=Paper=
{{Paper
|id=Vol-1170/CLEF2004wn-DomainSpecific-Kojima2004
|storemode=property
|title=Ricoh at CLEF 2004
|pdfUrl=https://ceur-ws.org/Vol-1170/CLEF2004wn-DomainSpecific-Kojima2004.pdf
|volume=Vol-1170
|dblpUrl=https://dblp.org/rec/conf/clef/Kojima04
}}
==Ricoh at CLEF 2004==
https://ceur-ws.org/Vol-1170/CLEF2004wn-DomainSpecific-Kojima2004.pdf
Ricoh at CLEF 2004
Yuichi Kojima
Software R&D Group, RICOH CO., Ltd.
1-1-17 Koishikawa, Bunkyo-ku, Tokyo 112-0002, JAPAN
ykoji@src.ricoh.co.jp
Abstract. 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.
1 Introduction
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.
2 Description of the System
The basic system is same as last year. Before describing our new modifications to European languages, we give
an outline of the system as background information. It uses a document ranking method based on the probabilistic
model [3] with query expansion using pseudo-relevance feedback [4] and was shown to be effective in TREC and
NTCIR experiments.
In the following sections, we explain the processing flow of the system [5, 6].
2.1 Query term extraction
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.
2.2 Initial retrieval
Each query term is assigned a weight wt, and documents are ranked according to the score sq,d as follows:
N
wt = log k 4' • + 1 (1)
nt
f t ,d
∑
wt
s q,d = t∈q • (2)
K + f t ,d k 4' • N + 1
l
K = k1 (1 − b ) + b d (3)
l ave
where N is the number of documents in the collection, nt is the document frequency of the term t, ft,d is the
in-document frequency of the term, ld is the document length, lave is the average document length, and k'4, k1 and b
are parameters.
Weights for phrasal terms are set lower than those for single terms.
2.3 Query expansion
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 = w2 t • t − t
r n
(4)
R N
rt + 0.5 (5)
R − rt + 0.5
w2 t = α • wt + (1 − α ) • log
n t − rt + 0.5
N − n t − R + rt + 0.5
where R is the number of relevant documents, rt is the number of relevant documents containing the term t and α
is a parameter.
The weight of the initial query term is re-calculated using the same formula as above, but with a different α
value and an additional adjustment to make the weight higher than the expansion terms.
2.4 Final retrieval
Using the initial query and expansion terms, the ranking module performs a second retrieval to produce the final
result.
2.5 Cross-lingual retrieval
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.
3 Experiments
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:
3.1 Stemming and tokenizing
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.
Table 1. Summary of the experiments
Word decompounding stemming Average precision *
No German Snowball stemmer 0.3149
Yes Stemmer A** 0.2944
Yes Stemmer A + German Snowball stemmer 0.3470
* Average precision using GIRT German monolingual task of CLEF 2003 after training
** German Stemmer in the analyzer
3.2 Stopword dictionary
This year, we used stopword dictionaries at Snowball site.
3.3 Parallel corpora
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.
3.4 Training
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.
Table 2. Average precision values after training
Language Average Prec.
DE -> DE 0.3470
EN -> DE 0.1370
4 Results
Table 3 shows the summary of our official results for CLEF 2004.
The result of monolingual task was worse than our expectation. According to our estimate using our
experiments, the average precision value can be 0.1 point higher than the current value if attributes of queries are
similar to last year. As for the result of cross-lingual task, although the value was bad, it was what we were
expected.
Table 3. Official runs for CLEF 2004
Language Run Relevant Rel. Ret. Average Prec. R-Precision
DE -> DE rdedetde04 1663 922 0.2381 0.2759
EN -> DE rendetde04 1663 684 0.1261 0.1678
5 Conclusions
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.
References
[1] Y. Kojima, H. Itoh, H. Mano, and Y. Ogawa. Ricoh at CLEF 2003. At
http://clef.iei.pi.cnr.it:2002/2003/WN_web/26.pdf.
[2] Y. Kojima and H. Itoh. Ricoh in the NTCIR-4 CLIR Tasks. At
http://research.nii.ac.jp/ntcir-ws4/NTCIR4-WN/CLIR/NTCIR4WN-CLIR-KojimaY.pdf
[3] S. E. Robertson and S. Walker. On relevance weights with little relevance information. In Proceedings of the
20th Annual International ACM SIGIR Conference (SIGIR ’97), pages 16-24, 1997.
[4] Y. Ogawa and H. Mano. RICOH at NTCIR-2. In Proceedings of the Second NTCIR Workshop Meeting, pages
121-123, 2001.
[5] H. Itoh, H. Mano and Y. Ogawa. RICOH at TREC-10. In the Tenth Text Retrieval Conference (TREC-2001),
page 457-464, 2001.
[6] M. Toyoda, M. Kitsuregawa, H. Mano, H. Itoh and Y. Ogawa. University of Tokyo/RICOH at NTCIR-3 Web
Retrieval Task. At http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings3/NTCIR3-WEB-ToyodaM.pdf
[7] S. E. Robertson. On term selection for query expansion. Journal of Documentation, 46(4):359-364, 1990.
[8] S. E. Robertson and K. Spark-Jones. Relevance weighting of search terms. Journal of ASIS, 27:129-146, 1976
[9] H. Itoh. NTCIR-4 Patent Retrieval Experiments at RICOH. At
http://research.nii.ac.jp/ntcir-ws4/NTCIR4-WN/PATENT/NTCIR4WN-PATENT-ItohH.pdf
[10] Snowball web site. At http://snowball.tartarus.org/ visited 7th November 2002.