BERT — Technology Wiki

Overview

Direct Answer

BERT (Bidirectional Encoder Representations from Transformers) is a transformer-based language model trained on masked language modelling that processes text bidirectionally to generate contextualised word embeddings. Released by Google in 2018, it represents a fundamental shift from unidirectional language models by simultaneously considering both preceding and following tokens when encoding meaning.

How It Works

BERT employs a multi-layer transformer encoder architecture that masks 15% of input tokens during training and learns to predict them using surrounding context. During inference, it produces contextualised embeddings where each token's representation depends on its full sentence context, not just sequential history. The model pretrains on two objectives: masked language modelling and next-sentence prediction, enabling it to capture deep syntactic and semantic relationships.

Why It Matters

The model achieved state-of-the-art results across multiple NLP benchmarks upon release, significantly improving accuracy for tasks like sentiment analysis, named entity recognition, and question answering. Organisations leverage it to reduce development time for language understanding systems and improve performance on domain-specific tasks through fine-tuning rather than training from scratch.

Common Applications

Applications include document classification, semantic similarity measurement, and information extraction in legal and financial document processing. Search engines and chatbot platforms utilise bidirectional representations to improve query understanding and relevance ranking.

Key Considerations

Computational cost during pretraining is substantial, though fine-tuning on task-specific data remains efficient. Bidirectional processing makes the model unsuitable for autoregressive generation tasks; practitioners must select architectures appropriate to their specific use case, whether comprehension or generation-focused.

Cross-References(1)

Natural Language Processing

Language Model

Related in Semantics & Representation

Large Language Model

A neural network trained on massive text corpora that can generate, understand, and reason about natural language.

GPT

Generative Pre-trained Transformer — a family of autoregressive language models that generate text by predicting the next token.

Tokenisation

The process of breaking text into smaller units (tokens) such as words, subwords, or characters for processing by language models.

Language Model

A probabilistic model that assigns probabilities to sequences of words, enabling prediction of the next word in a sequence.

Contextual Embedding

Word representations that change based on surrounding context, capturing polysemy and contextual meaning.

Word2Vec

A neural network model that learns distributed word representations by predicting surrounding context words.

GloVe

Global Vectors for Word Representation — an unsupervised learning algorithm for obtaining word vector representations from aggregated word co-occurrence statistics.

Instruction Tuning

Training a language model to follow natural language instructions by fine-tuning on instruction-response pairs.

RLHF

Reinforcement Learning from Human Feedback — a technique for aligning language models with human preferences through reward modelling.

Grounding

Connecting language model outputs to real-world knowledge, facts, or data sources to improve factual accuracy.

Hallucination Detection

Techniques for identifying when AI language models generate plausible but factually incorrect or unsupported content.

Prompt Injection

A security vulnerability where malicious inputs manipulate a language model into ignoring its instructions or producing unintended outputs.

More in Natural Language Processing

Text Embedding Model

Core NLP

A neural network trained to convert text passages into fixed-dimensional vectors that capture semantic meaning, enabling similarity search, clustering, and retrieval applications.

Topic Modelling

Text Analysis

An unsupervised technique for discovering abstract topics that occur in a collection of documents.

Speech Synthesis

Speech & Audio

The artificial production of human speech from text, also known as text-to-speech.

Extractive Summarisation

Generation & Translation

A summarisation technique that identifies and selects the most important sentences from a source document to compose a condensed version without generating new text.

Reranking

Core NLP

A two-stage retrieval process where an initial set of candidate documents is rescored by a more powerful model to improve the relevance ordering of search results.

Cross-Lingual Transfer

Core NLP

The application of models trained in one language to perform tasks in another language, leveraging shared multilingual representations learned during pre-training.

Abstractive Summarisation

Text Analysis

A text summarisation approach that generates novel sentences to capture the essential meaning of a document, rather than simply extracting and rearranging existing sentences.

Machine Translation

Generation & Translation

The use of AI to automatically translate text or speech from one natural language to another.