Translation Memory (TM)

The foundational process of dividing source text into logical units called segments—typically sentences or paragraphs—and pairing them with their corresponding target translations. This is governed by segmentation rules (SRX) that handle abbreviations, punctuation, and formatting.

• Rule-Based Splitting: Uses regular expressions to identify sentence boundaries, avoiding splits on abbreviations like "Dr." or "U.S."
• Structural Alignment: Maps source segments to target segments, even when the order changes due to linguistic differences
• Embedded Content Handling: Preserves inline tags for formatting, variables, and placeholders within segments to protect code integrity

The retrieval engine compares new source segments against the database to find previously translated content. The match type determines the level of human intervention required.

• Exact Match (100%): The source segment and its surrounding context are identical to a stored entry, requiring zero editing
• Context Match (101%): An exact match where the preceding and following segments also match, guaranteeing identical document context
• Fuzzy Match (50-99%): A partial similarity calculated via Levenshtein distance or similar algorithms; the translator edits the suggested target text rather than translating from scratch
• No Match: The segment is entirely new and must be translated, then stored for future reuse

An advanced retrieval technique that breaks segments into smaller chunks—phrases or terminology units—and matches them independently. This is critical for leveraging translations of recurring phrases even when the full sentence is new.

• Statistical Phrase Extraction: Identifies high-frequency n-grams within the TM and indexes them separately
• Concordance Search: Allows translators to query the TM for how a specific word or phrase was translated in all previous contexts
• Machine Learning Integration: Modern TMs use neural models to identify semantically similar sub-segments, not just string-identical ones, improving match rates for paraphrased content

A mechanism to protect specific text strings from translation, ensuring brand names, product codes, and regulatory statements remain identical across all locales.

• Inline Protection: Uses XML or HTML-like tags to mark untranslatable spans within a segment
• Locked Segments: Entire segments can be locked at the database level, preventing any modification even by administrators
• Regular Expression Filtering: Automatically identifies and protects patterns like email addresses, URLs, and alphanumeric product codes before segmentation occurs
• Compliance Assurance: Critical for regulated industries where specific legal disclaimers must appear verbatim in every language

Each translation unit stores rich metadata that enables precise filtering and retrieval based on domain, client, project, or quality status.

• Domain Classification: Tags segments by subject matter (legal, medical, technical) to prevent inappropriate reuse
• Quality Status Flags: Marks segments as "approved," "draft," or "rejected" to control which entries are leveraged
• Date and Author Stamps: Tracks when and by whom a translation was created, enabling time-based decay or preference for recent translations
• Custom Attributes: Supports user-defined fields like product line, document type, or regulatory region for granular workflow routing

The ability of a TM to operate in both language directions, automatically populating the reverse language pair when a translation is stored. This eliminates the need for separate databases for each direction.

• Automatic Inversion: When English→French is stored, the system immediately creates a French→English entry
• Consistency Enforcement: Ensures that the same term is translated identically in both directions, critical for technical documentation
• Multi-Language Projects: A single TM can store dozens of target languages for one source, enabling simultaneous leverage across all language pairs in a localization program

A retrieval technique that identifies previously translated segments which are similar but not identical to the current source text. The system calculates a percentage match score (e.g., 75%, 89%) based on edit distance. A human translator then modifies the suggested target text rather than translating from scratch. This is distinct from a 100% match (also called an exact match), which requires no editing. Fuzzy matching thresholds are configurable; setting them too low retrieves irrelevant suggestions, while setting them too high misses useful partial matches.

A centralized, structured repository of approved terms and their translations, often including usage rules, part-of-speech tags, and contextual notes. Unlike a Translation Memory, which stores full sentences, a termbase operates at the lexical level to enforce consistent terminology. When integrated with a TM, the system can automatically highlight or enforce a termbase entry during translation, overriding the TM suggestion if a conflict exists. This is critical for brand names, legal definitions, and technical jargon.

An algorithm for evaluating translation quality by measuring the precision of n-gram matches between a candidate translation and one or more human reference translations. It applies a brevity penalty to prevent artificially short translations from scoring highly. While widely used for benchmarking Machine Translation engines, BLEU has known limitations: it operates on surface-level string matching, penalizes legitimate synonym use, and correlates imperfectly with human judgment. It is a corpus-level metric, not a sentence-level quality indicator.

A machine learning task that predicts the quality of a translation without access to a human reference. QE models output confidence scores at the word, phrase, or sentence level, enabling workflows where low-confidence segments are automatically routed to human post-editors. Modern QE systems use cross-lingual pre-trained language models to assess fluency and adequacy. This is a critical component for reducing human review costs in high-volume, automated localization pipelines.

A software platform that centralizes the entire localization workflow. A TMS integrates Translation Memories, termbases, and Machine Translation engines into a single interface. It manages job assignment, version control, and review cycles. Key features include:

• Connector frameworks to pull content from CMS or code repositories
• Workflow automation for routing content based on QE scores
• Analytics dashboards tracking cost savings from TM reuse

A secondary machine learning task focused on automatically correcting errors in raw Machine Translation output. An APE model is trained on pairs of raw MT output and human post-edited versions. It learns to fix systematic errors like incorrect word order, missing negation, or wrong gender agreement. APE can be applied as a final processing step before delivery, reducing the cognitive load on human post-editors and further closing the quality gap between raw MT and publishable text.