AdaElectra — Adaptive pooling based approach with Electra model for Multi Label Relation Classification
In this project, I present an Adaptive pooling based method on top of Electra model — AdaElectra in order to classify relations. The proposed model — AdaElectra achieves F1 score of 0.88 on the NYT29 dataset
The complete documentation of the project can be found here
In this project, I use the NYT29 dataset. This dataset is derived from the New York Times dataset. It has 29 relation types which are shown in the relation.txt file. We have splited the whole dataset into training (63,306 sentences with 78,973 relation tuples), development (7,033
sentences) and test (4,006 sentences and 5,859 relation tuples) sets.
Each set consists of a .sent file containing all source sentences, and a .tup file indicating the relation tuples from each of the
source sentence. Each sentence comes with one or more relation tuples in the form of (entity 1; entity 2;relation). Multiple tuples are separated with |.
A sample sentence and its corresponding relation tuples are shown below:
Sentence: then terrorism struck again , this time in the indonesia capital of jakarta.
Relation Tuples (separated by |):
jakarta ; indonesia ; /location/administrative division/country | indonesia ; jakarta ; /location/country/capital | indonesia ; jakarta ; /location/country/administrative divisions
Given all the entity pairs, the task is to classify each pair of entities of a sentence to a particular relation type or other (indicating no relation between the two entity mentions).
The source code for this project can be found at Relation Classification.ipynb.