Outline

• 1: introduction, overview
  • background for the problem: WSD in machine translation, why CL-WSD matters and some good lexical choice examples
  • thesis statement: CL-WSD is a feasible and practical means for lexical selection in a hybrid MT system for a language pair with relatively modest resources.
  • Intuition: when translating into a lower-resourced language, our target-language LM cannot do as much work in lexical selection for us, so we need to do more work understanding the source-language text. Hence. WSD.
  • We'll demonstrate this with a number of CL-WSD experiments (showing we can do it), and integrate CL-WSD into a number of MT systems (to show that it helps).
  • research questions
    • Can we use monolingual evidence from the source language to help with CL-WSD?
    • Can we use multilingual evidence from the source language with other languages?
    • Can we use sequence labeling techniques?
    • Which kinds of MT systems can benefit from CL-WSD?
  • structure of the rest of the dissertation
    • here's what happens in the subsequent chapters
• 2: background
  • history of WSD and CL-WSD
  • hybrid MT
  • Guarani language
  • Quechua language
• 3: related work
  • CL-WSD approaches in vitro
  • WSD with sequence models
  • lexical selection in RBMT
  • CL-WSD for SMT
  • WSD for lower-resourced languages
  • SemEval and Senseval
  • Translation into MRLs
• 4: overview of tasks, evaluation, baseline system
  • measuring CL-WSD classification accuracy
  • measuring MT improvements (extrinsic evaluation)
  • using the Bible: how why
    • background on using Bible as corpus ...
    • preprocessing steps
  • exploring our bitext
  • description of the baseline system
  • source-side annotations
  • classification results: the baseline system
    • something Zipfian about most common words being most polysemous.
    • Test set on 100 most common words.
    • test set on Bible sentences: sample a bunch of sentences that use those 100 most common words
• 5: learning from monolingual data
  • unsupervised learning from big corpora: word representations
  • word2vec: embeddings
  • brown clusters
  • extracting features with existing NLP tools: analysis with FreeLing
  • CL-WSD experiments
• 6: learning from multilingual data
  • europarl
  • prepackaged multilingual data: the PPDB
  • other bitext for Spanish
  • joint classification with MRFs
  • classifier stacking
  • CL-WSD experiments
• sequence models
  • MEMMs
  • linear chain CRFs (maybe)
  • CL-WSD experiments
• all together now: combining different signals
  • CL-WSD experiments
• integration into MT systems
  • SQUOIA (skip?)
  • Terere: hybrid MT with cdec for es-gn
    • language models for Guarani
    • CL-WSD at runtime with cdec in Python
    • getting morphology right
    • important: don't spend too much time on this, like comparing different SMT styles
    • training by symmetrization to extract phrases: that's our training set for CL-PSD
      • maybe this is really important for es-gn -- Spanish phrases often translate to Guarani words.
  • MT experiments
• conclusions
  • we can do it
  • it works well
  • nobody had investigated these particular strategies before: now we can make better word choices when translating from a resource-rich language to a resource-poor one