Text Classification Algorithms: A Survey

1. Introduction

• Text classification systems are organized into feature extraction, dimensionality reduction, classifier selection, and evaluation phases across document, paragraph, sentence, and sub-sentence scopes.The paper frames these phases as the structure for reviewing text classification systems.Source: §1 · pp. 1–2

1. Document

• Feature extraction converts cleaned, unstructured text into structured features using weighted-word methods such as TF-IDF and TF or embeddings such as Word2Vec and GloVe.Dimensionality reduction can then reduce time and memory complexity when feature spaces contain many unique words, using methods including PCA, LDA, NMF, random projection, autoencoders, and t-SNE.Source: §1 · p. 2
• The survey covers traditional, ensemble, probabilistic, non-parametric, discriminative, tree-based, graphical, and deep learning classifiers, including SVM as a widely used baseline.It also highlights Naïve Bayes for low computational and memory requirements and deep learning for modeling complex, non-linear relationships.Source: §1 · pp. 2–3
• Evaluation is essential because accuracy is unsuitable for unbalanced datasets, so the survey discusses Fβ Score, MCC, ROC, and AUC.The paper also compares methods by architecture, novelty, feature extraction, corpus, validation measure, and limitations.Source: §1 · p. 3

2. Text Preprocessing

• Text preprocessing removes noise and supports informative featurization through cleaning operations such as tokenization, stop-word removal, capitalization normalization, slang and abbreviation conversion, noise removal, spelling correction, stemming, and lemmatization.Tokenization breaks text into meaningful tokens, while stemming and lemmatization consolidate variant word forms into shared representations.Source: §2 · pp. 3–4, §2.1 · p. 4, §2.1.1 · p. 4

2.1. Text Cleaning and Pre-processing

• Preprocessing choices can improve classification but may introduce tradeoffs, such as lowercasing collapsing “US” and “us” or removing punctuation that carries human-interpretable information.Spelling correction is optional and can use hashing-based, context-sensitive, Trie, or Damerau–Levenshtein methods; slang may be converted into formal language.Source: §2.1.3 · p. 4, §2.1.4 · p. 4, §2.1.5 · p. 4

2.2. Syntactic Word Representation

• Syntactic word representation addresses the loss of word-order and semantic relations by using n-grams as features, while BOW represents words as unordered 1-gram vectors.Common 2-gram and 3-gram features preserve more information than 1-grams by representing words occurring in sequence.Source: §2.2 · p. 5, §2.2.1 · p. 5

An Example of 2-Gram

• A 2-gram representation of “After sleeping for four hours, he decided to sleep for another four” consists of adjacent word pairs such as “After sleeping,” “sleeping for,” and “for four.”The example illustrates that n-grams are extracted as ordered token sequences rather than complete text representations.Source: An Example of 2-Gram · p. 5

An Example of 3-Gram

• A 3-gram represents text through overlapping sequences of three tokens, such as “After sleeping for” and “sleeping for four.”Syntactic n-grams instead follow paths in syntactic dependency or constituent trees rather than the text’s linear structure.Source: An Example of 3-Gram · p. 5, §2.2.2 · p. 6

2.3. Weighted Words

• TF maps each word to its corpus occurrence count and represents each document as a frequency vector, but common language words may dominate the representation.TF-based weighting methods can also use boolean or logarithmic frequency scaling.Source: §2.3 · p. 6

2.3.1. Bag of Words (BoW)

• Bag-of-words represents a document as selected word-frequency features, ignoring grammar, word order, and semantic relationships while retaining multiplicity.The representation is used in applications including NLP, Bayesian spam filtering, document classification, information retrieval, and computer vision.Source: §2.3.1 · p. 6

2.3.2. Limitation of Bag-of-Words

• Bag-of-words models face scalability and order limitations because vocabulary-sized one-hot vectors can reach millions of dimensions and reorderings receive identical representations.For example, “This is good” and “Is this good” have the same vector representation.Source: §2.3.2 · p. 6

