta-learning framework with first-level machine learning models and a second-level meta-model; detecting coherence break points using constraint-based analysis with binary indicators and continuous measures; calculating weighted sums compared against thresholds; generating constraint-specific feedback recommendations; and displaying results through a user interface.
[0021] In another aspect, the present disclosure relates to a method for automated text coherence analysis comprising: receiving text input using a processor; generating entity grid data, extracting topic representations using Latent Dirichlet Allocation and graph convolutional networks, and obtaining contextual embeddings from transformer-based language models; implementing a multi-feature meta-learning framework; detecting coherence break points using constraint-based analysis; calculating weighted sums against thresholds; generating improvement feedback through feature-wise analysis; providing constraint-specific recommendations including sentence splitting, transition phrases, and grammatical corrections; and displaying results through a user interface.
[0022] Various objects, features, aspects and advantages of the inventive subject matter will become more apparent from the following detailed description of preferred embodiments, along with the accompanying drawing figures in which like numerals represent like components.

BRIEF DESCRIPTION OF DRAWINGS
[0023] The accompanying drawings, which are incorporated herein, and constitute a part of this invention, illustrate exemplary embodiments of the disclosed methods and systems for automated text coherence analysis which like reference numerals refer to the same parts throughout the different drawings. Components in the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present invention. Some drawings may indicate the components employing block diagrams and may not represent the internal circuitry of each component. It will be appreciated by those skilled in the art that the invention of such drawings includes the invention of electrical components, electronic components or circuitry commonly used to implement such components, as well as the machine learning algorithms and computational methods used in the multi-feature meta-learning framework for coherence classification.
[0024] FIG. 1 illustrates an exemplary block diagram of an automated text coherence analysis system, in accordance with an embodiment of the present disclosure.
[0025] FIG. 2 illustrates an exemplary flowchart of the automated text coherence analysis method, in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION
[0026] While the present disclosure has been disclosed with reference to certain embodiments, it will be understood by those skilled in the art that various changes may be made, and equivalents may be substituted without departing from the scope of the invention. In addition, many modifications may be made to adapt to a particular situation or material to the teachings of the invention without departing from its scope.
[0027] The present disclosure provides systems and methods for automated text coherence analysis employing a multi-feature meta-learning approach. While the disclosure may be susceptible to embodiment in different forms, there is depicted in the drawings, and herein will be described in detail, specific embodiments with the understanding that the present disclosure is to be considered an exemplification of the principles of the invention, and is not intended to limit the invention to that as illustrated and described herein.
[0028] In the present disclosure, the term "coherence" refers to the logical flow and connectivity of ideas within a text, encompassing how sentences and paragraphs are organized and linked to convey meaning clearly. "Entity grid data" refers to a representation of how entities (nouns or noun phrases) appear and transition between grammatical roles across consecutive sentences in a text. "Topic representations" refer to the distribution of thematic elements across sentences and how these themes connect to create a unified discourse. "Contextual embeddings" refer to vector representations of text that capture semantic meaning and contextual relationships as derived from transformer-based language models, constraint-based analysis" refers to the evaluation of multiple linguistic and semantic constraints including binary indicators and continuous measures that are combined through weighted sum calculation to detect coherence degradation.
[0029] The present disclosure provides a system for automated text coherence analysis that addresses limitations of existing approaches by combining multiple feature types, employing a two-level machine learning architecture, detecting specific coherence break points through constraint-based analysis, and generating constraint-specific actionable improvement feedback. The system provides a comprehensive, automated solution for evaluating and enhancing text coherence in educational and professional contexts.
[0030] Various embodiments of the present disclosure are described employing FIGs. 1 to 2.
[0031] FIG. 1 illustrates an exemplary block diagram of an automated text coherence analysis system, in accordance with an embodiment of the present disclosure.
[0032] As illustrated in FIG. 1, the automated text coherence analysis system (100) includes several layers and components. The hardware layer includes a processor (101) and memory (102) for executing the computational operations required for coherence analysis. The processor may be a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), or any other suitable processing device capable of executing the required computations. The memory may include random access memory (RAM), read-only memory (ROM), cache memory, flash memory, hard disk drives, solid-state drives, or any combination thereof suitable for storing program instructions, data structures, and computational results.
[0033] In an embodiment, the system receives text input (103) in the form of paragraphs with sentences having grammatical structures and semantic content. This input may be provided through various channels including direct user input through a text editor interface, file upload of documents in formats such as .txt, .docx, .pdf, or .html, or through application programming interfaces (APIs) that enable integration with other software systems. The text input undergoes preprocessing to normalize formatting, correct minor typographical errors, and prepare the text for feature extraction.
