morningmoni / taxorl

Code for paper "End-to-End Reinforcement Learning for Automatic Taxonomy Induction", ACL 2018

Connect with me:

taxorl's Issues

Wordnet subtrees Visualization

Hey, How to visualize the wordnet subtrees, what tools did you use ? Thanks!

Use the saved model for inference on a new set of terms

Hi,
i builded everything from scratch for Italian language, ran the train_RL.py using the italian version of SemEval-2016 datasets for training, validation and test and saved the model. I can't figure out how to use the saved model for taxonomy induction on my own vocabulary. Thanks.

Hello, where is the relations.txt? Thanks very much.

Ferequency files

I am applying your method to my own dataset, so I need to generate my own frequency files.
How do you generate the twodatasets_freq_w.pkl & other frequency files?

corpus

The datasets in corpus are used in knowledge_resource.py, like _term_to_id.db, _id_to_term.db, path_to_id.db, _id_to_path.db, _l2r.db etc. I know from your source code these db files are created by yourself. However, I can’t understand what input is in the shell file in the corpus folder. For example, the parameters like $1, $2, $3, $4 are input from the command line, but I don’t What files do these parameters represent? I will be grateful you if you can provide your created db file. Thanks very much.

no generalization

Hi,

I am trying to run training using the features from the pickle file. The performance on the training set keeps going up, however there seems to be no generalization (the performance on the evaluation set remains constant). I was wondering if you used different hyperparameter values or training options than the default ones to get your results?

pickled_data missing

Hi,
Is there any place to download "pickled_data/lower2original.pkl" data mentioned in the utils_common.py file.

Could not find SemEval tree file

As you wrote in train_RL.py:

        trees_semeval = read_edge_files("../datasets/SemEval-2016/original/",
                                        given_root=True, filter_root=args.filter_root, allow_up=False)

It seems you are trying to read tree.ptb from a directory. So should there be a tree file for semEval and how could I find/generate it?

What is the training preprocess data for WordNet only/WordNet +SemEval

It makes me really confused when I try to do with the semEval .
For wordNet only, I used valname = dev_wnbo_hyper, model_prefix_file = 3in1_subseqFeat
For wordNet+SemEval, I used valname = dev_twodatasets, model_prefix_file = twodatasets_subseqFeat.
Am I doing right?
By the way, why is dev_twodatasets so big compared to dev_wnbo_hyper. And how do you generate it?

How to train with GPU

It seems the model_RL.py is based on cpu training? How could I train with GPU?
I commented dyparams.set_cpu_mode() and add the follow codes before import _dynet

import dynet_config
# Declare GPU as the default device type
dynet_config.set_gpu()

But the training speed is still slow. Please correct me if I do it wrong.

How to generate files lower2original ？

Could you tell me how to generate lower2original.pkl？
Thank you！

corpus

Hello, morningmoni, does the datasets in corpus need to create with ourselves? or can you open this datasets? Thank you very much.