Skip to content

Latest commit

 

History

History
121 lines (97 loc) · 4.24 KB

README.md

File metadata and controls

121 lines (97 loc) · 4.24 KB
Raw

Code for running the entity linking model. This is part of the code for the xelms project.

Requirements

  1. pytorch (0.2.0+21f8ad4): installed from source, and patched for sparse tensor operations (instructions below).
  2. python3.
  3. cogcomp-nlpy.
  4. Download the resources and trained models here and place them in the folder xling-el/data. Right now, pre-trained models are available for German, Spanish, French, Italian, and Chinese.

Resources for Candidate Generation

  1. First set up candidate generation and other resources as described in projects wikidump_preprocessing and wiki_candgen.
  2. A mongo daemon needs to be running. This is where the resources generated in wiki_candgen will be kept for fast (and parallel) access.

Note: These resources are provided in the resources directory downloaded in step 4. above, so ideally you do not need to regenerate them, unless you plan to use a newer Wikipedia dump or a larger knowledge base.

Patching Pytorch for Sparse Tensor Operations

This is best done in a new conda environment.

  1. First checkout the sparse_patch branch from this repository.
git clone https://github.com/shyamupa/pytorch
cd pytorch
git checkout sparse_patch
  1. Install the patched code from source using the following commands,
export CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" # [anaconda root directory]

# Install basic dependencies
conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing
conda install -c mingfeima mkldnn
cd pytorch_patched
python setup.py install

Ensure that the patched pytorch was successfully installed,

>>> import torch
>>> torch.__version__
'0.2.0+43662e7'

Mention Detection using NER

  1. For German, Spanish, French and Italian, download relevant Spacy NER Models
pip install spacy
python -m spacy download de_core_news_sm
python -m spacy download es_core_news_md
python -m spacy download fr_core_news_md
python -m spacy download it_core_news_sm
  1. For Chinese, download stanford corenlp jar and the chinese model jar and place them in a stanford_jars directory.
$ ls stanford_jars/
stanford-corenlp-full-2018-10-05
$ ls stanford_jars/stanford-corenlp-full-2018-10-05
...
...
stanford-chinese-corenlp-2018-10-05-models.jar
...

And set the bash environment variable CORENLP_HOME to path/to/stanford_jars/stanford-corenlp-full-2018-10-05.

export CORENLP_HOME=path/to/stanford_jars/stanford-corenlp-full-2018-10-05

Running the Model

To run the model, use the command,

./run_inference_on_doc.sh <lang> <infile> <outfile>

For instance, for running on a German document test_docs/de_doc.txt, one would run

./run_inference_on_doc.sh de test_docs/de_doc.txt test_docs/de_doc_output.txt

The json output will be produced in test_docs/de_doc_output.txt.

Output

The output file is a json serialized text annotation, with a view named NEURAL_XEL_<lang>. The view consists of a list of the constituents that have been linked to a Wikipedia title. Below is the output for the German test document provided in the repo,

...
"viewName": "NEURAL_XEL_de",
...
...
"constituents": [
      {
       "end": 2,
       "label": "en.wikipedia.org/wiki/Angela_Merkel",
       "score": 0.5128146075318596,
       "start": 0,
       "tokens": "Angela Merkel"
      },
      {
       "end": 5,
       "label": "NULLTITLE",
       "score": 0.05000000074505806,
       "start": 4,
       "tokens": "Elim-Krankenhaus"
      },
      ...

The label field for each constituent is the predicted Wikipedia entity for the span identified by the start and end token index. Here a label of NULLTITLE means that the named entity detected by the mention detection system could not be linked to any entity.