Label-Propagation with Seq2Seq

0. Introduction

This project implements label propagation with seq2seq. It applies Neural Graph Machines to Seq2Seq.

Problem Settings

• Semi-supervised learning techniques such as label propagation are used to solve classification (finite-categories) problems.
• And these approaches produce improvements at some extent.
• However, these approaches are not applied well to solve continuous target (infinite-categories) problems.
• Therefore, I want to tackle this problem using Neural Graph Machines in this project.

Some Details

• I test the performance in Neural Machine Translation problem.
• For calculating distance between nodes, I use L1, L2, and Mahalanobis distance metrics.
• Presentation info is given in https://goo.gl/whAbB1

---

You can explore the whole project code by following jupyter notebook codes.

toy_example.ipynb : contruct 2D sinc function with biased parallel data and unbiased non-parallel data

preprocessing.ipynb : preprocess sentences

graph_operations.ipynb : construct graph from source sentences

neural_graph_machines-benchmark.ipynb : Default Encoder-Attention-Decoder Neural Translation Model

neural_graph_machines.ipynb : Neural Graph Machine Model

1. Data Preparation

Experiments are done with IWSLT English-Vietnamese data set.

You can download using download.sh

I also use monolingual data from http://www.manythings.org/anki/.

After that, run preprocessing.ipynb

2. Graph Construction

run graph_operations.ipynb

3. Experiments

run neural_graph_machines-benchmark.ipynb and neural_graph_machines.ipynb