Naslov
Improved sentence retrieval using local context and sentence length

Autori
Doko, Alen ; Štula, Maja ; Šerić, Ljiljana

Izvornik
Information processing & management (0306-4573) 49 (2013), 6; 1301-1312

Vrsta, podvrsta i kategorija rada
Radovi u časopisima, članak, znanstveni

Ključne riječi
sentence retrieval; TF-ISF; context; sentence length

Sažetak
In this paper we propose improved variants of the sentence retrieval method TF-ISF (a TF-IDF or Term Frequency – Inverse Document Frequency variant for sentence retrieval). The improvement is achieved by using context consisting of neighboring sentences and at the same time promoting the retrieval of longer sentences. We thoroughly compare new modified TF-ISF methods to the TF-ISF baseline, to an earlier attempt to include context into TF-ISF named tfmix and to a language modeling based method that uses context and promoting retrieval of long sentences named 3MMPDS. Experimental results show that the TF-ISF method can be improved using local context. Results also show that the TF-ISF method can be improved by promoting the retrieval of longer sentences. Finally we show that the best results are achieved when combining both modifications. All new methods (TFISF variants) also show statistically significant better results than the other tested methods.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo

POVEZANOST RADA