[0034] In an embodiment, the feature extraction layer of the system includes three primary components: entity grid data generation (104), topic representation extraction (105), and contextual embedding generation (106). These components operate in parallel to extract different types of features from the text input, each capturing distinct aspects of coherence.
[0035] In an embodiment, the entity grid data generation component (104) identifies entities within the sentences of the text input and tracks their grammatical roles across consecutive sentences. This process begins with entity identification employing natural language processing techniques including part-of-speech tagging, named entity recognition, and noun phrase extraction. For each identified entity, the system determines its grammatical role in each sentence, categorizing it as Subject (S) when it serves as the subject of a main or subordinate clause, Object (O) when it serves as a direct or indirect object, Other (X) when it appears in any other grammatical role, or Absent (-) when it does not appear in the sentence.
[0036] In an embodiment, the entity grid data generation component then constructs a two-dimensional grid where rows represent sentences and columns represent entities. Each cell in the grid contains the grammatical role (S, O, X, or -) of the corresponding entity in the corresponding sentence. After constructing this grid, the system analyzes transitions of grammatical roles between consecutive sentences for each entity. These transitions include sixteen possible patterns: S-S, S-O, S-X, S--, O-S, O-O, O-X, O--, X-S, X-O, X-X, X--, --S, --O, --X, and ----. The system calculates the frequency of each transition type and normalizes these counts to produce a 16-dimensional feature vector representing entity transition probabilities. These probabilities capture patterns of entity continuity that are strongly correlated with perceived coherence, with coherent texts typically showing higher proportions of transitions that maintain entity presence (e.g., S-S, O-O, X-X) and incoherent texts showing higher proportions of transitions indicating entity discontinuity (e.g., --S, --O).
[0037] In an embodiment, the topic representation extraction component (105) applies topic modeling techniques to capture thematic structures within the text input. This process begins with text preprocessing including tokenization, stopword removal, and lemmatization to prepare the text for topic modeling. The system then applies Latent Dirichlet Allocation (LDA) to the preprocessed text to discover latent topics and generate probability distributions for each sentence across these topics. The LDA algorithm operates by modeling each document (in this case, sentence) as a mixture of topics and each topic as a distribution over words. The algorithm iteratively adjusts these distributions to maximize the likelihood of the observed word occurrences, ultimately producing topic probability distributions that indicate the likelihood of each sentence belonging to specific topics.
[0038] After generating topic distributions, the topic representation extraction component constructs a graph where sentences are represented as nodes. Edges between nodes are weighted according to cosine similarity between their respective topic distributions, with higher similarity indicating stronger thematic connections. This graph structure is then processed employing graph convolutional networks (GCNs), which apply convolution operations to node features by aggregating information from neighboring nodes. This process captures inter-sentence relationships by incorporating information about thematic connections between each sentence and its neighbors, producing refined topic representations that reflect not just individual sentence topics but also their contextual relationships within the discourse structure.
[0039] In an embodiment, the contextual embedding generation component (106) utilizes transformer-based language models to create vector representations of text that capture semantic meaning and contextual relationships. This process begins with tokenization of the text input according to the requirements of the chosen language model, which may include subword tokenization methods such as WordPiece, Byte-Pair Encoding, or SentencePiece. The tokenized text is then processed through pre-trained language models such as BERT, RoBERTa, or T5, which have been trained on large text corpora to understand linguistic patterns and semantic relationships.
[0040] These models process text through multiple layers of self-attention mechanisms that capture relationships between all tokens in the sequence, producing contextualized vector representations for each token. The system extracts these representations and aggregates them to produce sentence-level and paragraph-level embeddings. For sentence-level embeddings, the system may use the representation of special tokens such as [CLS] in BERT, average the representations of all tokens in the sentence, or apply more sophisticated pooling methods. For paragraph-level embeddings, the system may similarly aggregate sentence-level embeddings or process the entire paragraph through the language model to capture cross-sentence contextual relationships. These embeddings, typically 768-dimensional vectors for models like BERT, capture deep semantic relationships that extend beyond surface-level linguistic features, enabling assessment of coherence based on meaning and context.
[0041] In an embodiment, the machine learning pipeline of the system consists of the multi-feature meta-learning framework (107), which includes first-level machine learning models (108) and a second-level meta-model (109). This framework follows a stacked generalization approach, where multiple base models are trained on different feature sets and their outputs are combined by a meta-model to produce the final classification.
[0042] The first-level machine learning models (108) consist of at least five models selected from: support vector machine (SVM), random forest, k-nearest neighbors (KNN), logistic regression, multi-layer perceptron (MLP), decision tree, naïve Bayes, gradient boosting, AdaBoost, and XGBoost. Each model is trained on specific feature sets derived from the entity grid data, topic representations, and contextual embeddings. For example, SVMs may be trained on entity transition probabilities, random forests on topic distributions, and neural networks on contextual embeddings. Each model produces a probability distribution across coherence classes (Highly Coherent, Moderately Coherent, and Incoherent) for the input text.
