Co-occurrence measures which linguistic units appear together or within a shared sequence. Instead of counting words only across whole documents, it uses nearby context to preserve relationships that help represent meaning. section 6.2.4 section 2.3
The useful shift is to make documents very short, so nearby-word counts become informative. Tokenize and segment sentences, then let large collections of unlabeled text provide the co-occurrence patterns used to learn word embeddings. section 6.2.4
Co-occurrence supports several language-analysis workflows:
- Compare terms within a dataset to derive meanings and transform bag-of-words or TF–IDF vectors into topic vectors with SVD. section 4.11
- Learn word embeddings from a large body of text without dictionary definitions, thesauri, or supervised labels. section 6.2.4
- Avoid treating every token independently, since ignoring co-occurring token sequences can weaken classification, sentiment analysis, and related vector representations. section 2.3
