Latent semantic analysis
In September 1990 Scott Deerwester, Susan Dumais, George Furnas, Thomas Landauer and Richard Harshman described in JASIS document retrieval through the singular value decomposition of a term-by-document matrix: terms and documents get vectors of about a hundred factors, and a query is compared with them by cosine. On the MED medical collection precision rose 13% over plain word matching.
Why it matters
Words and documents became points in a space derived from which words occur in the same texts, so a document could be found even when it contained none of the query's words. Bengio's neural language model of 2003 and word2vec of 2013 both name the method as a forerunner of their word vectors.
The authors were at Bell Communications Research (Dumais, Furnas, Landauer), the University of Chicago (Deerwester) and the University of Western Ontario (Harshman). MED has 1,033 abstracts of medical papers and 30 queries, 5,823 terms; mean precision was .51 against .45 for plain term matching. Precision rose from 10 to 100 factors and was best at 100. CISI has 1,460 information science abstracts and 35 queries: .11 for both the method and term matching, while the SMART system with stemming got .14. The authors themselves warn that MED was assembled from keyword search results and may give unrealistically good results. What the record does not claim. The figures come from the authors' typescript of the paper, not from the journal pages, which are paywalled; the day the issue appeared is unknown. The text of the paper calls the method latent semantic indexing (LSI); "analysis" appears only in the title.