2.3.3. Term Frequency-Inverse Document Frequency

• TF-IDF combines term frequency with inverse document frequency to reduce the influence of common terms, but it cannot represent similarity between independently indexed words.Word embeddings address this limitation by incorporating word similarity and part-of-speech information.Source: §2.3.3 · p. 7

2.4.1. Word2Vec

• Word2Vec uses CBOW and Skip-gram architectures to learn high-dimensional word vectors that capture relationships and similarities between words.CBOW predicts a target word from context, whereas Skip-gram predicts surrounding words from a given word; embeddings can place “big” and “bigger” close together.Source: §2.4.1 · p. 8, Figure 2 · p. 8

Continuous Bag-of-Words Model

• CBOW predicts a target word from multiple context words, with the context window determining how many words contribute to the prediction.The model creates a vocabulary of unique corpus words and commonly uses a window size of 4–5 words.Source: Continuous Bag-of-Words Model · p. 8

Continuous Skip-Gram Model

• Continuous Skip-gram maximizes classification of a word based on another word in the same sentence, while continuous bag-of-words and continuous Skip-gram preserve syntactic and semantic information of sentences.However, preserving full meaning of coherent documents remains an issue for machine learning; Figure 4 shows per-sentence feature extraction fails, including references from “This day” to “July 4th” and “She” to “Maryam”.Source: Continuous Skip-Gram Model · p. 8, §2.5 · pp. 10–11

3. Dimensionality Reduction

• Dimensionality reduction reduces the feature-space size of term-based text models, addressing their expensive time complexity and memory consumption.Source: §3 · p. 11

3.1.1. Principal Component Analysis (PCA)

• PCA identifies an uncorrelated subspace that maximizes preserved variance, while supporting preprocessing, noise reduction, and overfitting avoidance; KPCA extends it to nonlinear cases.Source: §3.1.1 · pp. 11–12

3.1.2. Independent Component Analysis (ICA) / 3.2. Linear Discriminant Analysis (LDA)

• ICA models observed data as linear transformations of independent components, whereas LDA reduces dimensions using class information and between-class versus within-class variance ratios [72, 73, 74, 75, 76, 77].Source: §3.2 · p. 12, §3.1.2 · p. 12

3.3. Non-Negative Matrix Factorization (NMF)

• NMF compresses a nonnegative n × m matrix V into WH when (n + m)r < nm, and projects training and test documents into an r-dimensional space.Source: §3.3 · pp. 13–14

3.4. Random Projection / 3.4.1. Random Kitchen Sinks

• Random projection targets high-volume or high-dimensional text data; random kitchen sinks approximate shift-invariant kernels by Monte Carlo sampling.Source: §3.4 · p. 14, §3.4.1 · p. 14

3.4.2. Johnson Lindenstrauss Lemma

• The Johnson–Lindenstrauss lemma maps n-point Euclidean data to k = O(logn/ϵ^2) dimensions while bounding pairwise distances within multiplicative factors of 1 − ϵ and 1 + ϵ.Source: §3.4.2 · p. 15

3.5. Autoencoder / 3.5.1. General Framework / 3.5.2. Conventional Autoencoder Architecture / 3.5.3. Recurrent Autoencoder Architecture

• Autoencoders reduce feature dimensions through a hidden layer with p < n units, with conventional CNN and recurrent architectures providing encoding and decoding for text, document, and sequence data [91, 92, 93, 94, 95, 97, 100].The general framework reduces n dimensions to p dimensions, while CNN autoencoders reconstruct through convolutional feature maps and recurrent autoencoders process sequences iteratively.Source: §3.5 · pp. 16–17, §3.5.1 · p. 17, §3.5.2 · pp. 17–18

3.6. T-distributed Stochastic Neighbor Embedding (t-SNE)

• t-SNE embeds high-dimensional data for low-dimensional visualization by converting Euclidean distances into conditional similarities and optimizing a KL-divergence-based cost.Source: §3.6 · p. 19

