A program for aligning sentences in bilingual corpora is described. The method is based on a simple statistical model of character lengths. It was tested on a small trilingual sample of Swiss economic reports and later applied to a large sample of 90 million words of Canadian Hansards, which was donated to the ACL/DCI. The program aligns sentences between two languages by using a dynamic programming framework to find the maximum likelihood alignment. The alignment is based on the ratio of sentence lengths and the variance of this ratio. The model assumes that longer sentences in one language tend to be translated into longer sentences in the other language, and shorter sentences into shorter ones. The parameters of the model are determined empirically from the UBS data. The program was evaluated against human judgments and showed a 4.2% error rate on 1316 alignments, which was reduced to 0.7% by selecting the best scoring 80% of the alignments. The method is fairly language-independent and performs well with both English-French and English-German data. The program is efficient, taking 20 hours of real time on a Sun 4 to align 367 days of Hansards. The results suggest that character-based alignment is more accurate than word-based alignment. The method is simple yet effective, and could serve as a useful first step in building probabilistic dictionaries or bilingual concordances for machine translation and lexicography. The program's success in aligning sentences suggests that length could be a strong clue for sentence alignment.A program for aligning sentences in bilingual corpora is described. The method is based on a simple statistical model of character lengths. It was tested on a small trilingual sample of Swiss economic reports and later applied to a large sample of 90 million words of Canadian Hansards, which was donated to the ACL/DCI. The program aligns sentences between two languages by using a dynamic programming framework to find the maximum likelihood alignment. The alignment is based on the ratio of sentence lengths and the variance of this ratio. The model assumes that longer sentences in one language tend to be translated into longer sentences in the other language, and shorter sentences into shorter ones. The parameters of the model are determined empirically from the UBS data. The program was evaluated against human judgments and showed a 4.2% error rate on 1316 alignments, which was reduced to 0.7% by selecting the best scoring 80% of the alignments. The method is fairly language-independent and performs well with both English-French and English-German data. The program is efficient, taking 20 hours of real time on a Sun 4 to align 367 days of Hansards. The results suggest that character-based alignment is more accurate than word-based alignment. The method is simple yet effective, and could serve as a useful first step in building probabilistic dictionaries or bilingual concordances for machine translation and lexicography. The program's success in aligning sentences suggests that length could be a strong clue for sentence alignment.