Finding the Optimal Summary By Combining Extractive and Abstractive Summarization Methods.

In modern-day society, we are surrounded by data, most of which is too long and laborious for the everyday person to read comprehensively. With the art of generating extractive summaries becoming increasingly trivial, the challenge of producing abstractive summaries that can understand and convey the meaning of a document remains prominent within the field of natural language processing. Although summarisation is a task that humans can easily complete, developing automated approaches that can generalise well with different data formats remains a challenge. Despite a recent surge in research to improve abstractive summarisation methods, there are still outstanding problems regarding the validity of generated summaries. This method experiments with combining extractive and abstractive summarisation methods, to evaluate whether extractive methods can be effective in driving more human-like abstractive summaries. The overall aims and evaluators are the reduction of repetition, and the improvement of precision, recall and human-judged grammatical correctness.

Setup

• This project requires Python 3.0+ to run.
• Install required dependencies (needed for the first run to ensure NLTK packages are installed)

  make setup 
  

• Install required python packages
  only run if requirements have changed since make setup has been run + no new NLTK packages are required to be installed

  make requirements
  

• To run the GloVe model, download the pre-trained word embeddings found here and ensure they are placed on the specified path in a file named "glove".

Run

• To run data processing with default settings (as shown below)

  make run-data
  

• To run the model with default settings - bidirectional model

  make run-model
  

• To alter runtime variables follow the following examples

  make run-data TRAIN_DATA_PATH=<> OUTPUT_CSV=<> ...
  make run-data MODEL_ID="1" CSV_NAME="daily-mail.csv" ...
  

Note: For any machine running python 3.0 via the python3 command, append -labs to the end of any make commands (e.g. make-setup-labs)

The models can also be run as Jupyter Notebooks (found in the ./notebooks directory). To do this, you must have Jupyter installed.

MakeFile

The default settings for all runtime variables that can be altered are shown below.

    ## Variables
    ### Data Processing
    TRAIN_DATA_PATH="./cnn/originals"
    OUTPUT_CSV="./data/cnn-all.csv"

    STOP_WORDS=True
    LEMMATIZE=True
    LEMMATIZE_WITH_POS=True
    SENT_POS=False
    #### Only one of these can be true at any given time
    TEXT_RANK=False
    WORD_FREQ=False

    ### Model Running
    MODEL_ID=1  # 0 = unidirectional, 1=bidirectional, 2=GloVe model
    WORD_REMOVAL=False # remove words using uncommon_word_thr
    CSV_NAME="cnn-all.csv" # csv data to run model against

Notebooks

There are three notebooks for this project: bi_model.ipynb glove_model.ipynb uni_model.ipynb.
Each notebook contains the exact same code as the relevant class in models.py and the shared methods in modelCommon.py.
As with the python scripts, the required hyperparameters, packages, and directories are required in order to run these.

Flow

• dataProcesing.py
  • Read in Data
  • Clean Data
    • Stop Word Removal
    • POS
    • Lemmatization
  • Extractive Methods
    • TextRank?
    • Word Frequency?
    • Sentence Position?
  • Write Dataframme to CSV
• modelCommon.py and models.py
  • Read CSV into Dataframe
  • Data Cleaning
  • Uncommon Word Removal?
  • Max Text Lengths
  • Training Validation Split
  • Word Embeddings
    • Reshape Data
    • Learning Model
      • Encoder
      • Decoder
      • Combined LSTM Model
      • Training
    • Inference Model
      • Encoder
      • Decoder
      • Reverse Word Embeddings
    • Evaluation
      • Validation Data
      • Training Data
      • Test Data