4. Existing Classification Techniques

• Existing techniques include Rocchio, boosting, bagging, logistic regression, Naïve Bayes, k-nearest neighbor, SVMs, tree-based methods, neural networks, and combination techniques.Tree-based algorithms are described as fast and accurate for document categorization, while neural methods include DNNs, CNNs, RNNs, DBNs, and HANs.Source: §4 · pp. 20–21

4.1. Rocchio Classification

• Rocchio represents words with TF-IDF weights, averages training-document vectors into a prototype for each class, and assigns test documents to the most similar prototype.The algorithm originated as a relevance-feedback method for full-text databases and was later developed for text and document classification [104–107].Source: §4.1 · p. 21
• Rocchio predicts a document’s label using its Euclidean distance from class centroids, which may be normalized to unit length before label assignment.Each centroid is computed from the weighted vectors of documents belonging to its class.Source: §4.1 · p. 21

Limitation of Rocchio Algorithm

• Rocchio’s limitation is that users can retrieve only a few relevant documents, although its results consider semantics.Source: Limitation of Rocchio Algorithm · p. 21

4.2. Boosting and Bagging

• Voting methods including bagging and boosting have been applied successfully to document and text classification, with boosting adapting training distributions while bagging ignores previous classifiers.Source: §4.2 · p. 21

4.2.1. Boosting

• Boosting, introduced by R.E. Schapire and further developed by Freund, combines multiple trained models and led to AdaBoost.AdaBoost weights training instances, evaluates successive classifiers, and outputs an optimal classifier C∗.Source: §4.2.1 · pp. 21–22, Figure 9 · p. 22

4.2.2. Bagging

• Bagging trains classifiers on bootstrap samples and predicts the class receiving the most subclassifier votes, but it increases computational complexity and reduces interpretability.The loss of interpretability prevents feature importance from being discovered in these models.Source: §4.2.2 · p. 23, §4.2.3 · p. 23

4.3. Logistic Regression

• Logistic regression predicts class probabilities with a linear decision boundary, extends to multinomial classes through softmax, and assumes independent data points.Its instance-based combination uses distance and similarity functions, while multinomial parameters can be estimated by maximum likelihood or maximum a posteriori methods.Source: §4.3 · p. 23, §4.3.2 · p. 24, §4.3.3 · p. 25

4.4. Naïve Bayes Classifier

• Naïve Bayes is a generative, widely used document-classification baseline that commonly uses term-frequency bag-of-words features and multinomial word probabilities.It performs poorly with unbalanced classes, although class normalization with a centroid classifier produces experimental results very similar to the centroid classifier; it also assumes a particular data-distribution shape and requires likelihood estimates for possible feature values.Source: §4.4 · p. 25, §4.4.1 · p. 26, §4.4.2 · p. 26

4.5. K-Nearest Neighbor

• KNN assigns a test document to the highest-scoring class among its k nearest training neighbors, while weighted KNN learns feature weights for classification.KNN is easy to implement and naturally supports multiclass cases, but large searches require substantial storage and performance depends on a meaningful distance function.Source: §4.5.1 · p. 27, §4.5.2 · p. 28, §4.5.3 · p. 28

4.6. Support Vector Machine (SVM)

• SVM supports linear and nonlinear classification, extends binary models to multiclass schemes such as One-vs-One and All-vs-One, and uses word-sequence, TF-IDF, or string-kernel features.String kernels map strings into feature spaces and can support text categorization, while stacking SVM creates hierarchical classifiers and MIL labels bags positive when at least one instance is positive.Source: §4.6 · p. 28, §4.6.2 · p. 29, §4.6.3 · p. 30

4.6.6. Limitation of Support Vector Machine (SVM)

• SVM text classifiers can lack transparency in high-dimensional settings, and string-kernel classification is additionally limited by time complexity.These limitations make functional interpretations difficult and constrain computational efficiency.Source: §4.6.3 · p. 30, §4.6.6 · p. 31

