#disclaimer: This is work in progress. If you encounter any problems while compiling or using it, it is likely our mistake not yours. Please contact wammar@cs.cmu.edu with questions, comments, and suggestions.
#description: This is an implementation of the CRF autoencoder framework for four tasks:
- bitext word alignment
- part-of-speech tagging
- code switching
- dependency parsing
Our NIPS 2014 paper describes the CRF autoencoder framework as well as the bitext word alignment and part-of-speech induction tasks in detail. Details on code-switching can be found in our EMNLP shared task paper.
#dependencies:
#how to build I'm assuming your default compiler is either gcc 4.6.3, clang 3.1-8 (or later "fingers crossed")
- bitext word alignment: make -f Makefile-latentCrfAligner
- part-of-speech tagging: make -f Makefile-latentCrfPosTagger
- code switching: make -f Makefile-latentCrfPosTagger (this is not a typo)
- dependency parsing: make -f Makefile-latentCrfParser (still in the works)
--output-prefix prefix # just a filename prefix for files generated during training
--train-data sent-per-line-space-delimited-tokens.txt # example file below
--feat LABEL_BIGRAM --feat PRECOMPUTED --feat EMISSION
--feat BOUNDARY_LABELS --feat PRECOMPUTED_XIM2 --feat PRECOMPUTED_XIM1
--feat PRECOMPUTED_XI --feat PRECOMPUTED_XIP1 --feat PRECOMPUTED_XIP2
--feat OTHER_ALIGNERS
--min-relative-diff 0.001
--optimizer adagrad --minibatch-size 8000
--max-iter-count 50
--cache-feats true
--wordpair-feats word-level-features```
for a list of all options: execute ``latentCrfAligner --help``
### snippet of the file ``sent-per-line-space-delimited-tokens.txt``
Ms. Haag plays Elianti . Rolls-Royce Motor Cars Inc. said it expects its U.S. sales to remain steady at about 1,200 cars in 1990 .
### snippet of the file ``word-level-features``
expects starts-with-e 1 starts-with-ex 1 ends-with-ts 1 ends-with-s 1 plays starts-with-p 1 starts-with-pl 1 ends-with-ys 1 ends-with-s 1
## using multiprocesses:
```mpirun 32 train-latentCrfAligner [options]```