This paper describes the participation of the University of Alicante in the new CL-SR Track at CLEF conference. In this track we introduce a set of features in the topics processing applied by our IR-n system. This set of features are based in the application of logic forms to topics and in the increment of the terms weight of the topics according to a set of syntactic rules.
In the same line of active participation of University of Alicante in previous CLEF conferences,
IR-n system takes part in this new Cross-Language Speech Retrieval (CL-SR) Track at the present
CLEF 2005 conference. As novelty, IR-n system includes a new module that increments the terms
weights of the topics applying a set of rules based on the representation of the topics in the way
of logic forms [
Following section shows the main features of this new release of IR-n system. The logic form derivation module and the rules applied to these logic forms are also presented in following sections. Finally, we describe each one of the submitted runs, the scores obtained by the IR-n system in these submitted runs, the conclusions and the future works in our research activity.
These passages are usually composed for a ¯xed number of sentences, but the but the format of the document collection of this CL-SR track does not allow this feature. These documents are composed by a contiguous set of words without punctuation marks. Moreover, we can't know the limit between each sentence. As a result, we have chosen a ¯xed number of words to compose the passages. Furthermore, IR-n system uses overlapping passages in order to avoid that some documents could be considered not relevant if it appears words of the question in adjacent passages.
IR-n system allows the use of distinct similarity measures (Ex. Okapi [
Once the weights of the words have been calculated and with the aim of increment the weights of several words, IR-n system incorporates a new module that apply a set of heuristics to the representation of the topics in the way of logic forms.
According to others IR systems, IR-n system uses di®erent techniques of the query expansion.
Previous researches [
On the other hand, in present conference and for the ad-hoc track, a new technique called
variable passages [
Following section shows in detail the main features of the treatment of topics in the way of logic forms performed by IR-n system. The process that automatically derives the logic form applying a set of inference rules to the analysis of dependencies between the words of the topic is also described. 3
To enhance the performance of our IR-n system we use the logic form of the topics. Each one of
the terms of the topic in the logic form can modify its weight term according to the type of assert
of the term in the logic form and the relationships between these asserts of the topic in the logic
form. The logic form of a topic (or sentence) is calculated through the analysis of dependency
relationships between the words of the sentence. MINIPAR [
This task obtains the di®erent relationships between the words of the sentence. These dependency
relationships between words are calculated by MINIPAR [
The logic form of the sentence is calculated via this analysis of dependency relationships between
the words of the sentence. Our approach employs a set of rules that infer several aspects such
as the assert, the type of assert, the identi¯er of the assert and the relationships between the
di®erent asserts in the logic form. This technique improves the Moldovan technique [
Committee conj e lex-mod Emergency Rescue () fin rel t whn i p who saved subj obj
mod p () who thousands in pcomp-n p Marseille each assert has at least one argument. The ¯rst argument is usually instantiated with the identi¯er of the assert and the rest of the arguments are corresponded to the identi¯ers of other asserts that are related with this assert. As instance, in the assert \story:NN(x14)", its type is corresponded to noun (NN) and its identi¯er is instantiated to x14 ; in the assert \NNC(x11, x10, x12)", its type is corresponded to composed entity (NNC), its identi¯er is instantiated to x11, and the other two arguments indicate the relationships with other asserts: x10 and x12 ; and so on. 4
When the type of assert is a preposition (IN) which second argument instantiates an assert which type is noun (NN) or derives in a assert which type is noun, then the weight term associated to this last assert is modi¯ed. This rule generally describes those grammatically utterances that have a circumstantial behaviour in the sentence (ej. in Marseille, in concentration camps, in Sweden, of Holocaust experience and so on) and then we consider their words which POS are nouns (type of predicate NN) as very relevant words in the topic. This reason produces that we increment the weight terms of these words (terms) in 15%. Table 1 shows the terms weights that IR-n system associates to the topic the topic \The story of Variant Fly and the Emergency Rescue Committee who saved thousands in Marseille". These terms are expressed by their stem.
According to the rule described in this section, the logic form inferred for this topic (\story:NN(x14) of:IN(x14, x13) varian:NN(x10) NNC(x11, x10, x12) fry:NN(x12) and:CC(x13, x11, x6) emergency:NN(x5) NNC(x6, x5, x7) rescue:NN(x8) NNC(x7, x8, x9) committee:NN(x9) who:NN(x1) save:VB(e1, x1, x2) thousand:NN(x2) in:IN(e1, x3) marseille:NN(x3)") have two asserts which types are IN. The second argument of these asserts is instantiated to x13 and x3 respectively. x13 derives in the asserts x10, x12, x5, x8 and x9 which types are NN, while the type of x3 is directly NN. According to this rule, these fact produces that the terms weight associated to all these asserts increment their value in 15%. Table 2 shows the weight terms once this rule has been applied. This section describes the di®erent submitted runs of our IR-n system. The di®erences between these ¯ve submitted runs are basically based in the treatment of the topics and in the indexation of a combination of di®erent ¯eld of the segments in the document collection. In all submitted runs we use the indexing and searching processes developed by our IR-n system using the English as query language. There is not used any kind of thesaurus terms as keywords in the indexing and in the searching processes. Following subsections show the features of these ¯ve submitted runs according to the judgment pool priority order. 5.1
In this run IR-n system indexes the automatically created transcript using the best presently available ASR system (ASRTEXT2004A ¯eld of the segments in the document collection). The English title and description ¯elds of the topics are used in the construction of the queries. This is the unique submitted run in which we apply the rules based on the processing of queries in the way of logic forms described in previous section. 5.2
In this run, as previous submitted run, our IR-n system indexes the ASRTEXT2004A ¯eld of the segments in the document collection. The English description ¯eld of the topics is used in the construction of the queries. 5.3
In this run we index the ASRTEXT2004A ¯eld and a set of thesaurus keywords that were assigned automatically using a k-Nearest Neighbor (kNN) classi¯er using only words from the ASRTEXT2004A ¯eld of the segment (AUTOKEYWORD2004A1 ¯eld of the segments in the document collection). The English description ¯eld of the topic is used in the construction of the queries. 5.4
In this run IR-n system indexes the ASRTEXT2004A ¯eld and a a set of thesaurus keywords that were assigned using a di®erent kNN classi¯er that was trained (fairly) on di®erent data (AUTOKEYWORD2004A2 ¯eld of the segments in the document collection). The English description ¯eld of the topic is used in the construction of the queries. 5.5
In this run our IR-n system indexes the ASRTEXT2004A, AUTOKEYWORD2004A1 and AUTOKEYWORD2004A2 ¯elds of the segments in the document collection. The English description ¯eld of the topics is used in the construction of the queries. 6
In this new release of the CL-SR track at the CLEF 2005 conference we have participated applying our IR-n system to the English language. Our main aim is to evaluate the goodness of the new Logic Form Module of IR-n system. According to our foresight, the obtained scores applying this module (UATDASR04LF) are upper than the obtained scores without the use of this new module (UATDASR04).
According to the format of the document collection, each document is considered as a sentence (continuous set of words). Then, this fact produces that IR-n system runs as a document retrieval system and not as a passage retrieval system. This feature resumes that the powerful of the use of the new logic form module must be combined with the passage overlapping technique in document collections where documents are composed by many passages (see our paper in the bilingual IR track at present conference). The combination of these two techniques would obtain better scores.
Acknowledgment This research work has been partially funded by the Spanish Government under project CICyT number TIC2000-0664-C02-02 and PROFIT number FIT-340100- 2004-14 and by the Valencia Government under project numbers GV04B-276 and GV04B-268.