4.7. Decision Tree

• Decision trees hierarchically decompose the data space and select parent attributes by maximizing information gain, thereby reducing entropy after partitioning the training set.The central construction challenge is choosing which features belong at parent and child levels.Source: §4.7 · p. 32

Limitation of Decision Tree Algorithm

• Decision trees learn and predict quickly but are sensitive to small data perturbations, prone to overfitting, and problematic for out-of-sample prediction.Validation and pruning may mitigate these effects, although their effectiveness remains uncertain.Source: Limitation of Decision Tree Algorithm · p. 32

4.8. Random Forest

• Random forests generate parallel random decision trees as an ensemble method for text classification.The technique was introduced by T. K. Ho and further developed by L. Breiman.Source: §4.8 · p. 32, Figure 14 · p. 33

4.8.1. Voting

• After training the forest, random-forest predictions are assigned by voting.The voting procedure determines the final prediction from the individual trees’ outputs.Source: §4.8.1 · p. 33

4.8.2. Limitation of Random Forests

• Random forests train quickly on text datasets relative to deep learning but predict slowly after training.Reducing the number of trees can speed prediction because larger forests increase prediction-time complexity.Source: §4.8.2 · p. 33

4.9. Conditional Random Field (CRF)

• CRFs combine classification with graphical modeling to represent multivariate data and exploit high-dimensional feature spaces for text prediction.They model P(Y|X), the conditional probability of a label sequence given observations, rather than the joint probability P(X, Y), allowing complex observation features without violating the independence assumption [174,175].Source: §4.9 · p. 33
• CRFs compute sequence probabilities through clique potentials whose non-negative values are combined across graph cliques and normalized by Z [174].The feature-based potential function uses a weight vector w associated with feature vector f.Source: §4.9 · pp. 33–34

Limitation of Conditional Random Field (CRF)

• CRF training has high computational complexity, especially for text datasets with large feature spaces, and CRF does not perform with unseen words.These limitations concern both computational cost and words absent from the training sample.Source: Limitation of Conditional Random Field… · p. 34

4.10. Deep Learning

• Deep learning models achieve state-of-the-art results across many domains, including diverse natural-language-processing applications.For text and document classification, the survey presents three basic deep-learning architectures in parallel.Source: §4.10 · p. 34

4.10.1. Deep Neural Networks

• DNNs learn through sequentially connected hidden layers, using features such as TF-IDF or word embeddings as input and class-sized outputs for classification.They are discriminatively trained with back-propagation, using sigmoid or ReLU activations and Softmax for multiclass outputs [2].Source: §4.10.1 · pp. 34–35, Figure 16 · p. 35

4.10.2. Recurrent Neural Network (RNN)

• RNNs weight prior sequence data, supporting semantic analysis of text through LSTM or GRU architectures, but they are vulnerable to vanishing and exploding gradients.Their inputs can be word embeddings processed through hidden layers to an output layer.Source: §4.10.2 · pp. 35–36, Figure 17 · p. 36

• LSTM was introduced by S. Hochreiter and J. Schmidhuber and later augmented by other researchers.Source: Long Short-Term Memory (LSTM) · p. 36
• LSTM preserves long-term dependencies more effectively than basic RNNs and is particularly useful for overcoming the vanishing-gradient problem.Source: Long Short-Term Memory (LSTM) · p. 36
• Multiple gates regulate how much information enters each LSTM node state within its chain-like, RNN-inspired architecture.Source: Long Short-Term Memory (LSTM) · p. 36
• The LSTM equations define input-gate activation, candidate memory, forget-gate activation, updated memory, and final output-gate values.Source: Long Short-Term Memory (LSTM) · p. 36
• Bias vectors, weight matrices, and x_t denote the model parameters and input at time t, while i, c, f, and o index the four gate or memory components.Source: Long Short-Term Memory (LSTM) · p. 37
• CNNs address the bias toward later words in RNNs by using max-pooling to identify discriminative phrases in text data.Source: Long Short-Term Memory (LSTM) · p. 37

