An implementation of Aggregate and mixed-order Markov models for statistical language processing as described in the Saul & Pereira 1997 paper.
Harshal Godhia
hgodhia@umass.edu
lm.py
Contains all the python code for training and testing the model.
model_params.pkl
A pre-trained model whose parameters are serialized and saved into this file
test_sentences.txt
Sample sentences used througout the questions for the homework
Q1 Which file do I run?
A1 lm.py will use a pre-trained model and sample sentences from the file test_sentences.txt to evaluate sentence probability
Q2 What packages do i need to run?
A2 This code is tested in the environment Python 3.6.0 64-bit. You would need the following packages installed nltk, nltk.data(brown corpus), numpy, matplotlib
Q3Where do I add new test sentence whose probability I want?
A3Add new sentences, one sentence per line, in the file test_sentences.txt. the code will report sentence probability for the sentence and its reverse ordering. Incase you want a custom ordering then use add the delimiter || on the same line and another sentence.
- Some details
- There are lots of sentences already prepopulated (that can serve as examples if you like to add more test cases)
- Please ensure all words are in the vocabulary, it will throw an exception otherwise (not handled by code)
Q4 What is model_params.pkl
A4 It is a pre-trained model (emission and transition matrices) which I found to give good numbers for the questions in the assignment. Since EM approaches a local optimum and depends on random initialization, the results like the ratio for the colorless sentence may not be reproduced. Hence I saved the best model (after different tries) in this file.
Q5 How do I train a new model?
A5 Uncomment line #353. It will train a new model and save it into the file model_params.pkl replacing the pre-trained model
- Some details
- Training may take upto 4 mins overall
- Please re-run incase the results don't match as reported in the writeup