An efficient C++ implementation of the exchange word clustering algorithm
Optimizes bigram perplexity by swapping words between classes. The evaluations can be done in parallel in multiple threads. Word-class and Class-word statistics are used for efficiency.
One sentence per line is assumed. Sentence begin and end markers (<s> and
</s>) are added to each line if not present in the corpus. Perplexity values
include the sentence end symbol.
For more details:
Martin, Liermann, Ney: Algorithms for bigram and trigram word clustering, Speech Communication 1998
Botros, Irie, Sundermeyer, Ney: On efficient training of word classes and their application to recurrent neural network language models, Interspeech 2015
For converting the training and possible development sets to class sequences, the class_corpus.py script is provided.
To handle possible out-of-vocabulary words in these sets, the unk symbol must be specified.
If the unk symbol is written in capitals (for instance VariKN), use the --cap_unk switch and for lowercase
unk symbol (for instance SRILM) use the --lc_unk switch.
ngramppl/classppl/classintppl checks for the correct unk symbol.
Usage example:
exchange -c 1000 -a 1000 -m 10000 -t 2 corpus.txt exchange.c1000
class_corpus.py --cap_unk exchange.c1000.cmemprobs.gz <train.txt >train.classes.txt
class_corpus.py --cap_unk exchange.c1000.cmemprobs.gz <devel.txt >devel.classes.txt
varigram_kn -3 -C -Z -a -n 5 -D 0.02 -E 0.04 -o devel.classes.txt train.classes.txt exchange.vkn.5g.arpa.gz
classppl exchange.vkn.5g.arpa.gz exchange.c1000.cmemprobs.gz eval.txt