Gated Recurrent Unit (GRU)

• GRUs simplify LSTMs by using two gates without internal memory or a second non-linearity.The gates are the update and reset gates, while the output uses a hyperbolic tangent function.Source: Gated Recurrent Unit (GRU) · p. 37, Figure 18 · p. 37

4.10.3. Convolutional Neural Networks (CNN)

• CNNs classify text through word embeddings, one-dimensional convolutions and pooling, fully connected layers, and an output layer.Max pooling selects the maximum element in each pooling window, while stacked feature maps are flattened before subsequent layers.Source: §4.10.3 · pp. 37–38, Figure 19 · p. 38
• CNN text classification faces high dimensionality because the number of input channels can reach 50 K, unlike the few channels typical in image classification [189].During back-propagation, both weights and feature-detector filters are adjusted.Source: §4.10.3 · p. 38

4.10.4. Deep Belief Network (DBN)

• DBNs are deep structures superposed by restricted Boltzmann machines that learn probability distributions over samples, typically trained with contrastive divergence.The subsection also presents their energy, unit-configuration, logistic, and gradient-descent update formulations.Source: §4.10.4 · pp. 38–39

4.10.5. Hierarchical Attention Networks (HAN)

• HANs support document-level classification by applying word encoding and attention at the lower level and sentence encoding and attention at the upper level [194].The document is represented as L sentences, each containing Ti words.Source: §4.10.5 · p. 39, Figure 20 · p. 39

4.10.6. Combination Techniques

• Researchers combine or concatenate standard deep-learning architectures to develop novel techniques intended to provide more robust and accurate classification architectures.The subsection describes recent and popular combined deep-learning architectures and structures.Source: §4.10.6 · p. 40

Architecture

• RMDL randomly generates the layers and nodes of multiple DNN, CNN, and RNN classifiers, enabling heterogeneous model ensembles for classification.For example, nine models may comprise three CNNs, three RNNs, and three DNNs, each uniquely created through randomization.Source: Random Multimodel Deep Learning (RMDL) · p. 40, Figure 21 · p. 40

Prediction aggregation

• The final RMDL prediction is computed by applying majority vote to the outputs of the trained random models.This aggregation defines the final prediction for each data point from the individual model outputs.Source: Random Multimodel Deep Learning (RMDL) · pp. 40–41

Optimization

• Using multiple optimizers allows RMDL to ignore inefficient models when n > k and adapt better to different datasets.The framework can use optimizers such as SGD, Adam, RMSProp, Adagrad, and Adamax, although the evaluation used Adam and RMSProp.Source: Random Multimodel Deep Learning (RMDL) · p. 41
• SGD updates parameters using momentum on a rescaled gradient.The paper presents SGD’s basic equation and its momentum-based parameter-update equation.Source: Random Multimodel Deep Learning (RMDL) · p. 41
• RMSProp divides each weight’s learning rate by a running average of recent gradient magnitudes, but lacks bias correction for sparse gradients.The absence of bias correction causes significant problems when gradients are sparse.Source: Random Multimodel Deep Learning (RMDL) · p. 41

Adam Optimizer

• Adam uses estimated first and second gradient moments, m and v, with bias-corrected forms for optimization.The first moment is m_t and the second moment is v_t; their estimates are corrected by factors involving β_1^t and β_2^t.Source: Adam Optimizer · p. 41
• Adam handles non-stationary objective functions while overcoming RMSProp’s sparse-gradient limitation [197].Source: Adam Optimizer · p. 41
• AdaGrad dynamically incorporates data geometry to enable more informative gradient-based learning.It is described as a novel family of sub-gradient methods and an extension of stochastic gradient descent.Source: Adam Optimizer · p. 42
• AdaDelta uses an exponentially decaying average of the gradient as its second moment and updates AdaGrad’s first-order approach [199].It is presented as an updated version of AdaGrad that relies on first-order information.Source: Adam Optimizer · p. 42

