Machinevertaling

Neural systems often treat this as a sequence-to-sequence task: read one sequence and produce another. Building and evaluating such a system requires attention to data alignment and language-specific errors.

Key takeaways

• Check aligned data and document-level splits.
• Record evaluation configuration.
• Inspect meaning-changing errors by language pair.

Parallel training data pairs source passages with corresponding translations. Incorrect alignment, duplicated material, or mismatched language labels can teach the wrong relationship. Clean the pairs and keep documents together when splitting evaluation data to avoid near-duplicate leakage. Tokenization determines how text becomes model inputs. A tokenizer that handles one writing system efficiently may split another into many more units. Check the length limits in both languages and whether truncation removes the end of either the source or target passage. Automatic metrics make repeated experiments practical. BLEU compares patterns of word sequences with reference translations, but a metric is not a complete judgment of meaning, readability, or suitability for a domain. Evaluation settings and reference choices matter, so record them with the score. Use an error taxonomy alongside metrics: additions, omissions, changed numbers, inconsistent terminology, incorrect negation, and awkward phrasing. Assess each language pair and domain separately. An average across several well-resourced languages can conceal failures in a less-represented language or specialized document type.

1. Imagine a reference “The package did not arrive.” Candidate A says “The parcel never arrived.” Candidate B says “The package did arrive.”
2. Candidate A uses different words but preserves the main meaning. Candidate B resembles the reference while reversing the outcome.
3. Record the negation error explicitly instead of choosing a translation by appearance or overlap alone.

The constructed example demonstrates why metric-based comparisons need semantic review.