This project aims to create a paragraph from 6 random sentences.
We have half a million dictionaries in train data, including a list of sentences and a list of the index that shows the correct order.
Due to lack of memory and GPU, a sampled data is created from the original dataset. We map the ordering sentences problem to a classifier with two possible classes. Either two sentences are continuous or not. For each paragraph, we consider the combination of 2 from 6 sentences which 5 of them have the label True, and the rest are False. Then, we finetune a BertForSequenceClassification from huggingface library.
To order sentences in the test step, we consider all possible combinations of 6 sentences and, for all of them, calculate the sum of the probability of pair sentences, then choose the combination that leverages the max score.
To get the result, just run the following command.
chmod +x run.py
./run.py