A python library / model for creating co-references between AMR graph nodes.
Project description
amr_coref
A python library / model for creating co-references between AMR graph nodes.
Install
To install:
pip install zensols.amr_coref
About
amr_coref is a python library and trained model designed to do co-referencing between Abstract Meaning Representation graphs.
The project follows the general approach of the neuralcoref project and it's excellent blog on the co-referencing. However, the model is trained to do direct co-reference resolution between graph nodes and does not depend on the sentences the graphs were created from.
The trained model achieves the following scores
MUC : R=0.647 P=0.779 F₁=0.706
B³ : R=0.633 P=0.638 F₁=0.630
CEAF_m: R=0.515 P=0.744 F₁=0.609
CEAF_e: R=0.200 P=0.734 F₁=0.306
BLANC : R=0.524 P=0.799 F₁=0.542
CoNLL-2012 average score: 0.548
Project Status
!! The following papers have GitHub projects/code that are better scoring and may be a preferable solution. See the uploaded file in #1 for a quick view of scores.
This is a fork of Brad Jascob's
amr_coref
repository, and modified to address the multiprocessing issues on
non-Debian style OSs. See #3
for details on the issue.
Usage
To turn multi-threading off, create the Interface instance with use_multithreading=False
.
Installation and usage
There is currently no pip installation. To use the library, simply clone the code and use it in place.
The pre-trained model can be downloaded from the assets section in releases.
To use the model create a data
directory and un-tar the model in it.
The script 40_Run_Inference.py
, is an example of how to use the model.
Training
If you'd like to train the model from scratch, you'll need a copy of the AMR corpus. To complete training, run the scripts in order.
- 10_Build_Model_TData.py
- 12_Build_Embeddings.py
- 14_Build_Mention_Tokens.py
- 30_Train_Model.py.
You'll need amr_annotation_3.0
and GloVe/glove.6B.50d.txt
in your data
directory
The first few scripts will create the training data in data/tdata
and the model training
script will create data/model
. Training takes less than 4 hours.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distribution
Hashes for zensols.amr_coref-0.0.2-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | cfa948bdb34be87b05b94fb15cd7c9d7d56c57bb31b2d773c9725a1b1e26b5d0 |
|
MD5 | f7dfa54ca20fc534ec283f5d417e25ab |
|
BLAKE2b-256 | 25617aa08eb2521c966d9abec0085040e10f52c269e1acda2e4b821f36f197db |