I am trying to make a simple chatbot for Frequently Asked Questions about any particular thing,  for example some college admission. Most of the questions are unique and have no relation to each other. On being fed such questions and answers the bot tries to match the closest question. Currently I am using topic modelling( tfidf followed by lsa) with only the nouns and adj. in the input sentences. The topic value is set high. It is not very good in terms of accuracy.

I also tried parsing the sentence and finding out the root verb and its direct objects and its modifiers (using a dependancy parser) and then put a weight on each of them depending on their type, following which I did a straight match with the lemmatized input senteces and come up with a ratio based on the degree of overlapping words and their weights. But it was not scaling with a larger dataset.

Because there only 1 or 2 sentences about a topic I haven’t used classifiers.

Is there any one who has tried to solve a similar problem or someone knows some different approach to tackle this problem.

thanks

You may want to look at the Bot Libre AI engine on GitHub, it does something similar.

http://www.botlibre.org/

Basically it indexes all of the words from the questions, then for the user input finds the FAQ question with the most similar words.  Different words types (noun, verb, adjective) are scored differently to find the best matching question.

It also uses keywords, which if you can classify some words from the FAQ’s domain as keywords would help a lot.

Keyword and key phrase processing in ‘Narwhal’ might be an easy solution. If you are in Python, or can access it, take a look at:

https://github.com/peterwaksman/Narwhal

A Narwhal class to handle FAQ is pretty straightforward: organize the input keywords and phrases, and you are done. There is no need for parts of speech and grammar.

Narwhal is new, I am interesting in FAQ applications, and I would be glad to help if you feel like giving it a try.

Here is a new Github project QnA Baker, which uses answer retrieval and ranking with k-NN Search. https://github.com/a-mma/QnA-baker