Accountability in AI
Organisational responsibility for how AI systems function, make decisions, and impact users.
Resources
Key Concepts, Explained
This section provides concise, structured explanations of key concepts in artificial intelligence, machine translation, data protection, and translation workflows. Each entry links to a full article.
A
Algorithmic Bias
Systematic errors in AI outputs arising from skewed, imbalanced, or prejudiced training data.
API (Application Programming Interface)
A controlled interface enabling software systems to communicate and process requests, including translation queries.
API Key Security
Protection measures preventing unauthorised access or misuse of API credentials used to process translations.
Artificial Intelligence
Computational systems capable of performing tasks that traditionally require human intelligence.
Attention Mechanism
A neural network method that helps AI models focus on the most relevant parts of the input when generating output, improving context handling and translation quality.
B
Backpropagation
A learning process used in neural networks to adjust internal weights after comparing predictions with expected results.
Baseline System
A reference translation system used for comparison when evaluating improvements or alternative models.
Beam Search
A decoding method used in AI language models and neural machine translation to evaluate several candidate word sequences before selecting the most likely output.
BERTScore
A metric comparing contextual embeddings of a machine translation and a human reference.
Bilingual Corpus
A collection of aligned texts in two languages used to train, evaluate, or improve translation systems and language technologies.
BLEU
A traditional translation metric measuring n-gram overlap between machine output and a reference.
C
CAT Tools
Software environments supporting professional translation through TMs, termbases, and quality checks.
chrF++
A character-level machine translation metric evaluating similarity between system output and a reference.
COMET
A neural evaluation metric scoring translation quality based on semantic similarity and model judgments.
Confidentiality in Translation Workflows
Safeguards ensuring that customer texts remain private, protected, and not reused beyond authorised tasks.
Concordance Search
A tool function retrieving previous translations or term occurrences from a translation memory.
Context-aware Translation
Translation methods using surrounding sentences or document-wide information for better accuracy.
Context Window
The amount of text an AI model can process at once during translation or generation.
Continuous Localisation
A workflow where translation, QA, and updates occur continuously alongside product development.
D
Data Privacy
Protection of personal or sensitive information during processing, transfer, and storage.
Data Residency
A requirement for data to remain within a specific legal jurisdiction or region.
Deep Learning
A machine-learning approach using multi-layer neural networks to model complex patterns.
Discourse-level Translation
Translation that accounts for relationships across sentences, paragraphs, and whole documents.
Document-level Machine Translation
Translation approaches that process full documents rather than isolated segments.
E
Encryption
Technical measures ensuring data is secured during transfer and storage.
Ethical AI
AI developed and used in alignment with fairness, safety, and respect for rights.
EU AI Act
The European Union’s regulatory framework governing the development and use of AI.
Explainability (XAI)
Methods making AI model decisions understandable to human users.
F
Fairness and Bias
Principles ensuring AI systems behave without unjustified discrimination.
Federated Learning
A machine learning approach where models are trained across distributed devices or servers without centralising raw data.
File Parsing
Automated extraction of text and structure from formats such as DOCX, PDF, PPTX, or XLSX.
Fine-Tuning
The adaptation of a pre-trained model for a specific task through additional training on focused data.
G
GDPR (General Data Protection Regulation)
EU legislation regulating the processing and protection of personal data.
Gender Bias in AI
Differences in how AI treats or represents genders due to training-data patterns.
Generative AI
AI systems that generate new text, images, audio, code, or other content from learned patterns.
GPU (Graphics Processing Unit)
A specialised processor that accelerates parallel computations for training and running AI models.
Gradient Descent
An optimisation algorithm that iteratively updates model parameters to reduce prediction error.
Glossary-driven Translation
Translation guided by predefined terminology lists.
H
Hallucinations
Confident but incorrect outputs generated by AI models.
Human Evaluation
Assessment of translation quality performed manually by linguists.
Human-in-the-Loop
Workflows in which human specialists supervise or correct AI outputs.
Hybrid Translation
A translation approach combining machine output with human expertise or rule-based controls for higher quality.
I
Inference
The process by which a trained model generates translations or other outputs.
Information Retrieval
The process of searching for and identifying relevant documents, data, or information in response to a user query.
Intellectual Property Rights in MT
Rules defining ownership of source texts, translations, and AI-generated outputs.
Interoperability
The ability of different software systems, tools, or platforms to exchange and use information effectively.
J
Jaccard Similarity
A statistical measure used to evaluate the similarity between two sets, often applied in text analysis and information retrieval.
Joint Training
A machine learning training approach where a model is trained simultaneously on multiple tasks or datasets.
JSON (JavaScript Object Notation)
A lightweight data-interchange format widely used for transmitting structured information between systems and APIs.
Just-in-Time Compilation (JIT)
A runtime optimisation technique in which code is compiled during execution in order to improve performance.
K
Keyphrase Extraction
The automated identification of important phrases within a document to represent its main topics.
Knowledge Graph
A structured representation of entities and their relationships used by AI systems to organise and connect information.
L
Large Language Model (LLM)
A neural model trained on vast text corpora, capable of understanding and generating natural language.
Latency
The delay between sending a translation request and receiving a response.
Logging
Automatic recording of system events or requests used for monitoring and troubleshooting.
LSP (Language Service Provider)
A company offering translation, localisation, and linguistic workflow services.
M
Machine Learning
A field of AI where systems learn patterns from data to make predictions or generate content.
Machine Translation (MT)
Automatic translation of text from one language to another.
Machine Translation Post-Editing (MTPE)
Human editing of machine-generated translations to ensure accuracy and style compliance.
