TY - GEN
T1 - The feasibility analysis of re-ranking for N-best lists on English-Turkish machine translation
AU - Yildirim, Ezgi
AU - Tantug, Ahmet Cuneyd
PY - 2013
Y1 - 2013
N2 - In this paper, we present the results of re-ranking for N-best list on machine translations. The main purpose of this research is to determine the upper bound of MT success that can be gained by reordering possible candidate translations. We use Google Translate Research API1 as our Statistical Machine Translation (SMT) system to get the N-best lists consisting of possible Turkish translations for a given English sentence. We evaluate the effect of reordering using three simple methods: unigram count (UC), unigram ratio (UR), and first four characters match (FFCM). We collected 720 sentences in order to give to the SMT system, and then we used 3 different sets of Turkish translations of them to evaluate our work on the N-best lists. Success of re-ranking is determined by using BLEU metric, besides an inclusive investigation which is necessary especially for agglutinative languages (e.g. Turkish, Czech, Hungarian, and Finnish) is performed by using BLEU+ MT scoring tool. We observe an improvement in BLEU score from 31.71 for the baseline system to 35.46 which is about 11.81% relative for the re-ranked model using UR.
AB - In this paper, we present the results of re-ranking for N-best list on machine translations. The main purpose of this research is to determine the upper bound of MT success that can be gained by reordering possible candidate translations. We use Google Translate Research API1 as our Statistical Machine Translation (SMT) system to get the N-best lists consisting of possible Turkish translations for a given English sentence. We evaluate the effect of reordering using three simple methods: unigram count (UC), unigram ratio (UR), and first four characters match (FFCM). We collected 720 sentences in order to give to the SMT system, and then we used 3 different sets of Turkish translations of them to evaluate our work on the N-best lists. Success of re-ranking is determined by using BLEU metric, besides an inclusive investigation which is necessary especially for agglutinative languages (e.g. Turkish, Czech, Hungarian, and Finnish) is performed by using BLEU+ MT scoring tool. We observe an improvement in BLEU score from 31.71 for the baseline system to 35.46 which is about 11.81% relative for the re-ranked model using UR.
KW - BLEU+
KW - machine translation
KW - n-best list
KW - re-ranking
UR - https://www.scopus.com/pages/publications/84883435387
U2 - 10.1109/INISTA.2013.6577652
DO - 10.1109/INISTA.2013.6577652
M3 - Conference contribution
AN - SCOPUS:84883435387
SN - 9781479906611
T3 - 2013 IEEE International Symposium on Innovations in Intelligent Systems and Applications, IEEE INISTA 2013
BT - 2013 IEEE International Symposium on Innovations in Intelligent Systems and Applications, IEEE INISTA 2013
T2 - 2013 IEEE International Symposium on Innovations in Intelligent Systems and Applications, IEEE INISTA 2013
Y2 - 19 June 2013 through 21 June 2013
ER -