DeepScite - A Simple Convolutional-based Recommendation Model

Overview

DeepScite takes in papers (titles, abstracts) and emits recommendations on whether or not they should be scited by the particular users whose data we've used for training (in the case of this repo, it is me).

As output, it also gives a "goodness" score for each word; when this number is high, it has contributed strongly to the paper being (recommended) for sciting, when it is negative, it has contributed strongly to the paper not being recommended.

Below are some example outputs of the system:

The blue text are those words which are "good", and the red text are those which are "bad".

Installation

1. Clone this repository:

git clone https://github.com/silky/deep-scite.git

1. Use conda or (virtualenv) and create an environment that has Python 3.5.

   conda create -n deep-scite python=3.5

2. Activate the environment

   source activate deep-scite

3. Install the requirements

pip install -r requirements.txt

1. Install nltk language packs

In order to tokenise strings, we use the nltk package. It requires us to download some data before using it though. To do so, run:

python -c 'import nltk; nltk.download("punkt")'

1. Install this library in develop mode

python setup.py develop

Usage

From the root directory of this project:

1. Activate the deep-scite environment

source activate deep-scite

2. Train the model on the noon data set, and emit recommendations

./bin/run_model.py

This will run through the steps defined in model.yaml.

3. Open up ./data/noon/report.html in your browser and observe recommendations.

Misc

You can play around with the embedding by looking at it in TensorBoard. Run TensorBoard with:

tensorboard --logdir /tmp/tf-checkpoints/deepscite-noon

Then click on the "Embedding" tab.