[0043] The first-level models undergo extensive hyperparameter optimization to maximize performance. For SVMs, this includes tuning the C parameter (regularization strength), kernel type (linear, polynomial, radial basis function), gamma parameter (kernel coefficient), and degree (for polynomial kernels). For random forests, parameters include the number of estimators, maximum depth, minimum samples for splitting, and minimum samples per leaf. For neural networks, optimization covers learning rate, hidden layer sizes, activation functions, and regularization parameters. This tuning ensures that each base model achieves optimal performance on its specific feature set.
[0044] The second-level meta-model (109) combines outputs from all first-level models to produce a final coherence classification. This model is typically implemented as a k-nearest neighbors classifier, which has demonstrated superior performance in aggregating base-level decisions due to its ability to capture complex relationships between model outputs without overfitting. The meta-model takes as input the concatenated probability distributions from all first-level models, creating a feature vector that represents the collective "wisdom" of the diverse base models. By learning which base models perform best in different scenarios, the meta-model leverages the complementary strengths of each approach, producing more accurate classifications than any individual model.
[0045] In an embodiment, the output and analysis components of the system include coherence classification (110), coherence break point detection (111), coherence improvement feedback generation (112), and user interface presentation (113). These components translate the technical analysis results into actionable insights for users.
[0046] In an embodiment, the coherence classification component (110) assigns the text input to one of three coherence classes: Highly Coherent, Moderately Coherent, or Incoherent. Highly Coherent texts exhibit excellent logical flow with well-organized ideas and smooth transitions between sentences. Moderately Coherent texts maintain a general sense of topic continuity but may suffer from minor disruptions or weaker transitions. Incoherent texts display poor sentence-to-sentence linkage, abrupt topic shifts, or lack of structural clarity. This classification provides a high-level assessment of overall text quality regarding organization and flow.
[0047] In an embodiment, the coherence break point detection component (111) identifies specific locations within the text input where coherence decreases. This process involves several analytical approaches working in concert. The system calculates semantic similarity between adjacent sentences employing cosine similarity of their contextual embeddings, detecting abrupt drops that indicate potential coherence breaks. Grammatical inconsistencies are identified by analyzing changes in verb tense, subject continuity, and pronoun reference between consecutive sentences, employing part-of-speech tagging and dependency parsing to track these linguistic features. Sentence length variations are measured to identify structural incongruities, with abrupt changes from very short to very long sentences (or vice versa) potentially disrupting flow. Topic shifts are detected by comparing changes in topic distribution vectors between sentences, with substantial shifts indicating potential thematic disconnects.
[0048] In an embodiment, the coherence break point detection component (111) identifies specific locations within the text input where coherence decreases employing constraint-based analysis. This process involves calculating multiple constraints for each pair of consecutive sentences, including binary indicators for entity shift, pronoun shift, tense shift, presence of grammatical errors, and presence of discourse markers, along with continuous measures including part-of-speech (POS) similarity, embedding similarity, and topic similarity. For each sentence pair, the system computes values for all constraints and combines them employing a weighted sum, where each constraint is assigned a weight based on its relevance to coherence assessment. The resulting weighted sum is compared against a predefined threshold. A weighted sum below the threshold signifies a potential coherence issue between the sentence pair, whereas a sum above the threshold indicates that the sentence pair is likely coherent.
[0049] In an embodiment, the coherence break point detection specifically involves implementing the constraint analysis module that calculates constraint values, the weighted sum calculator that combines constraints employing assigned weights, and the threshold comparator that identifies break points where the weighted sum falls below the threshold. For example, when analyzing a paragraph, the system may detect that the weighted sum between sentences 3 and 4 yields a score of -0.494, which falls below the threshold of 0, indicating a coherence break point after sentence 3.
[0050] In an embodiment, the coherence improvement feedback generation component (112) creates constraint-specific actionable recommendations based on detected coherence break points. The system performs feature-wise analysis at each break point to identify which specific constraints contributed most to the coherence degradation. Based on this constraint-based evaluation, the system generates targeted recommendations: for low POS similarity, it suggests maintaining consistent grammatical structures; for entity shifts, it recommends employing pronouns or repeated references; for tense shifts, it advises maintaining temporal consistency; for high grammar error counts, it provides specific corrections; and for missing discourse markers, it suggests appropriate transitional phrases.