Hierarchical Deep Learning for Text (HDLTex)

• HDLTex specializes deep-learning architectures at each document-hierarchy level to address performance drops as the number of classes increases [2].Its architecture includes separate DNN, RNN, and CNN configurations for hierarchical classification.Source: Hierarchical Deep Learning for Text… · p. 42, Figure 22 · p. 43

Other Techniques

• RCNN combines recurrent contextual modeling with CNN-based text representations, while C-LSTM combines CNN phrase-level features with LSTM long-term dependency learning [201].Source: Other Techniques · p. 43

4.10.7. Limitation of Deep Learning

• Deep learning’s accuracy is associated with lower interpretability, while its black-box behavior, greater data requirements, and training complexity limit classification use [202–207].These limitations are especially consequential for applications requiring explanations, such as healthcare.Source: §4.10.7 · pp. 43–44

4.11. Semi-Supervised Learning for Text Classification

• Semi-supervised text classification uses unlabeled data when few labels are available, often combining clustering with methods such as TSVM, EM, translated features, SFE, or fuzzy DBN [210].Source: §4.11 · p. 44

5. Evaluation

• Evaluation comparability is hindered by inconsistent data-collection and train/test protocols, so classifiers are assessed with confusion-matrix measures including accuracy, recall, specificity, precision, and F-measure.Different metrics capture different aspects of classification performance and therefore may not provide identical information.Source: §5 · p. 45, Figure 24 · p. 45

5.1. Macro-Averaging and Micro-Averaging

• Micro-averaging weights every document equally, whereas macro-averaging weights each category equally without accounting for category frequency [217].Source: §5.1 · p. 46

5.2. Fβ Score; 5.3. Matthews Correlation Coefficient (MCC)

• Fβ balances recall and precision, but it does not fully represent the confusion matrix; MCC captures the full matrix and supports uneven class sizes, yet metrics can rank classifiers differently.Source: §5.2 · p. 46, §5.3 · p. 46

5.4. Receiver Operating Characteristics (ROC); 5.5. Area Under ROC Curve (AUC)

• ROC curves plot TPR against FPR but can misrepresent performance under class imbalance, whereas AUC is threshold-independent and invariant to a priori class probabilities [219–221].Evaluation guidance also calls for comparing methods, accounting for collection variability, assessing ranking and binary performance, and studying scalability in large category spaces.Source: §5.4 · pp. 46–47, §5.5 · p. 47

6. Discussion

• Efficient text-classification system design requires understanding similarities, differences, and limitations across pipeline steps.The discussion covers preprocessing, feature extraction, dimensionality reduction, classification algorithms, and evaluation methods.Source: §6 · p. 47

6.1. Text and Document Feature Extraction

• Weighted words count document terms directly, whereas word embeddings learn unsupervised vectors from word occurrence and co-occurrence information.Weighted-word similarity becomes computationally expensive for large vocabularies and misses semantic similarity; embeddings require large corpora and cannot represent out-of-vocabulary words absent from those corpora [224].Source: §6.1 · pp. 47–48
• Contextualized word representations address polysemy by training word meanings from document context, while text cleaning helps handle slang and abbreviations.GloVe, FastText, and Word2Vec can assign one meaning across different sentences, and pre-trained vectors may omit abbreviations and slang.Source: §6.1 · p. 48

6.2. Dimensionality Reduction

• Dimensionality reduction primarily improves computational time and reduces memory complexity, but methods differ in data, interpretability, and dataset-size requirements.PCA is computationally complex, LDA requires labeled data and manually selected components while producing less interpretable features, random projection is faster than PCA but performs poorly on small datasets [227], and autoencoders require more training data.Source: §6.2 · p. 49

