Xapian 1.0 Term Indexing/Querying Scheme

Table of contents

Introduction

In Xapian 1.0, the default indexing scheme has been changed significantly, to address lessons learned from observing the old scheme in real world use. This document describes the new scheme, with references to differences from the old.

Stemming

The most obvious difference is the handling of stemmed forms.

Previously all words were indexed stemmed without a prefix, and capitalised words were indexed unstemmed (but lower cased) with an 'R' prefix. The rationale for doing this was that people want to be able to search for exact proper nouns (e.g. the English stemmer conflates Tony and Toni). But of course this also indexes words at the start of sentences, words in titles, and in German all nouns are capitalised so will be indexed. Both the normal and R-prefixed terms were indexed with positional information.

Now we index all words lowercased with positional information, and also stemmed with a 'Z' prefix (unless they start with a digit), but without positional information. By default a Xapian::Stopper is used to avoid indexed stemmed forms of stopwords (tests show this shaves around 1% off the database size).

The new scheme allows exact phrase searching (which the old scheme didn't). NEAR now has to operate on unstemmed forms, but that's reasonable enough. We can also disable stemming of words which are capitalised in the query, to achieve good results for proper nouns. And Omega's $topterms will now always suggest unstemmed forms!

The main rationale for prefixing the stemmed forms is that there are simply fewer of them! As a side benefit, it opens the way for storing stemmed forms for multiple languages (e.g. Z:en:, Z:fr: or something like that).

The special handling of a trailing . in the QueryParser (which would often mistakenly trigger for pasted text) has been removed. This feature was there to support Omega's topterms adding stemmed forms, but Omega no longer needs to do this as it can suggest unstemmed forms instead.

Word Characters

By default, Unicode characters of category CONNECTOR_PUNCTUATION (_ and a handful of others) are now word characters, which provides better indexing of identifiers, without much degradation of other cases. Previously cases like time_t required a phrase search.

Trailing + and # are still included on terms (up to 3 characters at most), but - no longer is by default. The examples it benefits aren't compelling (nethack--, Cl-) and it tends to glue hyphens on to terms.

A single embedded ' (apostrophe) is now included in a term. Previously this caused a slow phrase search, and added junk terms to the index (didn't -> didn and t, etc). Various Unicode characters used for apostrophes are all mapped to the ASCII representation.

A few other characters (taken from the Unicode definition of a word) are included in terms if they occur between two word characters, and ., , and a few others are included in terms if they occur between two decimal digit characters.