deeplearnxmu / nseg

Code for “Graph based Neural Sentence Ordering” (IJCAI2019)

Graph based Neural Sentence Ordering

Installation

The following packages are needed:

• Python == 3.6
• Pytorch >= 1.0
• torchtext == 0.3
• Stanford POS tagger or Dependency Parser
• Glove (100 dim)

Dataset Format

*.lower: each line is a document: sentence_0 sentence_1 sentence_2

*.eg: entity1:i-r means entity1 is in the sentence_i and its role is r.

Other datasets are easy to access and process. We also recommand a high-quality dataset for sentence ordering, ROC story.

Preprocessing

Use a dependency parser to get POS and syntax

Select the word as entity if the POS is noun

Find the nsubj and dobj to get the roles ( or just use a POS tagger and ignore the roles if you think the dependency parser is time-consuming)

Training and Evaluation

bash run.sh

nseg's Issues

generator.py, line 31, in step candidate = prev_candidates[b_ix] + [t_ix]. TypeError: list indices must be integers or slices, not float

Hello, I'm interested in your work. But when I'm running the run.sh, I found this error and it seems that the beam_ix and token_ix in generator.py is float, not int. For example, the beam_ix is tensor([0.8333, 0.1667, 0.5000, 0.0000, 0.3333, 0.6667], device='cuda:0'), I wonder how I can fix it. Thanks

Can you provide the result generate by original data

nothing happen~

Could you please add some comments to the code

Mismatch between test data file and test data decoding

Hi

I want to test the model on paragraphs of sentences which are in some random order and want the correct sentence order predicted from the model.

For the NIPS dataset, I assume the test.lower file consists of gold paragraph sentences in correct order while the mdl.devorder file (created after model training) consists of the output from model (predicted values|||truth values).
But the gold paragraph sentence order does not match with the truth values order.
Could you please specify how does the decoding work for test data?

Cannot get the same result.

When training on nips dataset, my performance get stuck at around acc of 54%, much less than the result in the paper. I want to now how can I improve it.

How to create pre-processed data for a new dataset?

Could you share the script used for pre-processing the data for creating '.eg', '.lower' and 'vocab.pt' files?

Could you provide the Graph-SE version code?

Hi, recently we want to apply your model in a text ordering pipeline. But we found the code in the repo is a simplified version in the original paper called F-graph. Would you mind sharing us the Graph-SE version of code?

这个nips数据集不是原版的

测试集少了很多条。用这版数据集确实能复现效果，如果用原版的测试集，效果会差很多

How to run for a custom dataset

Hi,
I have sentences in this order(example):
S1,S2,S3
S4,S1,S5

1.How can I convert the above sentence order to run this algorithm??
2.Can I use this algorithm to predict next sentence??
3.If I give a new sentence S6 how to predict next sentence??

Clarification regarding entity roles

The preprocessed file for entities looks like this as mentioned in the ReadMe:
*.eg: entity1:i-r means entity1 is in the sentence_i and its role is r.
The values for roles (r) in the sample nips entity files are 1/2/3. Could you please clarify what exact role does each of these numbers refer to?
As I understand from the paper, are they labeled as: 1-subject, 2-object, 3-other?
Thanks!

Would you mind share the data or at least a sample of the data you use?

I and my colleges are trying to replicate this work recently, but we are confusing about the text format. Would you mind release some sample data for us?

Can't get attribute '_default_unk_index'

'bash run.sh' failed with following error:
Traceback (most recent call last):
File "main.py", line 203, in
DOC.vocab = torch.load(args.vocab)
File "anaconda3/envs/NSEG/lib/python3.6/site-packages/torch/serialization.py", line 387, in load
return _load(f, map_location, pickle_module, **pickle_load_args)
File "anaconda3/envs/NSEG/lib/python3.6/site-packages/torch/serialization.py", line 574, in _load
result = unpickler.load()
AttributeError: Can't get attribute '_default_unk_index' on <module 'torchtext.vocab' from '/anaconda3/envs/NSEG/lib/python3.6/site-packages/torchtext/vocab.py'>