[0051] In an embodiment, the user interface presentation component (113) displays the coherence classification results and improvement feedback in a visually interpretable format. The interface presents the original text with color-coded highlighting to indicate coherence break points, with different colors representing different types of coherence issues (e.g., red for semantic disconnects, yellow for topic shifts, blue for grammatical inconsistencies). A coherence score is displayed alongside explanatory metrics that break down the assessment into component factors such as entity continuity, topic consistency, and semantic flow. Improvement recommendations are listed with specific examples of how revised text might appear, allowing users to compare original and improved versions. The interface allows users to interactively explore different aspects of coherence analysis, toggling between views that emphasize different feature types to understand how each contributes to the overall assessment.
[0052] FIG. 2 illustrates an exemplary flowchart of the automated text coherence analysis method, in accordance with an embodiment of the present disclosure.
[0053] As illustrated in FIG. 2, the operational flow (200) of the automated text coherence analysis method begins with receiving text input (202). This input undergoes preprocessing to normalize formatting, correct minor typographical errors, and prepare the text for feature extraction. Preprocessing may include sentence segmentation to identify sentence boundaries, tokenization to split text into words or subwords, part-of-speech tagging to assign grammatical categories to words, lemmatization to reduce words to their base forms, and stopword removal to eliminate common words with limited semantic value.
[0054] In an embodiment, the preprocessed text then undergoes feature extraction through three parallel processes: generating entity grid data (204), extracting topic representations (206), and obtaining contextual embeddings (208). Each process captures different aspects of coherence as described in the system architecture.
[0055] In an embodiment, the entity grid data generation process (204) identifies entities and their grammatical roles, constructs the entity grid, and calculates transition probabilities as previously described. This process may be enhanced through coreference resolution to identify when different noun phrases refer to the same entity (e.g., recognizing that "the student," "she," and "Sarah" may all refer to the same person), improving the accuracy of entity tracking across sentences. The process may also incorporate entity salience detection to give greater weight to entities that play central roles in the discourse, as transitions involving these entities may have greater impact on perceived coherence.
[0056] In an embodiment, the topic representation extraction process (206) applies LDA to discover latent topics, constructs a graph based on topic similarities, and processes this graph employing GCNs as previously described. This process may be enhanced through dynamic topic modeling techniques that account for how topics evolve throughout a text, capturing narrative progression and thematic development. The process may also incorporate hierarchical topic modeling to identify both broad thematic areas and specific subtopics, providing a more nuanced understanding of thematic structure.
[0057] In an embodiment, the contextual embedding generation process (208) utilizes transformer-based language models to create vector representations of text as previously described. This process may be enhanced through fine-tuning of language models on domain-specific corpora (e.g., academic writing, technical documentation, creative fiction) to better capture semantic relationships relevant to particular genres. The process may also incorporate cross-sentence attention mechanisms that explicitly model relationships between sentences, strengthening the representation of discourse-level coherence.
[0058] In an embodiment, the extracted features are then processed through the multi-feature meta-learning framework (210) to produce coherence classification results. This framework combines the strengths of diverse machine learning approaches through a two-level architecture as previously described. The framework may be enhanced through ensemble diversity optimization techniques that select base models to maximize complementarity, ensuring that each model contributes unique information to the meta-learning process. The framework may also incorporate model confidence weighting to give greater influence to base models that demonstrate higher confidence in their predictions for specific inputs.
[0059] Based on the coherence classification, the system detects coherence break points (212) within the text employing the constraint-based analytical approaches previously described. The detection process calculates constraint values for each sentence pair, computes the weighted sum of all constraints, compares the weighted sum against the threshold, and identifies sentence break points where coherence degrades, such as after sentence 3 when the score falls to -0.494.
[0060] Employing the detected break points, the system generates improvement feedback (214) through feature-wise analysis mechanisms previously described. The feedback generation process performs constraint-specific analysis at break points, generates targeted recommendations based on which constraints failed, and creates split or reorganization suggestions to address the identified issues.
[0061] In an embodiment, the system displays the results through a user interface (216) as previously described. This interface presentation may be enhanced through progressive disclosure that initially presents high-level results and allows users to drill down into specific areas of interest, preventing information overload. The interface may also incorporate interactive editing capabilities that allow users to implement recommended changes directly within the interface and see real-time updates to coherence assessments as they revise their text.
[0062] In an embodiment, the system implements Local Interpretable Model-Agnostic Explanations (LIME) to enhance explainability of coherence classifications. The LIME process begins by perturbing the input text to create multiple variations with slight modifications to features. These variations are then classified by the coherence classification system, and a local linear model is trained to approximate how the classification system behaves in the neighborhood of the original input. This linear model identifies which features most strongly influence the classification outcome, providing insight into the specific aspects of the text that contribute to its coherence assessment. The results are presented visually, showing the relative importance of different features with color coding and numerical weights. This explainability component helps users understand not just what coherence level their text achieves, but why it receives that assessment, facilitating more targeted revisions.
