Natural Language Processing

Natural Language Processing (NLP) encompasses computational methods for analysing, interpreting and generating human language. Early systems relied on handcrafted rules and lexicons, but the rise of statistical techniques in the 1990s introduced probabilistic parsers and n-gram models. In the past decade, deep learning architectures—initially recurrent neural networks and subsequently attention-based Transformers—have transformed the field. Pretrained language models capture vast contextual knowledge and can be fine-tuned for diverse tasks without extensive labelled data. Self-supervised objectives, such as masked-token prediction, underpin this trend. Contemporary research addresses domain adaptation, low-resource scenarios and cross-lingual transfer, while also advancing multilingual understanding, zero-shot inference and model explainability. Applications span machine translation, summarisation, sentiment analysis, information extraction and conversational agents. Emerging directions explore integration with structured knowledge bases, robust dialogue systems, voice-enabled assistants and specialised pipelines for biomedical, legal and technical texts. With global reliance on digital communication, NLP has become indispensable for knowledge discovery, automated support services and enhanced human–computer interaction.

Research from Nature Portfolio

An innovative deep learning framework has been proposed for the automated classification of non-functional requirements in engineering documents. By employing a multi-layer neural architecture that captures long-range dependencies and contextual cues, the system achieves precision and recall exceeding 80 percent across multiple NFR categories—such as performance, security and usability—without manual feature engineering. Another study has elucidated the linguistic markers of deceptive online reviews through the lens of reality monitoring theory. By analysing perceptual, affective and cognitive cues in large-scale review corpora, researchers identified consistent patterns—such as reduced perceptual detail and heightened sentiment extremity—in fabricated postings. These insights refine feature sets for NLP-based fraud detection and inform the design of more transparent and accountable e-commerce review platforms.

Research from all publishers

A zero-shot learning approach for requirements classification has demonstrated that transformer-based models, when equipped with semantic descriptions of target categories, can accurately assign functional, non-functional and security labels without task-specific training examples. This method reduces reliance on extensive annotations while maintaining F1 scores above 0.66 in preliminary evaluations. In the speech domain, self-supervised embeddings from wav2vec2.0 have been integrated into neural architectures to detect disfluencies—such as repetitions, prolongations and filled pauses—in multiple languages and variable-length utterances. The resulting systems outperform traditional feature-based classifiers, particularly in low-resource settings. A further advance employs a purely attention-driven Transformer, “TranStutter”, to classify stuttering events from spectrogram representations. By eschewing convolutional layers, the model achieves accuracies exceeding 85 percent on clinical and podcast datasets, underscoring the versatility of attention mechanisms in capturing temporal speech patterns.

Natural Language Processing publication trend

The graph below shows the total number of articles in natural language processing across all publications each year (not limited to Nature Index journals).

Technical terms

Automatic speech recognition (ASR): A technology that converts spoken language into text by analysing acoustic signals and linguistic patterns.

Self-supervised learning: A training paradigm in which models derive supervisory signals from unlabeled data—such as predicting masked tokens—to learn useful representations without manual annotation.

Transformer: A neural network architecture based on self-attention mechanisms and positional encoding, enabling parallel processing of sequential data and capturing long-range dependencies.

Zero-shot learning: A technique allowing a model to generalise to classes or tasks for which no direct training examples were provided, often by leveraging semantic embeddings or auxiliary descriptions.

Non-functional requirement (NFR): A specification of system qualities or constraints—such as performance, security or maintainability—rather than its core functions.

Speech disfluency: Interruptions in fluent speech, including hesitations, repetitions, interjections and prolongations, which can be automatically detected for clinical or assistive applications.

References

1. A deep learning framework for non-functional requirement classification. Scientific Reports (2024).
2. What makes deceptive online reviews? A linguistic analysis perspective. Humanities and Social Sciences Communications (2023).
3. Zero-shot learning for requirements classification: An exploratory study. Information and Software Technology (2023).
4. Automatic Speech Disfluency Detection Using wav2vec2.0 for Different Languages with Variable Lengths. Applied Sciences (2023).
5. TranStutter: A Convolution-Free Transformer-Based Deep Learning Method to Classify Stuttered Speech Using 2D Mel-Spectrogram Visualization and Attention-Based Feature Representation. Sensors (2023).