Python implementation of an N-gram language model with Laplace smoothing and sentence generation.
Some NLTK functions are used (nltk.ngrams, nltk.FreqDist), but most everything is implemented by hand.
Note: the LanguageModel class expects to be given data which is already tokenized by sentences. If using the included load_data function, the train.txt and test.txt files should already be processed such that:
- punctuation is removed
- each sentence is on its own line
See the data/ directory for examples.
Example output for a trigram model trained on data/train.txt and tested against data/test.txt:
Loading 3-gram model...
Vocabulary size: 23505
Generating sentences...
...
<s> <s> the company said it has agreed to sell its shares in a statement </s> (0.03163)
<s> <s> he said the company also announced measures to boost its domestic economy and could be a long term debt </s> (0.01418)
<s> <s> this is a major trade bill that would be the first quarter of 1987 </s> (0.02182)
...
Model perplexity: 51.555
The numbers in parentheses beside the generated sentences are the cumulative probabilities of those sentences occurring.
Usage info:
usage: N-gram Language Model [-h] --data DATA --n N [--laplace LAPLACE] [--num NUM]
optional arguments:
-h, --help show this help message and exit
--data DATA Location of the data directory containing train.txt and test.txt
--n N Order of N-gram model to create (i.e. 1 for unigram, 2 for bigram, etc.)
--laplace LAPLACE Lambda parameter for Laplace smoothing (default is 0.01 -- use 1 for add-1 smoothing)
--num NUM Number of sentences to generate (default 10)
Originally authored by Josh Loehr and Robin Cosbey, with slight modifications. Last edited Feb. 8, 2018.