[0063] In an embodiment, the system includes a retraining mechanism for continuous improvement. This mechanism collects feedback from users on coherence classifications and improvement recommendations, including confirmations or corrections of assessments and indications of which recommendations were found useful. This feedback is used to create an expanded training dataset that includes challenging cases and edge conditions that might not be well represented in the original training data. The system periodically retrains both first-level machine learning models and the second-level meta-model employing this expanded dataset, employing cross-validation techniques to ensure robust performance. This retraining process may include curriculum learning approaches that gradually introduce more complex examples, allowing the models to develop increasingly sophisticated understanding of coherence patterns.
[0064] In an embodiment, the system includes domain adaptation capabilities to customize coherence assessment for specific genres or fields. This adaptation process begins with the collection of domain-specific corpora representing exemplary texts within particular genres (e.g., scientific research papers, legal documentation, creative narrative). Features extracted from these corpora are used to fine-tune the coherence analysis models, adjusting parameters to reflect domain-specific conventions and expectations regarding text organization and flow. Domain-specific coherence markers may be identified and incorporated into the analysis, such as the use of technical terminology in scientific writing or narrative point-of-view consistency in fiction. This adaptation enables more accurate and relevant coherence assessment across diverse writing contexts, accounting for how coherence conventions may vary between domains.
[0065] In an embodiment, the system supports multi-lingual coherence analysis through language-specific modeling and cross-lingual transfer learning. For each supported language, language-specific models are developed for entity grid analysis, topic modeling, and contextual embedding generation, accounting for grammatical and discourse differences between languages. Cross-lingual embeddings and alignment techniques enable knowledge transfer between high-resource and low-resource languages, allowing the system to leverage patterns learned from languages with abundant training data to improve performance on languages with limited data. Language-specific coherence conventions are incorporated into the assessment criteria, recognizing that features such as entity transitions, topic flow, and semantic relationships may manifest differently across languages due to varying syntactic structures and discourse norms.
[0066] In an embodiment, the system integrates with writing assistance tools to provide real-time coherence feedback during the composition process. This integration involves implementing lightweight coherence analysis algorithms that can operate efficiently on partial or in-progress texts, providing immediate feedback without requiring full document processing. Progressive coherence assessment examines relationships between the current sentence being written and previous content, highlighting potential coherence issues as they emerge rather than after document completion. Predictive coherence guidance suggests potential next sentences or transitions that would maintain or enhance coherence based on the current text state. This real-time assistance helps writers address coherence issues during initial composition rather than through later revision, potentially improving writing efficiency and quality.
[0067] In an embodiment, the system includes comparative coherence analysis capabilities for evaluating multiple versions of a text or comparing a text to reference standards. This comparative analysis aligns corresponding segments between text versions and identifies changes in coherence metrics, highlighting improvements or regressions in specific areas. Version tracking visualizations show coherence evolution across document drafts, helping users understand how their revisions impact overall text flow and organization. Benchmark comparison features evaluate a text's coherence against domain-specific reference standards derived from exemplary texts, providing context for interpretation of coherence scores. This comparative capability enables more nuanced assessment of writing development and revision effectiveness.
[0068] In an embodiment, the system implements cognitive load estimation to evaluate the processing demands that a text places on readers. This estimation combines multiple factors including sentence complexity, information density, conceptual difficulty, and coherence quality to predict the cognitive resources required for comprehension. Cognitive load maps visualize how processing demands vary throughout a text, identifying sections that may require excessive mental effort. Targeted simplification recommendations suggest revisions to reduce cognitive load while maintaining informational content, such as splitting complex sentences, clarifying references, or strengthening thematic transitions. This capability helps writers optimize their text not just for technical coherence but for actual readability and comprehensibility.
[0069] The present invention has been described with reference to specific embodiments. However, it will be appreciated that various modifications and changes can be made without departing from the scope of the invention as set forth in the claims. Accordingly, the specification and figures are to be regarded as illustrative rather than restrictive, and modifications are intended to be included within the scope of the present invention. All such variations and modifications are intended to be included within the scope of the invention as defined in the appended claims.
EXAMPLES
Example 1: Implementation and Evaluation of the Automated Text Coherence Analysis System
[0070] The automated text coherence analysis system was implemented and evaluated employing a dataset of 515 paragraphs extracted from undergraduate computer science student reports. The paragraphs were manually classified into three coherence levels: Highly Coherent (254 paragraphs), Moderately Coherent (210 paragraphs), and Incoherent (51 paragraphs). This classification served as the ground truth for evaluating system performance.
