This paper describes in detail the combination of NLP methods applied to the treatment of logic forms in the topic processing and statistical methods applied to the search engine in the frame of the CL-SR performance. The method that infers the logic form of a topic is based on dependency analysis between the words of the topic. These dependencies between the words of the topic are calculated using the MINIPAR parser. Di®erent combinations of the topic, description and narrative ¯elds are used in the runs to perform the retrieval process. The based on logic forms method processes the description and narrative ¯elds of the topics. This processing task consists on the removal of several terms according to the logic structure of the processed ¯eld in the logic form. On the other hand, the statistical processing applied to the search engine consists on using IR-n system. IR-n system is a passage retrieval system that manages overlapping of variable passages that are composed by a number of sentences. Di®erent statistical similarity measures are managed by IR-n system to acquire the topic terms weight. The removal of several topic terms according to the logic structure of the topic origines that the rest of the topic terms acquire a better relevance.
Di®erent combinations of the topic, description and narrative ¯elds can be applied to perform the information retrieval process. This fact implies that all the terms of these ¯elds are used by the search engine to accomplish its goal. The search engine usually removes many terms that can be considered as stop-words (prepositions, articles and so on). If we have a look to the structure of the description and narrative ¯elds of a topic (see table 1), we can deduce that there exists many terms that would not be as relevant as other terms in the information retrieval process. Our system processes the topics according to an NLP based approach. The topic processing basically consists on removing several terms of the description and narrative ¯elds of the topic. Obviously, these removed terms are consider as not relevant and then they will not be processed by the information retrieval engine.
In this new Cross-Language Speech Retrieval (CL-SR) Track, our research e®ort has been focused on combining the use of NLP and statistical methods in the CL-SR performance. Concretely,
of children born in 1930-1933 who spend the war in concentration camps or in ...
describe the circumstances and inner resources of the
surviving children
our main research goal has been centered on demonstrating that the use of NLP methods by way
of processing the topics according to the logic structure of their associated logic form increases
the results obtained by the statistical search engine. We applied the new version of IR-n system
[
The following section shows the topic processing by way of applying NLP rules based on the logic structure of their associated logic forms. Finally, we describe the submitted runs, the obtained results in these submitted runs, and discuss the application of NLP methods to the statistical IR-n system. 2
This section presents the topic processing applying NLP rules based on logic forms [
The topic processing by way of logic forms consists on removing many terms of the topic according to the logic structure of its logic form. A combination of the text, description and narrative ¯elds of the topic has been employed to perform the information retrieval process according to the submitted runs. The rules are only applied to the description and narrative ¯elds of the topic because these ¯elds contains a lot of information (see table 1) that would be previously ¯ltered before to be processed by the information retrieval process. These rules are independently applied to the description and narrative ¯elds of the topic when the number of words of these ¯elds are upper to 10 words. In other case, there is not necessary to remove any word (term).
These rules consist on the removal of the ¯rsts words until a preposition (predicate type IN), or a main verb (predicate type VB or VBE), or a compositional structure (predicate type CC), both included. If a preposition or a verb are the ¯rsts words in the sentence, we removed them and then the processing continues until ¯nding another preposition, main verb or compositional structure. If in this search process the system detects a noun (predicate type NN) coinciding with the nouns of the topic ¯eld then the search process is aborted until this noun. Table 2 shows how an example of the application of these rules.
Once these rules are applied, the next process consists on performing the search in the document
collection according to combination of updated ¯elds by the application of these rules. The
statistical IR-n system [
This section describes the submitted runs in which our system has participated in. The di®erences between these ¯ve submitted runs are basically based on the combination of the topic ¯elds and
Field Describe survival mechanisms of children born in 1930-1933
who spend the war in concentration camps or in ...
The relevant material should describe the circumstances and inner resources of the surviving children ...
describe:VB(e2, x11, e1) survival:NN(x1)
NNC(x8, x1, x9) mechanism:NN(x9)
of:IN(x8, x2) child:NN(x2) bear:VB(e1, x8, x10) in:IN(e1, x4) ...
relevant:JJ(x1) material:NN(x1) describe:VB(e1, x1, x5) circumstance:NN(x6)
and:CC(x5, x6, x3) inner:NN(x2)
NNC(x3, x2, x4) resource:NN(x4) ... on the indexation of a combination of di®erent segment ¯elds from 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: ² UA TDN FL ASR06BA1A2 Run. In this run IR-n system indexes the combination of the ASRTEXT2006B, AUTOKEYWORD2004A1 and AUTOKEYWORD2004A2 segment ¯elds of the document collection. The English title, description and narrative topic ¯elds are used in the construction of the queries. This was the unique submitted run in which we apply the rules based on the topic processing by way of logic forms described in previous section. ² UA TDN ASR06BA1A2 Run. In this run, as previous submitted run, IR-n system indexes the combination of the ASRTEXT2006B, AUTOKEYWORD2004A1 and AUTOKEYWORD2004A2 segment ¯elds of the document collection. The English title, description and narrative topic ¯elds are used in the construction of the queries. ² UA TD ASR06BA2 Run. In this run IR-n system indexes the combination of the ASRTEXT2006B and AUTOKEYWORD2004A2 segment ¯elds of the document collection.
Only the title and description topic ¯elds are used in the construction of the queries. ² UA TDN ASR06BA2 Run. In this run, as in previous run, IR-n system indexes the combination of the ASRTEXT2006B and AUTOKEYWORD2004A2 segment ¯elds of the document collection. The English title, description and narrative topic ¯elds are used in the construction of the queries. ² UA TD ASR06B Run. In this required run, IR-n system only indexes the ASRTEXT2006B segment ¯eld of the document collection. Only the title and description topic ¯elds are used in the construction of the queries. 4
In this research we have demonstrated that the previous preprocessing of the topics according to NLP methods produces an improvement in the statistical retrieval process. The NLP methods are based on the logic structure of the narrative and description topic ¯elds.