gabrielstanovsky / supervised-oie Goto Github PK

Code for training a Neural Open IE model (NAACL2018)

License: MIT License

Python 85.45% Shell 14.55%

supervised-oie's Introduction

Deprecated!

The maintenance of this project has moved to the AllenNLP framework.
Where you can use the model and an online demo. This thin wrapper may also be useful if you want to run the pretrained model.

Table of Contents generated with DocToc

supervised-oie

supervised-oie

Code for training a supervised Neural Open IE model, as described in our NAACL2018 paper.
🚧 Still under construction 🚧

Citing 🔖

If you use this software, please cite:

@InProceedings{Stanovsky2018NAACL,
  author    = {Gabriel Stanovsky and Julian Michael and Luke Zettlemoyer and Ido Dagan},
  title     = {Supervised Open Information Extraction},
  booktitle = {Proceedings of The 16th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL HLT)},
  month     = {June},
  year      = {2018},
  address   = {New Orleans, Louisiana},
  publisher = {Association for Computational Linguistics},
  pages     = {(to appear)},
}

Quickstart 🐣

Install requirements 🙇

pip install requirements.txt

Download embeddings 🚶

cd ./pretrained_word_embeddings/
./download_external.sh

Train model 🏃

cd ./src
python  ./rnn/confidence_model.py  --train=../data/train.conll  --dev=../data/dev.conll  --test=../data/test.conll --load_hyperparams=../hyerparams/confidence.json```

NOTE: Models are saved by default to the models dir, unless a "--saveto" command line argument is passed. See confidence_model.py for more details.

Predict with a trained model 👏

python ./trained_oie_extractor.py \
    --model=path/to/model \
    --in=path/to/raw/sentences
    --out=path/to/output/file
    --conll

More scripts 🚴

See src/scripts for more handy scripts. Additional documentation coming soon!

supervised-oie's People

Contributors

Stargazers

Watchers

supervised-oie's Issues

'NoneType' object is not iterable

Hi,

When I am trying to train the model. I got the TypeError problem. Can anyone help? Thanks in advance!!!

Traceback (most recent call last): File "./rnn/confidence_model.py", line 454, in <module> rnn.plot("./confidence_model.jpg", train_fn) File "./rnn/confidence_model.py", line 122, in plot X, Y = self.load_dataset(train_fn) File "./rnn/confidence_model.py", line 139, in load_dataset self.encode_outputs(sents)) File "./rnn/confidence_model.py", line 208, in encode_outputs list(self.transform_labels(sent.label.values)), File "./rnn/confidence_model.py", line 110, in transform_labels classes = list(self.classes_()) TypeError: 'NoneType' object is not iterable

Some questions about training

Dear author,

Firstly, is your training code in Allennlp open? In the link you give, I can only find the prediction instruction and an online demo but without the training instruction.

Secondly, in your training data "train.oie.conll", some instances are exactly the same including the predicate except for their labels. Taking the example from your paper "Barack Obama, a former U.S president, was born in Hawaii.". It can generate 2 training instances with the same predicate "was born in" but with different A0 tags("Brarack Obama" and "a former U.S president"). According to your model, the inputs to the model of this 2 instances are exactly the same but you expect that it will output 2 different sets of label. Could you shed some light to this issue?

Many thanks!

can't train the model

there is no training process when I run the code following the readme.
I find something weird in the confidence_model.py,

X, Y = self.load_dataset(train_fn)

but there is no place to use X, Y?

Error while training

Traceback (most recent call last):
File "./rnn/confidence_model.py", line 462, in
rnn.set_model()
File "./rnn/confidence_model.py", line 378, in set_model
self.model = Model(input = map(itemgetter(0), true_input + corrupt_input),output = [output])
File "/home/amrit/anaconda/lib/python3.6/site-packages/keras/engine/topology.py", line 1794, in init
' (missing Keras metadata).')
TypeError: Input tensors to a Model must be Keras tensors. Found: <map object at 0x7fbb3ee77ef0> (missing Keras metadata)

How to get oie triples from QAMR

Hi,

Could you please provide the oie triples from QAMR?

./question_distributions/dist_wh_sbj_obj1.json can't find

hello,
Please tell me where can find the file of the ./question_distributions/dist_wh_sbj_obj1.json

evaluation script

Hi,

From your NAACL paper,

To that end, we follow He et al. (2015) which judge an argument
as correct if and only if it includes the syntactic head of the gold argument (and similarly for
predicates).For OIE2016, we use the available Penn Treebank gold syntactic trees (Marcus et al., 1993),
 while for the other test sets, we use predicted trees instead.

I wonder if you have the matching evaluation script with the tree as the input in this repo?
Seems in supervised-oie-benchmark/matcher.py
Matcher.argMatch
Matcher.predMatch
Matcher.argMatch
Matcher.lexicalMatch

None of them correspond to that?

Thanks!

Installation documentation

Hi @gabrielStanovsky,

I just installed this software package and found a couple instructions that need to be updated.

pip install -r requirements.txt
hyperparam directory is missing a p in the command line example
The training, dev, and test files should have a oie in the middle, e.g. train.conll should be train.oie.conll

Users will also need to install a tagger using NLTK:
nltk.download("averaged_perceptron_tagger")

Thanks for releasing this!

the file of dist_wh_sbj_obj1.json can't find

@gabrielStanovsky
can you tell me where can find the --dist=./question_distributions/dist_wh_sbj_obj1.json

pretraining model

hello everyone, hope that you are doing well,

I want just to know if exists an pretrainnig model for OIE

thanks,

Statistics of AW-OIE dataset

What are the #sents and #extractions in train, test, and dev set respectively?

I only found that 12,952 tuples are yielded from QAMR in Section 5.2, but no overall statistics.

Another paper, LSOIE: A Large-Scale Dataset for Supervised Open Information Extraction, shows statistics of several datasets in Table 1. However, their #sents and #extractions (3,180 & 8,477) seem to be incorrect for OIE2016 because 3,200 & 10,359 are reported in the original paper.

Therefore, I am wondering if 3,300 & 17,165 are correct for AW-OIE or not.

How to convert BIO output into open IE (subject, predicate, object) tuples

Hi - I'm using the allennlp implementation and trying to figure out how to properly convert from BIO output into (subject, predicate, object) tuples.

It is clear when you have ARG0 V ARG1, but less clear with the following:

[ARG1: England] is [V: separated] [ARG2: from continental Europe] [ARG0: by The Irish Sea] [ARG2: to] the east and the [ARG0: English Channel] [ARG2: to the south]

I would appreciate some guidance on how to parse this.