SMAI@IIIT-H (Monsoon 2017)
- Kritika Prakash (20161039)
- Aditya Srivastava (201425112)
- Karthik Chintapalli (201501207)
- Vighnesh Chenthil Kumar (201501201)
Course Instructor
Project Mentor
Classify emails from the Enron email dataset based on their predicted authorship, and used the trained classifier to identify authors of test samples.
Available here, the dataset contains 0.5 million emails from about 150 users, who were employees of Enron.
The classifers use the authors as classess and the emails as samples to be assigned to those classes by authorship.
The number of author classes were fixed while maximising the number of emails per author, and while keeping the emails-per-author ratio similar for every author class.
This number was found to be 10 authors with 800-1000 emails each.
The Enron corpus contains all emails in raw form, including not only the message but also all the email metadata.
The data is cleaned to keep only the subject and body of the mails. All attached forward chains are removed, including forwarded threads, and salutations.
The data is also tokenised by word, sentence and paragraph, and is case normalised.
The following different models have been implemented and tested:
The CNN can identify commonly used groups of words and phrases by an author. Also, the CNN captures localized chunks of information which is useful for finding phrasal units within long texts. There are three layers to the CNN
- First, the embedding layer generates a sequence of word-embeddings from a sequence of words
- Second, the conv layer performs the convolution operation using 128 5x5 filter
- Third, the dense layer is used for classification
The Bi-LSTM is a commonly used technique for text classification.
LSTMs are a special kind of RNN which are more capable of remembering long term dependencies in a sequence. This gives more context to the classifier which helps in author identification while processing a sequence of text.
There are three layers to the model
- First, the embedding layer generates a sequence of word-embeddings from a sequence of words
- Second, the bidirectional LSTM generates email embeddings from the sequence of word embeddings
- Third, the dense layer is performs the classification
LSTMs are known to work best for a sequence of length of 10-15 elements. However, in this implementation the model can take the entire document, increasing the length and hence the overall context for classification.
There are four layers to this model
- First, the embedding layer generates a sequence of word-embeddings from a sequence of words
- Second, the first bidirectional LSTM generates sentence embeddings from the sequence of word embeddings
- Third second bidirectional LSTM generates email embeddings from sentence embeddings
- Fourth, the dense layer is performs the classification
This model appends stylometric features to the final document embedding in the hierarchical Bi-LSTM, right before it is passed on to the dense layer. The classification is now performed these augmented documenting-embeddings.
The stylometric features extracted from the data and experimented with are
- Lexical
- Average sentence length
- Average word length
- total number of words
- Ratio of unique words to total number of words
- Total number of characters
- Syntactic
- Total number of function words
- Total number of personal pronouns
- Total number of adjectives
root/
| data_preprocessing_scripts/
- dataProcessing.py
| extracted_features/
- adjperemail.txt
- avgsentlenperemail.txt
- avgwordlenperemail.txt
- charsperemail.txt
- funcwordsperemail.txt
- perpronperemail.txt
- stylometricVector.txt
- uniqbytotperemail.txt
- wordsperemail.txt
| feature_extraction_scripts/
- adjperemail.py
- avgsentlenperemail.py
- avgwordlenperemail.py
- charsperemail.py
- funcwordsperemail.py
- perpronperemail.py
- stylometricVector.py
- uniqbytotperemail.py
- wordsperemail.py
| models/
- CNN.py
- HierLSTM_withStylometry.py
- HierLSTM.py
- LSTM_final.py
- README.md