Word alignment try 2

[johnml1135]

Add word alignment engine to IInteractiveTranslationEngine.

---

This change is 

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 0% with 20 lines in your changes missing coverage. Please review.

Project coverage is 69.92%. Comparing base (7f2af4e) to head (08a2719).

Files with missing lines	Patch %	Lines
...SIL.Machine/Translation/HybridTranslationEngine.cs	0.00%	20 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #267      +/-   ##
==========================================
- Coverage   69.96%   69.92%   -0.05%     
==========================================
  Files         379      379              
  Lines       31778    31798      +20     
  Branches     4456     4456              
==========================================
  Hits        22235    22235              
- Misses       8509     8529      +20     
  Partials     1034     1034              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ddaspit]

What is this change for?

Reviewable status: 0 of 3 files reviewed, all discussions resolved

[johnml1135]

This is needed for adding the word alignment engine to Serval. Just exposing the alignment endpoints to the interactive engine.

[johnml1135]

This needs to be merged and released before the Serval changes will be able to compile.

[ddaspit]

I'm still not sure I understand what this is for. There are already interfaces for word alignment models. Also, phrase alignment isn't word alignment. That is specific to the Thot SMT engine.

Reviewable status: 0 of 3 files reviewed, all discussions resolved

[johnml1135]

The ThotSmtModel appears to be the best place to add the alignment routines onto - as the "phrase alignment" just means that the tokenizer can be configured. If I don't use ThotSmtModel, what specific things would I use? IWordAligner assumes that the source and target are already tokenized. Also, how would it interact with loading models built by machine.py?

[ddaspit]

For word alignment, you should use one of the classes that inherits from ThotWordAlignmentModel. For SMT and word alignment models, you will need to tokenize the text. We should just use the LatinWordTokenizer like we do for the SMT engine.

Reviewable status: 0 of 3 files reviewed, all discussions resolved

[johnml1135]

Hmm. It wold be quite a bit of reworking. I would have to use a different wording than ThotWordAlignmentModel because that is just referring to the asymmetrical alignment, not the symmetrical alignment with tokenizer. In python, the word aligner has the tokenizer connected to it. I could rework the Machine word aligner to have the tokenizer in it, but that would be a fair amount of work. The solution I have appears to be a good minimal solution - treat the ThotSmtModel as a SymmetrizedWordAlignmentModel with tokenizers - it already has the capability of having the truecaser as null.

Otherwise, I think I would have to create base class of ThotSmtModel called ThotSymmetrizedWordAlignmentModelWithTokenizer? in which 1/2 of the functionality of ThotSmtModel is implemented. And even then, all the configurations and trainers and everything else would need to be torn apart and rewritten.

I think this minimal change is the best solution - it looks like a word aligner on Serval but is just an SMT model underneath.

[ddaspit]

The ThotSmtModel is a full phrased-based SMT system and takes a lot more computation and time to train. The phrase alignment from the SMT model uses a different algorithm than the word alignment models and is much more expensive. Unfortunately, it is not a replacement for the word alignment models. We should meet to discuss how best to proceed. I'm sure if I had a better understanding of what you are trying to achieve, we can come up with a good solution.

Reviewable status: 0 of 3 files reviewed, all discussions resolved