Multiple-choice question answering for ARC Challenge

• Structure
• Stats
• Results
• Running the model
• Paper

Structure

• answer/ - the neural network that combines all the solvers;
• essential_terms/ - the model that assigns essential scores for each term in a question;
• multinli/ - the neural network trained on the MultiNLI dataset;
• nlp_inference/ - the neural network trained on the SNLI dataset;
• scitail/ - the neural network trained on the SciTail dataset;
• rephrase/ - the module that transforms questions into affirmative sentences;
• qa/ - pre-trained neural networks on the SQuAD v1 dataset;
• protos/ - protocol buffer definitions (for passing data around);
• wikipedia_indexer/ - code for indexing Wikipedia dumps and science book collection. It also includes the IR solvers and candidate context extraction tool.

Stats

--------------------------------------------------------------------------------
 Language             Files        Lines        Blank      Comment         Code
--------------------------------------------------------------------------------
 Java                    63        19577         1807         2480        15290
 Python                 139        20634         2809         4141        13684
 XML                      1          197            5            0          192
 Protobuf                 4          118           25            2           91
 Markdown                 2          104           29            0           75
 Bourne Shell             2           50            9            6           35
 Makefile                 2           47           17            0           30
--------------------------------------------------------------------------------
 Total                  213        40727         4701         6629        29397
--------------------------------------------------------------------------------

Collected using cgag/loc

Results

Dataset	Accuracy
ARC-Easy Test	60.943%
ARC-Challenge Test	26.706%

Running the model

Please contact [email protected] if you want to build and run the model. There are a lot of dependencies (both software and data sources) that need to be installed.

Paper

Improving Retrieval-Based Question Answering with Deep Inference Models