Doubt regarding semantic parsing

[bhi5hmaraj]

First I would like to thank the authors of this tutorial/repo for this great resource . It has helped me tremendously in understanding what's semantic parsing really is .

For my project , I need to extract certain entities like name, location, organisation , etc and create a query out of it . The recognition part can be done by training a custom NER using Stanford's corenlp. The doubt I have is if my intent is narrow , then is it required to write a grammar and create the semantics recursively ? Can't I just create the semantics for a question by just checking if it has so and so entities and directly create one. For example , let's say my query is "who knows John Doe from New York" , then I can just create a semantic by just recognising the named entities . Is a grammar a overkill here or am I missing something

[usptact]

For some applications, semantic parsing can be overkill. In those cases an intent classifier + couple NER models do to slot filling will do the job. You will need to map (intent, NER tags) tuples to the commands or queries in your case.

[bhi5hmaraj]

Thought this issue was dead.

We used exactly the same approach. But the problem that we are facing now is that it's hard to train the NER and classifier to detect entities in small sentences which are worded as commands , for e.g. "get me x " . Right now we are just adding these sentences manually to the training data. Will having a hybrid system of both semantic parsing and intent classifier + NER method make it more robust ? At least we can be sure that some questions are sure to work this way .

As the project we are working deals with conversational text , we don't expect the users to follow casing rules. Do you have any good resources to tackle NER on noisy text (uncased, bad grammar , spelling errors). The closest I could find was Workshop on Noisy User-generated Text (W-NUT) and this blog . But I am yet to try the methods stated there in our project. CoreNLP tackles it by either using a trucaser or by training a caseless models. Truecaser is introducing more errors and hence affecting the accuracy . So we created a data set with entities in all 3 cases (upper, lower, title) . Do you know of any noisy text conversational annotated data-set which has POL tags . If there are any standard ways of handling this situation, please let me know.

[usptact]

Looks like we are working on similar problems :)

Unfortunately, I can't answer your questions. To my best knowledge, those are still research questions. If one can provide robust solutions, those would be worth of pure gold.

A hybrid system can be one approach. I see building an arbitrator that decides which system output to use or even better, try to merge the results into one.

The availability of public datasets is a great challenge, especially for imperative language (eg. commands to Alexa). One way would be to build scenarios and ask crowd workers to have short, guided conversations using a keyboard with an agent. Just some wild ideas...

Thanks for the two links!