This is an old revision of the document!
Jacana word aligner
Slides accompanying the article are located at http://cs.jhu.edu/~xuchen/paper/yao-jacana-wordalign-acl2013.ppt
1/ Describe parameters of feature function in section 3.1. Is there something unclear?
2/ Section 3.1 describes a problem with unknown number of states.
For example, we have the following source sentence: “Linda is married to Charles.”
and target sentence: “Charles likes his new role as the husband.”
and tags: NOUN, VERB, ADJ, ADV, PRON, DET, ADP, NUM, CONJ, PRT
How many states will be used?
3/ Search WordNet (http://wordnetweb.princeton.edu/perl/webwn) for following words and write answers:
a / Imagine, that source sentence contains word “well” and target “part”.
Find differences with incorporating POS tagger and without it. How big will be gain in number of comparisons (just roughly)?
b/ Is there a synset containing both “well” and “water” words?
4/ Have a look at Table 3. How big was the gain from incorporating Wordnet with POS tagger? Would you use two external resources for this gain (refer to the discussion in section 4.6)?
5/ Why do you think, the authors add Hamming distance into the first formula in section 3.1. Can normalization function work without it?