* complain about state of the art

SMT is lexical, where each surface form in the source and target is an
independent entity.
bad because of sparsity. 

data sparsity is bad enough for popular languages like english, but
it's a nightmare for any language with a rich morphology (and ~anything
is richer than english).

goals:
generalize over morphology to reduce data sparsity
model syntactic coherence (morph agreement) in  the target language.

this has been done before, but mostly by using a target LM
(including related work here: Mei+Katrin, Kevin+Katrin, Amittai)

there's also been a lot of work done on modifying the translation
input to make it better approximate the structure of the desired
output, with the hope that this will improve alignments and thus
translation. (see section 3 in the paper).

* this paper

trained a model that predicts inflected forms, given the source sentence
and a sequence of stems in the target sentence.

used:

word and word alignments, plus "lexical resources that provide morph
info about the words"-- so whatever you have (probably a parser and a
dictionary. they also had projected dependency structures, presumably
created by the MSR treelet translation project).

worked on english--russian and english--arabic.
[discussion of russian + arabic morphology omitted]

a characteristic of morph complex languages is a richer system of
agreements-- so some non-lexical notions are needed if we're going to
have agreement spread out over a larger chunk of a sentence than just
adjacent words.

* inflection prediction

  - stemming (set of possible lemmas for a word w)
  - inflection (set of surface forms with the same stem as w)
  - morph analysis (set of analysis vectors for w)

* the learning model

  - MaxEnt Markov model, which allows you to decompose the prob of a
    predicted inflection sequence into a product of local probs for
  individual predictions, conditioned on some small number of previous
  predictions. here they used 2nd order (trigram) prediction.

  - the features they use pair items from the context (what you have
    so far on the target side, plus the source element for the current
    slot) and the target label for the current element.

  - structured prediction: the new target word has to match the
    predicted morphological analysis vector, not just be the highest
    prob word given the previous two elements.

* features and their categories 

  two axes for kinds of features:

  - monolingual (target language only) 
    vs bilingual (obtained from following the word alignments back
                into to the source sentence)

  - lexical (surface forms + stems: this looks kinda like a trigram LM,
             except it can also use information from the right context)
    vs morphological (used to describe the target label and its
                      context, so a sequence of POS tags would count),
    vs syntactic (derived from a parser: stem of the parent of a word, eg)


monolingual lexical:
stem of word and its adjacent words, plus LM features
monolingual morph:
attributes of current prediction plus two previous ones
monolingual syntactic:
stem of parent node

bilingual lexical:
words aligned to current one, plus all words aligned to immediate neighbors

bilingual morph and syntactic:
it's complicated.

* experiments 

  to cut down on noise, they used aligned sentencs pairs from
  reference translations as the input to their postprocessing, rather
  than SMT output.

 -  data: technical (limited) domain, 1million aligned en-ru, 500k
    en-ar.

 -  lexicon: russian general-domain available, buckwalter for arabic.

 - baseline: compared to picking a morphological inflection at random
   from the possible ones for the current word. that is, "how does our
   system fare compared to chance?"
   also, a trigram LM.