homework solutions for:

Homework #8:  Language Modeling






submitted to:

Dr. Joseph Picone
ECE_8993 Fundamentals of Speech Recognition

September 24, 1998


submitted by:

Janna Shaffer








Department of Electrical and Computer Engineering
Mississippi State University
Box 9571, 216 Simrall, Hardy Rd.
Mississippi State, Mississippi 39762
Email: janna@cs.msstate.edu
Results
The following model was built with a vocabulary of the 1000 most frequent words and using the entire data set.
evallm : perplexity -text ../hamaker/data/all_data.text
Computing perplexity of the language model with respect
   to the text ../hamaker/data/all_data.text
Perplexity = 69.12, Entropy = 6.11 bits
Computation based on 17455054 words.
Number of 3-grams hit = 17455053  (100.00%)
Number of 2-grams hit = 1  (0.00%)
Number of 1-grams hit = 0  (0.00%)
8256284 OOVs (32.11%) and 0 context cues were removed from the calculation.

These models were built using different kept portions (80% of all the data) and then evaluated on the held-out data corresponding to each unique kept portion.
evallm : perplexity -text data/20_percent_00.text 
Computing perplexity of the language model with respect
   to the text data/20_percent_00.text
Perplexity = 94.05, Entropy = 6.56 bits
Computation based on 3480540 words.
Number of 3-grams hit = 3116271  (89.53%)
Number of 2-grams hit = 333716  (9.59%)
Number of 1-grams hit = 30553  (0.88%)
1657462 OOVs (32.26%) and 0 context cues were removed from the calculation.
evallm : perplexity -text data/xaa
Computing perplexity of the language model with respect
   to the text data/xaa
Perplexity = 94.87, Entropy = 6.57 bits
Computation based on 4363160 words.
Number of 3-grams hit = 3875987  (88.83%)
Number of 2-grams hit = 444714  (10.19%)
Number of 1-grams hit = 42459  (0.97%)
2060282 OOVs (32.07%) and 0 context cues were removed from the calculation.

evallm : perplexity -text data/xac
Computing perplexity of the language model with respect
   to the text data/xac
Perplexity = 67.56, Entropy = 6.08 bits
Computation based on 4366169 words.
Number of 3-grams hit = 4366168  (100.00%)
Number of 2-grams hit = 0  (0.00%)
Number of 1-grams hit = 1  (0.00%)
2066872 OOVs (32.13%) and 0 context cues were removed from the calculation.

evallm : perplexity -text data/xab
Computing perplexity of the language model with respect
   to the text data/xab
Perplexity = 68.06, Entropy = 6.09 bits
Computation based on 4374071 words.
Number of 3-grams hit = 4374070  (100.00%)
Number of 2-grams hit = 1  (0.00%)
Number of 1-grams hit = 0  (0.00%)
2060120 OOVs (32.02%) and 0 context cues were removed from the calculation.



















APPENDIX
Histograms for unigrams, bigrams, and trigrams for the entire data set.