Relevance is a key concept in searching large volumes of unstructured text – and a key tool in maximising the recall of content related to a search.
Within Solr (the underlying index technology), relevance is determined (against a search) by a mathematical calculation known as ‘BM25’, which essentially determines within a large body of material which of the (typically hundreds or thousands) results are most relevant to the search input. There are several factors which feed into this, but by and large the calculation looks for words which are ‘rare’ in the whole archive but ‘common’ in the individual result – and scores them more highly. So words with high ‘discriminatory value’ (such as proper noun place names) will drive a higher relevance score – especially if mentioned multiple times in an article.
As we have substantially expanded the recall effectiveness of our search, the relevance score has become increasingly important to help deal with large volumes of results. In addition, there are two other factors which are very important in scoring relevance for results.
‘Prioritise within’ (#)
'Prioritise within’ is explicitly designed to bring specific results within a wider results set to the fore without the loss of recall that would result from making them a ‘required’ word.
So we have separated the concept of ‘must include’ anchor words and phrases (which must occur) and ‘prioritise within’ words and phrases – which may occur, and if so, will be promoted to the top of search results.
Note: ‘prioritise within’ will only affect results if they are ordered by relevance.
Title weighting
Each article has both body text (all of the text) and title text (typically the first few hundred characters). Typically, the title text will capture the headline/title and the opening sentence or so.
Searches interrogate both the body and title text for an article – and a match on defined keywords and phrases within the title text is weighted to increase relevance. This means that article content which mentions the specific target terms or phrases in the title appears high in the results (by relevance) – which is appropriate as the article is likely to be focussed on the specific content if it occurs in the title.
As you can see in the top results for this example search:
