... algorithms), but is trained entirely from parallel sentences without gold-standard word alignments Thus, it addresses the two limitations of current word alignment approaches This paper is structured as ... trained to maximize likelihood: infrequent source words act as “garbage collectors”, with many target words aligned to them (the word dislike in the Model alignment in Figure is an example) Thus, we ... for word alignment models since, unlike English, Czech words have a complex inflectional morphology, and the syntax permits relatively free word order For this language pair, we evaluate alignment...