In continuation with the problems with the version-1 of our model (which was getting 11% BLEU accuracy), I discussed the output with mentors. After a day or two of error analysis, we realised that there were few trivial problems with our dataset, due to which our BLEU score was so low:
There were a few minor problems: such as some NLQs just did not made sense were noisy - for which we can keep a threshold frequency, but we chose to not do it right now.
After solving the above two troubles, we were good to prepare data and train the version-2 of our model. We both were hoping we get far better than mere 11%, because we guessed that “1” should be the major problem.
After again a wait of 9 hours, we got ~80% BLUE accuracy for our model and this marked the correctness and completeness of the project. Training Setup and Graph Analysis:
This marks the end of a end-to-end system which generates dataset automatically for NSpM learner.
With nearly two and a half week until the final deadline of GSoC tenure, we decided to try a ML experiment to test whether a machine learning model can learn compositionality. Precisely, the goal of this experiment is as follows: The goal is to train a model on data {‘a1’, …, ‘an’, ‘b’, ‘a1○b’, …, ‘an○b’, ‘c’,’d’} and test on data {‘c○d’}. The idea is to check whether the Neural MT model can learn how to translate composite NLQ queries into SPARQL by learning on other composite NLQ-SPARQL examples.
For now, me and my mentors mutually decided to perform our experiment on a simpler setting, where ‘d’ is equal to ‘b’.
Formally setting one becomes: Train a model on data {‘a1’, …, ‘an’, ‘b’, ‘a1○b’, …, ‘an○b’, ‘c’} and test on data {‘c○b’}.
For eg:
Trainset:
a1 := “what is the county of
Testset:
c○b := “where is the district of