This paper describes the participation of RICOH in the monolingual and cross-lingual information retrieval tasks on German Indexing and Retrieval Testdatabase (GIRT) in the Cross-Language Evaluation Forum (CLEF) 2004. We used a morphological analyzer for word decompounding and parallel corpora for cross-lingual information retrieval. The performance of cross-lingual information retrieval was poor and that of monolingual information retrieval was not good. We need to check our modules and procedures. We are enhancing our information retrieval system for some languages [1, 2]. Our approach to the enhancement is to use same basic system and modify language depend modules. Our system showed reasonable performance for some European languages and the importance of word decompounding for the compound rich languages such as German in the participation in CLEF 2003 tasks [2]. This is our second participation in CLEF tasks. We used a commercial morphological analyzer for word decompounding and participated in GIRT tasks. Our focuses for this year were: 1. to confirm the performance of word decompounding 2. to find the problems in applying our approach to cross-lingual information retrieval Section 2 of this paper outlines our system, section 3 describes the modifications made for the experiments, section 4 gives the results, and section 5 contains some conclusions.
We used “title” and “description” fields of each topic. An input topic string is transformed into a sequence of stemmed tokens using a tokenizer and stemmer. Stop words are eliminated using a stopword dictionary. Two kinds of terms are extracted from stemmed tokens for initial retrieval: a “single term” is each stemmed token and a “phrasal term” consists of two adjacent tokens in a stemmed query string.
(
Weights for phrasal terms are set lower than those for single terms.
As a result of the initial retrieval, the top 10 documents are assumed to be relevant (pseudo-relevance) to the query and selected as a “seed” for query expansion. Candidates for expansion terms are extracted from the seed documents in the same way as for the query term extraction mentioned above. Phrasal terms are not used for query expansion. The candidates are ranked on the Robertson's Selection Value [7], or RSVt and the top ranked terms are selected as expansion terms. The weight is re-calculated as w2t using the Robertson/Sparck-Jones formula [8].
RSVt = w2t • rt R − nt N (
Using the initial query and expansion terms, the ranking module performs a second retrieval to produce the final result.
We performed English-to-German retrieval using well-known strategy and parallel corpora [9]. In the cross-lingual retrieval process, the English query is submitted against the English database and top-n documents are obtained. Their counterparts in the German database are exploited as seed documents to extract German query terms. The extraction can be performed using completely same mechanism for query expansion in pseudo-relevance feedback.
There are five items in the system which needs adjustments depending on the language, 1) the tokenizer, 2) the stemmer, 3) the stopword dictionary, 4) the training data and 5) the parallel corpora.
We used mostly the same modules as last year and a commercial morphological analyzer which can tokenize a sentence, decompose a compound word, and stem a word.
Details of the items in the system are as follows:
We had a selection of possible combinations of the stemmers and the tokenizers. The system can use Snowball stemmer [10] and simple tokenizer which were used for the last year’s CLEF experiments, and the morphological analyzer which is imported into the system this year.
The possible combinations are limited by the behavior of the analyzer. It decomposes a compound word into its single words and stems each single word in the same procedure. So there is no selection of word decompounding without the stemming in the analyzer.
After some experiments, we selected the combination of 1) word decompounding and 2) a two step stemming which consists of the first stemming step of the decompounding and the second stemming step using snowball stemmer.
Table 1 shows the summary of our experiments.
This year, we used stopword dictionaries at Snowball site.
We prepared additional two document databases using English and German GIRT corpus. First database was made from the English corpus by extracting each tagged entity (TITLE, AUTHOR and ABSTRACT) as a document and used for making lists of seed documents. Second database was made from the German corpus by the same method and used for making German queries from the lists of seed documents.
Each document was tokenized and stemmed depend on its language using the method mentioned above.
Although we cannot expect all of corpus to be parallel in the real situation, we used all parallel corpora because we could not get average score even using them.
We searched the parameters of the system by the hill-climbing method, using average precision values of search results with query expansion for the monolingual and cross-lingual retrieval task.
Table 2 shows the average precision values after training.
We tested our new module for word decompounding and checked problems in applying our approach to cross-lingual retrieval. According to our experiments, word decompounding is effective. But the result of the official experiment was not good. We need check of our procedure to make data for submission. On the other hand, we could not achieve the average score in cross-lingual retrieval tasks (both our experiment and official experiment). We will check modules and the procedure for cross-lingual retrieval.
Further analysis with additional experiments will be shown in the CLEF workshop.