Table 1 shows the distribution of paragraphs across coherence classes in the dataset used for system evaluation:
Table 1: Distribution of paragraphs across coherence classes
Coherence Class Label Paragraph Count
Highly Coherent 0 254
Moderately Coherent 1 210
Incoherent 2 51
Total - 515

[0071] The system was configured to generate entity grid data by tracking grammatical role transitions across sentences. Table 2 illustrates an example entity grid for a three-sentence paragraph, showing how entities transition between roles:
Table 2: Entity grid representation for a sample paragraph
Sentence machine learning algorithms datasets models accuracy
S1 S O X X -
S2 - - S S X
S3 - S - O O

[0072] From this entity grid, transition probabilities were calculated as shown in Table 3:
Table 3: Transition counts and probabilities from the entity grid
Transition Count Probability Entity Example
S → S 1 0.083 datasets
S → O 1 0.083 models
S → X 1 0.083 algorithms
S → – 1 0.083 machine learning
O → S 1 0.083 algorithms
O → O 1 0.083 models
O → X 0 0.000 -
O → – 0 0.000 -
X → S 0 0.000 -
X → O 0 0.000 -
X → X 0 0.000 -
X → – 1 0.083 machine learning
– → S 1 0.083 algorithms
– → O 1 0.083 accuracy
– → X 1 0.083 accuracy
– → – 2 0.167 machine learning, datasets

[0073] A comprehensive evaluation assessed the performance of various feature combinations and modeling approaches. Table 4 summarizes the performance of different system configurations:
Table 4: Performance comparison of different coherence classification models
System Model Dimension SMOTE Test F1 Test Accuracy
Entity Transitions + Stacking KNN 16 With 90.0 90.0
BERT + Entity Transitions + Stacking KNN 784 With 92.7 92.7
LDA with GCN - - Without 35.7 45.7
Sequential Model with Entity Transitions GRU 9x16 Without 31.4 46.5
Multi-Channel Input Fusion Dense Network - With 38.0 40.0
BERT + Stacking AdaBoost 768 With 90.0 90.0
LDA Topic Features + Stacking SVM 15 With 78.7 78.8
T5 + Stacking KNN 768 With 89.4 89.4
Normalized Entity Transitions + Stacking MLP 9x16 With 78.7 78.5
Multi-Feature Meta-Learning Framework LGBM 30 Without 92.2 92.3
Double Meta-Learning LGBM 50 Without 92.2 92.2

[0074] The initial coherence break point detection employing cosine similarity revealed unexpected results. Table 5 shows similarity scores between consecutive sentences in paragraphs of different coherence levels:
Table 5: Sentence similarities of paragraphs with different coherence levels
sim_0_1 sim_1_2 sim_2_3 sim_3_4 Label
0.62625 0.536287 0.576211 0.843517 1 (coherent)
0.946134 0.917875 0.832921 0.863998 2 (moderately coherent)
0.975551 0.909581 0.949544 0.988289 3 (incoherent)

[0075] This analysis revealed that contrary to initial expectations, coherent paragraphs sometimes showed lower average cosine similarity scores than less coherent ones. This finding emphasized the limitations of relying solely on semantic similarity and led to the development of the constraint-based feedback mechanism.
[0076] To address these limitations, the enhanced feedback mechanism was implemented employing constraint-based analysis. An exemplary paragraph demonstrating coherence issues was analyzed to illustrate the system's capabilities:
"Next, we perform similarity scoring, which allows to measure the degree of similarity between different users based on their characteristics and behaviours. This scoring helps in classifying the data into distinct classes and labels, effectively grouping users with similar traits. By visualizing these trends and patterns, we can gain valuable insights into user behaviour on the IRCTC platform, ultimately enhancing the overall user experience and informing targeted marketing strategies. The user inputs his information in the website using the user interface provided by IRCTC. The back-end service of the application stores the information given by the user in a storage unit. The back-end also processes several requests by the user and in the process stores additional information on Fig 2 System Architecture the user."
[0077] In this exemplary paragraph, the first three sentences discuss similarity scoring, user behavior analysis, and strategic insights, maintaining thematic coherence around data analysis and user behavior. However, the fourth sentence abruptly shifts to describing user interface and backend workflow, which pertains to system architecture rather than behavior analysis, creating a coherence break point that the system successfully identifies.
