KVmemnn

Key-Value Memory Networks for Directly Reading Documents

This project contains code for the Key-Value MemN2N setup in the following paper: "Key-Value Memory Networks for Directly Reading Documents".

Setup

This code requires Torch7 and its luarocks package tds. You need to compile the c code in library/c--a script containing a default gcc command is provided as setup.sh.

Examples

This directory contains scripts for running this code on specific datasets. The initial release will include the WikiMovies dataset.

Library

This directory includes the main memory network files, listed below:

base_model.lua: top-level shared model functions, extended by specific models
cmd.lua: file for parsing options
data.lua: file for iterating over data
dict.lua: file for accessing dictionary
eval_lib.lua: functions for evaluating dev/test-time evaluation
hash.lua: provides hashing system for accessing knowledge entries
interactive_lib.lua: interactive library for stepping through individual examples
kvmemnn_model.lua: key-value memory network model
memnn_model.lua: memory network model
parse.lua: parsing methods for building dictionary and data vectors from text
PositionalEncoder.lua: implements positional encoding system from "[End-To-End Memory Networks](http://arxiv.org/abs/1503.08895)"
SumVecarr.lua: implementation of Sum nn module for vector_arrays instead of Tensors
thread_utils.lua: utilities for torch multithreading
util.lua: utility functions
vector_array.lua: auto-resizable vector array implementation
WeightedLookupTableSkinny.lua: weighted lookup table optimized for fixed dimensions

It also includes a directory named "c" which includes a number of .c files that speed up the performance of the library code. The c files need to be compiled into libmemnn.so--default gcc parameters are provided in a script setup.sh in the top-level directory.

Scripts

This directory includes scripts to access the library, listed below:

build_dict.lua: run this on text first to create a dictionary
build_data.lua: run this on text second to build vector arrays from a dictionary
build_hash.lua: run this on text third to create hash access to knowledge entries (like the KB or wikipedia for WikiMovies)
eval.lua: run this to evaluate a dataset
interactive.lua: run this to walk through examples in a dataset
train.lua: run this to start training

For examples of using these scripts, check out the examples directory.

References

• Alexander H. Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston, "Key-Value Memory Networks for Directly Reading Documents", arXiv:1604.06045 [cs.CL].