REPORT.txt

I have implemented the trigram language models using the Katz back off smoothing technique which are trained on the
BROWN corpus and REUTERS corpus.
The results are as below:

1. For the language model that is trained on the BROWN corpus:
- I use the first 490 brown corpus files to train (it takes about 7s for training)
- After that, I evaluate this language model on both BROWN corpus and REUTERS corpus:
    + Computing the perplexity for the rest 10 brown corpus files (it takes about 33s)
    ==> perplexity = 274.4732873901855

    + Computing the perplexity for the test set of the REUTERS corpus (3019 files, it takes ~800s)
    ==> perplexity = 564.916121809559

2. For the language model that is trained on the REUTERS corpus:
- I use the training set of the REUTERS corpus to train (7769 files, it takes ~8s)
- After that, I evaluate this language model on both BROWN corpus and REUTERS corpus:
    + Computing the perplexity for the test set of the REUTERS corpus (3019 files, it takes ~295s)
    ==> perplexity = 135.91220816643056

    + Computing the perplexity for the rest 10 brown corpus files (it takes ~39s)
    ==> perplexity = 657.2699009793142