[0078] Table 6 demonstrates the feature-wise analysis for a paragraph where coherence issues were detected:
Table 6: Feature-wise Analysis and Coherence Score Between Consecutive Sentences
Sentence Index POS Similarity Tense Shift Embedding Similarity Entity Shift Pronoun Shift Length Difference Grammar Errors S1 Grammar Errors S2 Discourse Marker Coherence Score
1→2 0.818 1 0.585 1 1 5 1 0 0 0.138
2→3 0.750 1 0.361 0 1 11 0 1 1 0.050
3→4 0.583 1 0.579 1 1 14 1 2 0 -0.494
4→5 0.750 1 0.411 1 0 2 2 0 0 0.197
5→6 0.636 1 0.649 1 0 6 0 0 1 0.292

[0079] The constraint-based analysis of the exemplary paragraph revealed specific coherence degradation patterns. The system identified that sentences 1-3 maintained topical coherence with similarity scores above the threshold, discussing user behavior analysis and strategic insights. However, the transition from sentence 3 to sentence 4 showed a weighted coherence score of -0.494, falling below the threshold of 0. This break point occurred due to multiple constraint violations: (i) entity shift from "user behaviour" and "insights" to "user interface" and "back-end service," (ii) topic similarity decrease from behavior analysis to system architecture, (iii) absence of discourse markers to signal the topic transition, and (iv) grammatical inconsistencies in sentence structure. The system generated targeted feedback recommending paragraph splitting after sentence 3, insertion of transitional phrases such as "From a technical implementation perspective" to bridge the thematic gap, and reorganization of content to group system architecture details in a separate paragraph.
[0080] Employing a threshold of 0, the constraint-based analysis successfully identified a coherence break point after sentence 3, where the weighted sum score of -0.494 indicated significant coherence degradation. The analyzed paragraph exhibited a topic shift from user behavior analysis and strategic insights (sentences 1-3) to system architecture and UI description (sentences 4-5). The system generated constraint-specific feedback: recommending paragraph splitting after sentence 3, suggesting transition phrases to bridge the thematic gap, and highlighting grammatical corrections needed based on the detected errors.
[0081] When deployed in a classroom environment for a semester-long technical writing course, the constraint-based system demonstrated superior performance over similarity-only approaches when analyzing paragraphs similar to the IRCTC platform example. Student surveys indicated that 76% of users reported improved understanding of coherence principles after employing the system, with particular appreciation for the system's ability to detect subtle topic shifts between analytical content and technical implementation details. Instructors noted a 23% reduction in time spent providing structural feedback on writing assignments, allowing more focus on content-related guidance. The system proved especially effective in identifying coherence breaks in technical writing where students frequently transition between conceptual analysis and system implementation without appropriate discourse markers or logical organization. The constraint-based approach proved especially effective in identifying subtle coherence issues that traditional similarity measures missed.

ADVANTAGES OF THE PRESENT DISCLOSURE
[0082] The present disclosure offers a comprehensive multi-feature approach to coherence analysis that eliminates the limitations of single-feature methods, making coherence assessment significantly more accurate and reliable for educational and professional writing contexts.
[0083] The present disclosure provides specific, actionable feedback through automated break point detection that identifies exact locations of coherence issues, enabling targeted improvements rather than general recommendations that leave writers uncertain about how to enhance their text.
[0084] The present disclosure offers explainable coherence evaluations through Local Interpretable Model-Agnostic Explanations, maximizing educational value by helping writers understand specific factors influencing coherence assessment, unlike black-box approaches that provide scores without explanatory context.
, Claims:1. A system for automated text coherence analysis (100), comprising:
a processor (101); and
a memory (102) storing computer-executable instructions that, when executed by the processor (101), cause the system (100) to:
receive a text input (103) comprising paragraphs with sentences;
generate entity grid data (104) representing entity continuity patterns from the text input (103);
extract topic representations (105) from the text input (103) employing Latent Dirichlet Allocation and graph convolutional networks;
obtain contextual embeddings (106) from the text input (103) employing a transformer-based language model;
implement a multi-feature meta-learning framework (107) employing the entity grid data (104), the topic representations (105), and the contextual embeddings (106), wherein the framework comprises first-level machine learning models (108) and a second-level meta-model (109) to generate a coherence classification output (110);
detect coherence break points (111) in the text input (103) employing a constraint-based analysis comprising binary indicators for entity shift, pronoun shift, tense shift, grammatical errors, and discourse markers, and continuous measures for part-of-speech similarity, embedding similarity, and topic similarity, wherein the constraints are combined employing a weighted sum compared against a threshold;
generate coherence improvement feedback (112) for the detected coherence break points (111), wherein the feedback comprises constraint-specific recommendations based on feature-wise analysis at each detected break point; and
display the coherence classification output (110) and the coherence improvement feedback (112) through a user interface (113).
2. The system for automated text coherence analysis (100) as claimed in claim 1, wherein the first-level machine learning models (108) of the multi-feature meta-learning framework (107) comprise at least five models selected from a group consisting of: support vector machine, random forest, k-nearest neighbors, logistic regression, multi-layer perceptron, decision tree, naïve Bayes, gradient boosting, AdaBoost, and XGBoost.
