Methods & evaluation
How NLP systems turn text into something a model can use, and how to tell whether they actually work: tokenization, word senses, benchmarks and machine translation metrics.
- Benchmark Data Contamination: When Test Sets Leak into Training Data
What benchmark data contamination is, how test examples end up in the training data of language models, the methods used to detect it (n-gram overlap, guided completion, Min-K% Prob, ordering tests) and the practical defences: canary strings, encrypted test sets and fresh benchmarks.
- How NLP Models Are Evaluated: Benchmarks, Metrics and Their Blind Spots
A practical guide to evaluating natural language processing systems: test sets and splits, accuracy, precision, recall and F1, benchmark suites such as GLUE and HELM, significance testing, annotation artefacts and data contamination.
- How Tokenization Works: Words, Subwords and Bytes in NLP
A clear guide to tokenization in natural language processing: why models split text into subword units, how BPE, WordPiece, unigram and SentencePiece differ, why some languages cost more tokens than others, and what to check in practice.
- Machine Translation Evaluation: BLEU, chrF, COMET and Human Judgement
How machine translation quality is measured: BLEU and its pitfalls, chrF, learned metrics such as COMET, direct assessment and MQM human evaluation, quality estimation without references, and how to compare two systems fairly.
- Named Entity Recognition: How NLP Systems Find People, Places and Organisations
What named entity recognition (NER) does, how entity types and BIO tags are defined, how methods moved from rules and CRFs to BiLSTMs and fine-tuned transformers, why exact-match F1 can mislead, and where NER still struggles: nested entities, new names and languages with little annotated data.
- Retrieval-Augmented Generation (RAG): How It Works and How to Evaluate It
What retrieval-augmented generation is, from the 2020 RAG paper by Lewis and colleagues to today's pipelines: sparse, dense and late-interaction retrieval such as BM25, DPR and ColBERT, chunking and reranking, why models lose information in long contexts, and how to evaluate retrieval and grounded answers separately.
- Word Sense Disambiguation: How Machines Decide What a Word Means
What word sense disambiguation is, from Lesk's dictionary overlaps to contextual embeddings, how it is evaluated with WordNet, SemCor and the Senseval/SemEval benchmarks, and why ambiguity still trips up translation and search.