Model Training
Adjusting model parameters through exposure to data so it can learn linguistic patterns.
Multimodal AI Models
AI systems that combine text, image, audio, and video inputs to improve understanding and generation across tasks.
Multilingual Workflows
Processes for coordinating translation and localisation across multiple languages.
N
Named Entity Recognition (NER)
A natural language processing method used to identify and classify entities such as names, organisations, locations, and dates in text.
Natural Language Processing (NLP)
A field of artificial intelligence focused on enabling computers to understand, analyse, and generate human language.
Neural Machine Translation (NMT)
A machine-translation approach based on neural networks, enabling more fluent and context-aware results.
Neural Network
A computational model made up of interconnected layers that learns patterns from data and powers many modern AI systems.
O
OCR (Optical Character Recognition)
Technology that extracts text from scanned documents or images.
OOV (Out-of-Vocabulary)
Words or tokens that do not appear in a model’s training vocabulary and therefore cannot be directly recognised or translated by the system.
Open-Source Model
An AI model whose architecture, code, or weights are publicly available, allowing researchers and developers to inspect, modify, and deploy it.
Overfitting
A machine learning problem where a model learns the training data too closely and performs poorly on new or unseen inputs.
P
Parallel Corpus
A collection of texts and their translations in two or more languages used to train machine translation systems.
Post-Editing (MTPE)
The process of reviewing and correcting machine translation output to achieve publishable quality.
Pretraining
The initial phase of training a machine learning model on large datasets before adapting it to specific tasks.
Project Management in Translation
Organising translation tasks, resources, deadlines, and quality processes.
Prompt
A structured instruction or input guiding an AI model’s behaviour.
Prompt Engineering
The practice of designing and structuring prompts to obtain more accurate and useful outputs from AI models.
Q
Quantisation
A model optimisation technique that reduces the numerical precision of neural network parameters to decrease memory usage and improve inference speed.
Quality Assurance (QA)
Systematic checks ensuring accuracy, consistency, and compliance with project requirements.
R
Rate Limits
Restrictions on how many API requests can be processed within a time window.
Reinforcement Learning
A machine learning method in which an agent learns decision-making through rewards and penalties.
Retrieval-Augmented Generation (RAG)
An AI architecture combining language generation with external knowledge retrieval for more accurate outputs.
Responsible AI
An umbrella concept combining ethics, transparency, fairness, and safety in AI operations.
Revision vs Review
Quality control stages: revision checks against the source; review checks monolingually.
Rule-Based Translation
A translation approach based on linguistic rules and dictionaries instead of statistical or neural methods.
S
Semantic Similarity
A measure used in natural language processing to determine how similar two texts are in meaning.
Segment-based Translation
Translation performed segment by segment rather than on the full document.
Speech Recognition
Technology that converts spoken language into written text using machine learning models.
Style Guide
A document defining linguistic, stylistic, and formatting rules for translation.
Supervised Learning
A machine learning method in which models are trained using labelled data.
Sustainability of AI Systems
Practices aimed at reducing the environmental and computational impact of AI.
T
Terminology Extraction
Automated or manual identification of key terms for project use.
Terminology Management
Organising, updating, and maintaining termbases for consistency.
TER (Translation Edit Rate)
A metric showing how many edits a human would need to correct a translation.
Token
A unit of text processed by a model, such as a word, subword, or punctuation mark.
Tokenisation in Natural Language Processing
How text is segmented into machine-readable units for NLP pipelines and large language models.
Training Large Language Models
How LLMs are pretrained, optimised, and adapted using large datasets and distributed compute.
Transformer Architecture
The neural architecture used in modern LLMs and NMT systems.
Transformer Architecture Explained
A practical explanation of attention, encoder-decoder design, and why transformers power modern AI language systems.
Translation Memory (TM)
A database storing previously translated segments for reuse.
TMX (Translation Memory eXchange)
A standard XML format for sharing translation memory data.
U
Universal Language Models
Multilingual language models that use shared representations to support many languages and NLP tasks.
Unsupervised Machine Translation
A machine translation approach that learns to translate between languages without using parallel bilingual corpora.
Unsupervised Learning
A type of machine learning in which models identify patterns in data without labelled training examples.
V
Validation Dataset
A dataset used during model training to evaluate performance and detect problems such as overfitting before final testing.
Vector Database
A specialised database designed to store and retrieve vector embeddings efficiently for similarity search and semantic retrieval.
Vendor Neutrality
A commitment to flexible, non-proprietary technologies that avoid vendor lock-in.
Vocabulary
The set of tokens or words that a language model can recognise and process when analysing or generating text.
W
Weight Parameters (Model Weights)
The numerical parameters inside a neural network that determine how input data is processed and how predictions or generated text are produced.
Wide-context Translation
Translation using extended context windows or whole-document information to improve coherence and accuracy.
X
XLIFF (XML Localisation Interchange File Format)
An XML-based standard designed to exchange localisation data between translation tools and software systems.
XML (Extensible Markup Language)
A structured markup language used to encode documents and data in a format that can be processed by both humans and machines.
Z
Zero Data Retention
A mode in which user data is deleted almost immediately after processing.
Zero-Shot Learning
A machine learning capability that allows models to perform tasks they were not explicitly trained on by leveraging generalised knowledge.
Zero-Shot Translation
A multilingual translation capability where models translate between language pairs not directly present in training data.
Z-Score
A statistical measure indicating how many standard deviations a data point is from a dataset mean.
Where to go next
Take these concepts into practice with product tours, pricing comparisons, and current updates from Trad AI.