Generic Bi-LSTM Model for Word Importance Prediction in Spoken Dialogues:

This project demonstrates the use of generic bi-directional LSTM models for predicting importance of words in a spoken dialgoue for understanding its meaning. The model operates on human-annotated corpus of word importance for its training and evaluation. The corpus can be downloaded from: http://latlab.ist.rit.edu/lrec2018

Performance Summary (will-be-updated-soon):

Model	Classes	f-score	rms
bi-LSTM (char-embedding + CRF)	3	0.73	0.598
bi-LSTM (char-embedding + CRF)	6	0.60	0.154

You can cite this work and/or the corpus using:

Sushant Kafle and Matt Huenerfauth. 2018. A Corpus for Modeling Word Importance in Spoken Dialogue Transcripts. In Proceedings of the 11th edition of the Language Resources and Evaluation Conference (LREC). ACM.

I. Data Preparation

1. Download and locate the text transcripts of the Switchboard Corpus and the corresponding word importance corpus:

   • The word importance corpus can be downloaded from this website: http://latlab.ist.rit.edu/lrec2018

   • The text transcripts form the Switchboard corpus can be downloaded via this link: https://www.isip.piconepress.com/projects/switchboard/releases/switchboard_word_alignments.tar.gz

   • In the “config.py” file, update the varibles shown below:

     # location of the Word Importance Corpus "annotations folder"
     wimp_corpus = --HERE-- 
     
     # location of the Switchboard transcripts
     swd_transcripts = --AND HERE--
     

2. Download glove vectors glove.6B.300d.txt from http://nlp.stanford.edu/data/glove.6B.zip and update glove_filename in config.py

3. Run the ‘build_data.py’ to prepare data for training, development and testing as:

   python build_data.py

This will create all the necessary files (such as the word vocabulary, character vocabulary and the training, development and test files) in the “$PROJECT_HOME/data/“ directory.

II. Install Python Dependencies

pip install -r requirements.txt

III. Running the model

1. Traverse inside the model you want to train and open the ‘config.py’ file and review the configurations:

   • model : type of model to run (options: lstm_crf or lstm_sig)

   • wimp_corpus : Path to the Word Importance Corpus

   • swd_transcripts : Path to the Switchboard Transcripts

   • output_path : Path to the output directory

   • model_output : Path to save the best performing model

   • log_path : Path to store the log

   • confusion_mat : Path to save the image of the confusion matrix (part of analysis on the test data)

   • compare_predictions : Path to save the predictions of the model (.csv file is produced)

   • random_seed : Random seed

   • opt_metric : Metric to evaluate the progress at each epoch

   • nclass : Num of classes for prediction

   • dim : Size of the word embeddings used in the model

   • dim_char : Size of the character embeddings

   • glove_filename : Path to the glove-embeddings file

   • trimmed_filename : Path to save the trimmed glove-embeddings

   • dev_filename : Path to the development data, used to select the best epoch

   • test_filename : Path to the test data, use for evaluating the model performance

   • train_filename : Path to the train data

   • words_filename : Path to the word vocabulary

   • tags_filename : Path to the vocabulary of the tags

   • chars_filename : Path to the vocabulary of the characters

   • train_embeddings : If True, trains the word-level embeddings

   • nepochs : Maximum number of epoches to run

   • dropout : The probability of applying dropout during training

   • batch_size : Number of examples in each batch

   • lr_method : Optimization strategy (options: adam, adagrad, sgd, rmsprop)

   • lr : Learning rate

   • lr_decay : Rate of decay of the learning rate

   • clip : Gradient clipping, if negative no clipping

   • nepoch_no_imprv : Number of epoch without improvement for early termination

   • reload : Reload the latest trained model

   • word_rnn_size : Size of the word-level LSTM hidden layers

   • char_rnn_size : Size of the char-level LSTM hidden layers

2. Run the model by:

   python main.py

Summary:

• trained model saved at “model_output” (declared inside config.py).
• log of the analysis at “log_path” (declared inside config.py) - contains train, dev and test performance.
• confusion matrix at “confusion_mat” (declared inside config.py).
• CSV file containing the actual scores and the predicted score at “compare_predictions” (declared inside config.py).

IV. Running the agreement analysis

1. Locate the csv file containing the actual scores annotated by the annotators and the predicted scores.

2. Open up “compare-annotation.R” file and update the “annotation_src” variable with this new location.