spreads an output signal [2].
The query evaluation is based on spreading activation. Each node computes an input and the number of entries in each dictionary. tionaries were used all of which were actually simply a list of terms in language l1 that Translation : is based on \dictionaries". For the CLEF1 experiments, three bilingual dicwere paired with some equivalent terms in language l2. Table 1, shows the source and an input value from this initial query: = and then an activation value : In(ti) qui s 1. The query is the input of the network. Each node from the term layer computes Qu = where g is the identity function. Out(ti) g(In(ti)) top (X=12) retrieved by the query source. The terms are sorted according the following A context of the target query is built using an aligned corpus. It consists of selecting the formula : best terms appearing in the top (X=12) documents in target language aligned to the retained we do not select only the best translation as it is done in some other works [1]. appear in the list of terms of the target context. However, if a specic term has an target query. Note that in this process all the terms appearing in the target context are The desambiguisation of the translated query consists of retaining only terms that unique substitution this term is retained even though it not exists in the context of the four languages (German, French, Italian and English), were submitted. The queries were Two runs using English topics and retrieving documents from the pool of documents in all Documents
Indexing
Index L1 Index L2 Index Ln Source query
Dico.
Aligned corpus Translation sustitustion list
Target query Matching merging List of documents
Desambiguisation dk2Dx dik Two runs were submitted : irit1men2a based on normalised merging and irit2men2a based on naive merging. 16 (best 0) 24 (worst 1) irit2men2a Table 2 compares our runs against the published median runs. We notice that for both runs the number of topics better and less than median are slightly the same. documents are then sorted according to their RSV. The top 1000 were submitted. naive strategy : all the documents resulting from the pair searches join a nal list. These irit1men2a Run-Id irit2men2a ference at average precision. Nothing was gained from the normalised strategy. better than the normalised strategy in the top document, and at Exact precision but no difTable 3 compares the merging strategies. It can be seen that the naive strategy is slightly E2G (37 queries) E2E (33 queries) E2I (34 queries) E2F (34 queries) Pair language better than median Avg. Prec. : 15 (best 0) irit1men2a worse than median at Avg. Prec. : 25 (worst 2) merged and sorted according to their normalised RSV. The nal list corresponds to the normalised strategy : each list of retrieved documents resulting from the pair search was top 1000 documents. normalised. The normalisation consists simply of dividing the RSV of each document by the maximum of RSVs in that list. The documents of the dieren t lists are then translated using the downloaded dictionaries. No desambiguisation, all the translated words were tested : were retained in the target queries. The runs were performed by doing individual runs for pair languages and merging the results to form the nal rank ed list. Two merging strategies better than median Avg. Prec. : 22 (best 3) irit1bfr2en worse than median at Avg. Prec. : 11 (worst 2) Dico+Des.
Run-id (33 queries) Impr (%) Dico 152 296 228 467 E2I E2G 76 171 Table 7 compares the results between the runs Dico+desambiguisation and Dico only.
The desambiguisation is eectiv e the average precision improves of 6%. 0.2841 5.8 0.2685 Exact gual (E2E) search performs much more better than all the pair (E2F, E2G, E2I) searches. best multilingual search. The merging strategy caused the loss of relevant documents, Table Moreover, all the pair searches (except E2G) have their average precision better than the Table 4 shows the results of pair language (example, E2F means English queries translated to French and compared to French documents, etc.). We can easily notice that the monolin5 shows the total number of relevant in the pair list and the numuber of document which was kept in the nal list lost when merging. Relev ant documents were lost from all the pair lists.
Table 6 compares our run against the published median runs. Most queries give results better than the median and 3 were the best. built by RALI Lab (http://www-rali.iro.umontreal.ca/wac/). desambiguisation was performed using WAC (Word-wide-web Aligned Corpus) parallel corpus The bilingual experiment was carried on using F2E free dictionary + desambiguisation. The Three runs were submitted in monolingual tasks : iritmonofr, iritmonoit, iritmonoge multilingual and the bilingual searches. Secondly, French monolingual results seem to be First of all, we notice clearly that the monolingual search is much better than both the better than both Italian and the German. Italian results are better than German. These Rel. kept in the nal list Rel. Ret. by pair list Rel. lost. E. M. Voorhees and Harman D.K. (Ed.), NIST SP 500-236, Nov. 1997. ceedings of the 7th International Conference on Text REtrieval TREC7, [3] M. Boughanem, T. Dkaki, J. Mothe & C. Soule-Dupuy, Mercure at trec7, Proevance backpropagation in Adhoc environment, Information Processing and Man[2] M. Boughanem, C. Chrisment & C. Soule-Dupuy, Query modic ation based on relagment. April 1999. runs were done using exactly the same procedures the only dierence concerns the stemming which was used only for French.