Neural Entity Typing with Knowledge Attention

This repo contains the source code and dataset for the following paper:

• Ji Xin, Yankai Lin, Zhiyuan Liu, Maosong Sun. Improving Neural Fine-Grained Entity Typing with Knowledge Attention. The 32nd AAAI Conference on Artificial Intelligence (AAAI 2018) pdf.

How to use our code for KNET

Prerequisite

• python 2.7.6
• numpy >=1.13.3
• tensorflow 0.12.1
  • can be done by pip install https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl

All the codes are tested under Ubuntu 16.04.

Data

Data files should be put in the data/ folder.

• disamb_file, containing information for disambiguation, is already in data/. Please unzip it.
• Train, valid and test set data are also in data/. Please unzip them.
• For the word vector file, we recommend using Glove from http://nlp.stanford.edu/data/glove.840B.300d.zip . Please download, unzip, and put it in data/.
• types records all they types in the taxonomy (only for recording; not used in the code).

Parameters

• Parameters saved from training is in the parameter/ folder, but you can also choose a new location.
• We provide parameters for the model shown in our paper in the paper_parameter/ folder.

Usage

Detailed usage can be found by running python src/run.py --help.

Quick start: simply run ./run.sh.

For training and testing, follow the example of line 5 and 6 in run.sh.

How to direclty use the code for typing

1. Organize input data in .npy format. See https://github.com/thunlp/KNET/issues/1 for instructions.

   Another example is in the direct/ folder.

   • every sentence occupies three lines in raw. The first line is the entity mention, the second is left context, the third is right context. Words are separated with spaces.
   • run raw2npy.py. It's better to use the same python version with step 2 to avoid encoding issues.

2. Follow the example of line 7 in run.sh.