HINT GENERATION
Hint Generation for short answer Questions.
Recently working mostly in NLP related task and was involved in the project where we had to generate hint for the short questions. Here I will be discussing the approach rather than solution, What went wrong and how was it corrected.
So, as fast as you think about the task your mind will get clicked with Context aware answer generation approach, where we provide the answer and context as the input to the encoder and decoder outputs the hint as the answer.(like typical question answer generation) model, but using this approach it require more hardware power and should also have low inference time.
Using typically BERT model will cost you time, and computation as, BERT itself is around 1gb in hugging face and average inference time is around 2–4 second, which is far more.
So, we should think of the other approach rather than this.
APPROACHS
As, hint are with in the answer and should also make sense to the people using it so why not take it out ourselves, but how ?
Here, we can use the skipgrams, word2vec kind of concept where calculating the dependencies between 2 words, we could extract a part of sentence which could hold the meaning. If the two words are far close to each other, we extract the words between those two words, inclusive. This way we get the part of the sentence which hold meaning.
But her is one issue with this approach, that it might extract the part of sentence which has punctuation in it.
For example: sentence like
Ram is the good boy, and he is good at sports.
Here, just assume “is” is dependent in he, so our hint might be “is the good boy, and he” so here hint is not given in structurally. The possible solution for this can be removing the stop words before finding dependencies. That way we can remove those issue.
Also,
We can iterate till we get the noun/verb phrase in staring from front and back and exclude all before that word. That will make it more better.
The solutions are endless.
Want to connect …
