Meaning Representations in Statistical Word Alignment

As a testbed of statistical word aligner, we implemented the prototype of statistical word aligner by graphical models [2, 10]. The advantage of using graphica l method resides in its extensibility compared to the traditional approach for statistical word a lignment [3, 22, 14]. Although there are semi-supervised word aligner [6], we only talk about uns upervised word aligner [3, 22, 14]. The capabilities of this word aligner include that 1) it supp orts IBM / HMM models as well as tree-based Models [13], 2) it can extend easily to support MA P assignment-based decoder (Viterbi decoding [21] as well as posterior decoding [10]) in these mo dels [16, 15], 3) it can be used for the input in the lattice-based decoding [1, 4] which are reinter pr ted as the partial model selection, 4) it supports flexible on / off capability of random variables whi ch has advantageous in the lemma-based alignment [5] and the morpheme avoided alignment, and 4) it c an be used for the forced alignment [18]. This comes from the fact that the inference algorithms , such as sum-product and max-product algorithms, are not affected by the form of network structur es. Note that the traditional statistical word alignment is built purely counting the frequency of wor ds where syntax / semantics are not considered [11].