Bilgiye erisimin en yaygin ve en kolay yollarindan biri Internet, Internet’teki bilgiye erisimin en kolay yolu ise arama motorlaridir. Ancak arama motorlarindan istenilen bilgiye erismek icin once uygun anahtar kelimeler secmeli, daha sonra sonuc sayfalari yada sonuc sayfalarindaki baglantilardaki sayfalar kullanici tarafindan bizzat. incelenmelidir. Oysa kullanicilarin belirli tipteki sorulari icin tum bu islemleri otomatik hale donusturmek mumkundur ve Ingilizce icin bircok sistem gelistirilmistir. Kullanicisindan sorusunu dogal dillerde kabul edebilen ve kullanicisina cevabi bulmasi icin sayfa adresleri listesi yerine cevabin kendisini verebilen sistemlere soru cevaplama sistemleri denmektedir. Bu calismada; Turkce icin dogal dille calisan bir soru cevaplama sistemi gerceklestirilmistir. Sistem oncelikle kullanicisinin dogal dille sordugu sorusunu arama motoru sorgusuna cevirmekte ve arama motorunun sonuc sayfasindan yada baglantilarindaki sayfalardan olasi cevap cumlelerini secmektedir. Olasi cevap cumlelerini cesitli kriterlere gore puanlandirip en yuksek puani alan ilk bes cumle kullaniciya iletilmektedir. Sistem 524 adet soruyla degerlendirilmis ve arama motorunun sonuc sayfasi kullanildiginda sorularin yaklasik %43’ une, sonuc sayfasinin baglantilarindaki sayfalar kullanildiginda %60’ina ilk bes cumle icinde dogru cevap verebilmistir. Absract The easiest and common way to achieve information is Internet. The easiest way to achieve information on Internet is using a search engine. But user most do extra processes (selection of keywords, research on result pages and web sites) while using a search engine. These extra processes can be done automatically for some kind of queries. The systems, that accepts user queries in natural language and results in only answer not a document list, are named Question-Answering Systems. The instances of these systems exist for English. In this study, a Question-Answering system is built for Turkish. The system converts user question in natural language to search engine query and sends these queries to a search engine and receives result page. In result page and the web pages which web addresses are in result page, the answer candidate sentences are selected. The candidate answers are ranked according to four criteria. The five candidates which have the highest scores are the answers. The system is examined with 524 questions. 43% of questions are answers correctly in first five sentences when the result page of search engine was used. 60% of questions are answers correctly in first five sentences when the real pages (at the result page’s links) was used.
[1]
Oren Etzioni,et al.
Scaling question answering to the Web
,
2001,
WWW '01.
[2]
Susan T. Dumais,et al.
An Analysis of the AskMSR Question-Answering System
,
2002,
EMNLP.
[3]
Marc Najork,et al.
Measuring Index Quality Using Random Walks on the Web
,
1999,
Comput. Networks.
[4]
Stephen E. Robertson,et al.
Okapi/Keenbow at TREC-8
,
1999,
TREC.
[5]
Eduard H. Hovy,et al.
The Use of External Knowledge of Factoid QA
,
2001,
TREC.
[6]
Michael Collins,et al.
Answer Extraction
,
2000,
ANLP.
[7]
Leila Kosseim,et al.
QUANTUM: A Function-Based Question Answering System
,
2002,
Canadian Conference on AI.
[8]
Wei Li,et al.
QuASM: a system for question answering using semi-structured data
,
2002,
JCDL '02.
[9]
Eugene Charniak,et al.
A Maximum-Entropy-Inspired Parser
,
2000,
ANLP.