3. The system for automated text coherence analysis (100) as claimed in claim 1, wherein detecting the coherence break points (111) employing the constraint-based analysis further comprises:
calculating a weighted sum of the binary indicators and continuous measures for each pair of consecutive sentences in the text input (103), wherein weights are assigned based on relevance to coherence assessment;
comparing the weighted sum against the predefined threshold value;
identifying sentence pairs where the weighted sum falls below the threshold as the coherence break points (111); and
determining specific constraints from the binary indicators and continuous measures that contributed most to the coherence degradation at each detected break point.
4. The system for automated text coherence analysis (100) as claimed in claim 1, wherein the processor (101) further executes instructions to: apply Local Interpretable Model-Agnostic Explanations (LIME) to the coherence classification output (110) to identify specific features from the entity grid data (104), topic representations (105), and contextual embeddings (106) that contribute to the coherence classification output (110).
5. The system for automated text coherence analysis (100) as claimed in claim 1, wherein generating the coherence improvement feedback (112) based on the detected coherence break points (111) further comprises:
perform feature-wise analysis of the constraints to identify which constraints caused each coherence break point (111);
generate targeted recommendations based on the specific constraint violations identified at each break point;
suggest sentence splitting when the semantic similarity measure falls below the threshold;
recommend transition phrases for detected topic shifts based on the topic similarity measure; and
provide grammatical corrections for detected tense shifts and pronoun shift inconsistencies.
6. A method for automated text coherence analysis, comprising:
receiving, by a processor (101), a text input (103) comprising paragraphs with sentences;
generating, by the processor (101), entity grid data (104) representing entity continuity patterns from the text input (103);
extracting, by the processor (101), topic representations (105) from the text input (103) employing Latent Dirichlet Allocation and graph convolutional networks;
obtaining, by the processor (101), contextual embeddings (106) from the text input (103) employing a transformer-based language model;
implementing, by the processor (101), a multi-feature meta-learning framework (107) employing the generated entity grid data (104), the extracted topic representations (105), and the obtained contextual embeddings (106), wherein the framework comprises first-level machine learning models (108) and a second-level meta-model (109) to generate a coherence classification output (110);
detecting, by the processor (101), coherence break points (111) in the text input (103) employing a constraint-based analysis comprising binary indicators for entity shift, pronoun shift, tense shift, grammatical errors, and discourse markers, and continuous measures for part-of-speech similarity, embedding similarity, and topic similarity, wherein the constraints are combined employing a weighted sum compared against a threshold;
generating, by the processor (101), coherence improvement feedback (112) for the detected coherence break points (111), wherein the feedback comprises constraint-specific recommendations based on feature-wise analysis of the constraints at each detected break point; and
displaying, by the processor (101), the coherence classification output (110) and the generated coherence improvement feedback (112) through a user interface (113).
7. The method for automated text coherence analysis as claimed in claim 6, wherein training the second-level meta-model (109) of the multi-feature meta-learning framework (107) employing the outputs from the first-level machine learning models (108) as input features comprises:
collecting a labeled dataset of reports with human-annotated coherence ratings for training;
processing the labeled dataset through the first-level machine learning models (108) to generate model outputs;
employing the generated model outputs as training data for the second-level meta-model (109); and
optimizing the second-level meta-model (109) employing cross-validation techniques on the training data.
8. The method for automated text coherence analysis as claimed in claim 6, wherein the graph convolutional networks used in extracting the topic representations (105) process the topic probability distributions by:
applying graph convolution operations to node features of the constructed graph, wherein each node represents a sentence from the text input (103);
aggregating information from neighboring nodes in the graph to capture contextual relationships between the sentences;
generating graph embeddings that encode structural information of the topic probability distributions; and
integrating the generated graph embeddings with the generated entity grid data (104) and obtained contextual embeddings (106) for input to the multi-feature meta-learning framework (107).
9. The method for automated text coherence analysis as claimed in claim 6, further comprising: performing, by the processor (101), data augmentation on training data for the first-level machine learning models (108) and the second-level meta-model (109) comprising:
applying Synthetic Minority Over-sampling Technique (SMOTE) to the training data for addressing class imbalance;
normalizing feature vectors extracted from the training data before processing through the first-level machine learning models (108); and
implementing feature selection on the normalized feature vectors to identify discriminative features for the coherence classification.
10. The method for automated text coherence analysis as claimed in claim 6, further comprising: retraining, by the processor (101), the first-level machine learning models (108) and the second-level meta-model (109) employing new labeled report data by:
collecting feedback on the generated coherence classification output (110);
incorporating the collected feedback to create an expanded training dataset; and
periodically updating the models employing the expanded training dataset to improve coherence classification accuracy.