6.3.1. Limitations and Advantages

• Existing classifiers present distinct trade-offs involving retrieval scope, semantic modeling, computational complexity, interpretability, data assumptions, robustness, and prediction speed.Rocchio retrieves few relevant documents, classification results motivate incorporating semantics, boosting and bagging can be computationally complex and less interpretable, logistic regression assumes independent data points, and Naïve Bayes assumes a data-distribution shape and suffers from data scarcity.Source: §6.3.1 · p. 50
• KNN is simple and supports multiclass classification but depends on storage and meaningful distance functions, while SVM, trees, forests, and CRF have transparency, overfitting, prediction-time, or unknown-word limitations.Decision trees are fast but sensitive to perturbations and overfitting, random forests predict slowly as tree counts increase, and CRF training is computationally expensive and fails on unknown words.Source: §6.3.1 · p. 50

6.3.2. State-of-the-Art Techniques’ Comparison

• State-of-the-art comparisons organize techniques by architecture, authorship, model, novelty, feature extraction, corpus, validation measure, and limitations.The comparison treats each system as a classifier paired with a feature-extraction method and evaluated using a validation measure.Source: §6.3.2 · p. 52, Table 4 · p. 53, Table 5 · p. 54

7.1.1. Information Retrieval; 7.1.2. Information Filtering

• Information retrieval and filtering use text classification to manage large document collections, select relevant information, and rank documents for users’ information needs.Methods include Naïve Bayes, SVM, decision trees, J48, KNN, IBK, Bayesian inference networks, and vector-space models.Source: §7.1.1 · p. 55, §7.1.2 · p. 55

7.1.3. Sentiment Analysis

• Sentiment classification identifies opinion polarity using supervised methods such as Naive Bayesian classification and SVM, with lexical, frequency, part-of-speech, negation, and syntactic features.The task models opinions associated with an entity and opinion holder.Source: §7.1.3 · pp. 55–56

7.1.4. Recommender Systems

• Content-based recommender systems use user profiles and item descriptions, combining free-text attributes with specified conditions and methods such as decision trees, nearest neighbors, Rocchio, linear classifiers, probabilistic methods, and Naive Bayes.Profiles can be learned from search history, self-reports, or user-specified filters.Source: §7.1.4 · p. 56

7.1.5. Knowledge Management; 7.1.6. Document Summarization

• Text classification supports knowledge management and document summarization by categorizing document-based representations, mining relationships in reports, and extracting important features for single- and multi-document summaries.Nearly 80% of corporate information is described as existing in unstructured textual formats, and summaries may use words or phrases absent from the source document.Source: §7.1.5 · p. 56, §7.1.6 · p. 56

7.2.1. Health

• Healthcare applications address unstructured, ambiguous, and error-prone medical text through medical coding, personalized EHR representations, and development of MeSH and Gene Ontology resources.Patient2Vec uses recurrent neural networks and attention to learn interpretable, personalized longitudinal EHR representations.Source: §7.2.1 · p. 56

7.2.2. Social Sciences; 7.2.3. Business and Marketing

• Text classification is used to study human behavior from SMS, clinical notes, and social media, commonly relying on word-occurrence frequencies and psychologically relevant LIWC categories.Business and marketing applications mine social media to help companies find customers and increase profits.Source: §7.2.2 · p. 57, §7.2.3 · p. 57

7.2.4. Law

• Legal text classification helps retrieve and automatically categorize large volumes of government-generated legal documents for lawyers and clients.The continual creation of legal documents makes categorization a major challenge for the legal community.Source: §7.2.4 · p. 57

8. Conclusions

• The survey covers feature extraction, dimensionality reduction, classification algorithms, evaluation metrics, pipeline limitations, and applications across fields such as law and medicine.Feature methods include TF-IDF, TF, and word embeddings; dimensionality reduction includes PCA, LDA, NMF, random projection, autoencoders, and t-SNE; classifiers include Rocchio, ensemble methods, LR, NBC, KNN, SVM, decision trees, random forests, CRF, and deep learning; evaluation includes accuracy, Fβ, MCC, ROC, and AUC.Source: §8 · pp. 57–58