Conventions for search result scoring
- by DeaconDesperado
I assume this type of question is more on-topic here than on regular SO.
I have been working on a search feature for my team's web application and have had a lot of success building a multithreaded, "divide and conquer" processing system to work through a large amount of fulltext.
Our problem domain is pretty specific. Users of the app generate posts, and as a general rule, posts that are more recent are considered to be of greater relevance. Some of the data we are trying to extract from search is very specific (user's feelings about specific items or things) and we are using python nltk to do named-entity extraction to find interesting likely query terms. Essentially we look for descriptive adjective-noun pairs and generate a general picture of a user's expressed sentiment as a list of tokens. This search is intended as an internal tool for our team to draw out a local picture of sentiments like "soggy pizza." There's some machine learning in there too to do entity resolution on terms like "soggy" to all manner of adjectives expressing nastiness.
My problem is I am at a loss for how to go about scoring these results. The text being searched is split up into tokens in a list, so my initial approach would be to normalize a float score between 0.0-1.0 generated off of how far into the list the terms appear and how often they are repeated (a later mention of the term being worth less, earlier more, greater frequency-greater score, etc.) A certain amount of weight could be given to the timestamp as well, though I am not certain how to calculate this.
I am curious if anyone has had to solve a similar problem in a search relevance grading between appreciable metrics (frequency, term location/colocation, recency) and if there are and guidelines for how to weight each.
I should mention as well that the final fallback procedure in the search is to pipe the query to Sphinx, which has its own scoring practices. Sphinx operates as the last resort in case our application specific processing can't find any eligible candidates.
