End-To-End Memory Networks for Question Answering on Tables

For more details see TableQA: Question Answering on Tabular Data

Demo

• The demo is deployed at https://svakulenko.ai.wu.ac.at/tableqa

• Command to run it locally:

python -m demo.qa

Requirements

• Python 2.7

• Numpy, Flask (only for web-based demo)

• fastText (+cython)

can be installed via pip:

$ sudo pip install -r requirements.txt

Pre-requisite

1. Provide the correct path to the pre-trained fastText model (clone fastText Github repository and run ./word-vector-example.sh to train a model on the English Wikipedia) download pre-trained model ('make clean' and 'unzip enwik9.zip' before)

• To train and evaluate the model use tableQA_runner.py

Experiments

Data

Synthetic data based on a real table (limiting vocabulary size and producing more training examples)

Cell-based formatting

Dictionary: 65

1 Row1 LAU2_NAME Allhaming

2 Row1 YEAR 2002

3 Row1 INTERNAL_MIG_IMMIGRATION 2

4 Row1 INTERNATIONAL_MIG_IMMIGRATION 4

5 Row2 LAU2_NAME Geretsberg

6 Row2 YEAR 2005

7 Row2 INTERNAL_MIG_IMMIGRATION 3

8 Row2 INTERNATIONAL_MIG_IMMIGRATION 5

9 What is the INTERNAL_MIG_IMMIGRATION for Geretsberg? 3 5 7

Settings

• linear start

• 2 question templates

• BOW

Results

1. Simple key

What is the EMIGRATION_TOTAL for Helfenberg? 2 13 20

IMMIGRATION_TOTAL in Burgkirchen? 4 3 7

Number of training examples 5949

20 + 9 epochs

train error: 0 | val error: 0

2. Complex key

What is the INTERNAL_MIG_IMMIGRATION for Grieskirchen in 2004? 4 13 14 15

IMMIGRATION_TOTAL in Burgkirchen for 2002? 10240 23 24 27

Number of training examples 18953

20 + 68 epochs

train error: 0 | val error: 0

Acknowledgments

• Based on Vinh Khuc: MemN2N for babi tasks.

• Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, Rob Fergus,End-To-End Memory Networks

Related Work

• YerevaNN: Dynamic-memory-networks-in-Theano

• Kumar et al. Dynamic memory networks