Minimalistic test runs of the Eidetica indexer
                                    Teresita Frizzarin (frizzarin@eidetica.com)
                                    Annius Groenink (groenink@eidetica.com)



                                                      Abstract

Participating in a text retrieval conference for the first time, Eidetica has run six minimalistic tests with its
t·repository indexer, doing as little tuning as possible, in order to evaluate its “performance baseline”. Since no
tuning was done, we will only discuss the general properties of our indexing software and how it was run on the
CLEF topic sets for the monolingual German and Dutch tasks.



1. Background
Eidetica is a service provider of search and text mining technology on the basis of an application hosting model.
We took part in the monolingual German and Dutch tracks of CLEF 2001, with the t·repository software – the
core database and indexing software that drives Eidetica’s hosting applications. These applications include web
site search, newspaper archives, subject-based personal alerting services, automatic enrichment of XML data
streams with subject keywords, and internet filtering.

The t·repository software subjected to the test, has the following primary characteristics:

•   Built for speed, reliability and stability
•   Native data input format is flat-record XML
•   The source XML is compiled to a generalized index: a mathematically motivated set of string lexicons and
    matrices that represent relations between these lexicons. Among these matrices are both forward and
    backward term indexes.
•   Record identifiers, data elements (authors, subjects, dates), terms, words and trigrams are all living in the
    same, unified, space. This allows for virtually umlimited text mining applications.
•   On-the-fly context-free tagging: using simple UNIX shell pattern matching (for German/Dutch participle
    forms) and suffix co-occurrence rules (for singular/plural etc.), unknown words are automatically tagged
    and stemmed. Thus, the Eidetica software can operate with very minimalistic dictionaries. Only 5 different
    tags are used (noun, adword, verb, det, coor), and the tags are used purely to aid term extraction.
•   Use of dictionaries is very limited. Support for force/kill lists for search terms is used very sparingly.
•   Compounding: words consisting of two parts that also exist as words, are split in their “internal
    representation”. A compound form with and without a space or dash is considered to be identical (air craft =
    air-craft = aircraft). This is especially useful for German and Dutch.
•   The indexer performs various types of term extraction (tagging-based, proper name recognition, and
    extraction from so-called full term only - fields in the source data: controlled keyword entries, authors,
    etcetera). Extracted terms have a length between 1 and 4 words (or parts-of-word in the case of
    compounds).
•   Because indexes take terms as entities, rather than single words, stop words and other irrelevant parts of text
    are automatically skipped.


2. Technique description
The topics and text were first converted (with minimal changes) to fit the profile of a single database, where
records happen to contain either the fields TI, LE, TE and CAPTION (article type), or the fields TITLE, DESC
and NARR (topic type). Small CLEF-specific term kill lists were made with terms such as “relevant documents”
and “information” that hadn’t previously been any threat to Eidetica mining applications. Then, the standard
Eidetica t·repository indexer was run on the full document set without modifications to software or dictionaries.
In other words, we have implemented automatic runs.
The indexer produces individual forward and backward indexes in the form of matrices between
Document/Topic IDs and the terms appearing in the fields TI, LE, TE, CAPTION, TITLE, DESC and NARR.
The vectors in these indexes are normalized, so that there sums are 1.

We then compute a linear combination of these indexes, where the DESC and TE fields are given a very high
weight, and the remaining fields are blended in as “back-ups”. The result is a matrix from document/topic IDs to
terms. This matrix is transposed to a matrix from terms to document/topic IDs. Finally, we infer a Topic ID –
Document ID matrix by composing the elementwise 4th power root of both matrices, producing a Topic-
Document rank that is proportional to the sum of the square roots of individual term co-occurrences.

This ID – ID rank matrix is the basis for the Topic-ID – Article-ID lists that we submitted in various forms.

3. Submitted runs
We have submitted the same three runs for both the Dutch monolingual task, and the German monolingual task.
These runs are:

•   EidNL2001A: use all topic fields (TITLE, DESC and NARR) and keep only results above a threshold score
    (0.80)
•   EidNL2001B: use all topic fields, but produce the best 1000 results.
•   EidNL2001C: use only the TITLE and DESC fields of the topics, and produce the best 1000 results.

•   EidDE2001A-C: as their Dutch counterparts, but for German.