You can keep things very easy by analyzing the

sentences with regular expressions. I think it

is possible to access so-called named groups

in Java like (?X), where X stands for

a pattern. For each type of sentence, write one

regular expression. Present you user a list

of example sentences that could be matched

by some regex. For example (C#):

When does analysis II start in summer of 2016?

"When does (?.+I*?) start in summer of \d+\?"

private void button1_Click(object sender, EventArgs e)

{

string s1 = "When does (?.+I*?) start in summer of (?\\d+)\\?";

string s2 = "When does analysis II start in summer of 2016?";

Regex regex = new Regex(s1);

Match m = regex.Match(s2);

if (!m.Success)

return; // error;

string course = m.Groups["course"].Value;

string year = m.Groups["year"].Value;

string query = "select * from SYLLABUS where COURSE=" + course + " AND YEAR=" + year;

MessageBox.Show(query);

}

Of course you can also use a natural language parser,

which is, but that is rather rarely needed in

modern speech, especially when it comes to such

simple topics as schedules, able to handle recursion.

Regards,

Joachim

Thank you so much Joachim for replying to my question and i now have a general idea about how i am supposed to do my implementation , one little problem i still have is what if notion of question asked can be changed and i don't want it to limit on number of patterns to match a single notion of question , i want the system to learn different patterns himself  that's why i wanted to perform NLP and train my system for such type of questions.

Actually, I do not agree with Joachim's comment above - rule-based systems, including regexp-pattern-based ones, will ON THEIR OWN not be able to deliver a satisfactory performance. Reason is that language is too variable to be captured in these patterns (a lot of mateial will just not be matched) and adding patterns gets increasingly difficult since matching regions start to overlap and interact with each other. Do this only if you have only very little time and you want to show the general feasibility in the short term.

Much more promising is machine learning, as impressively demonstrated by IBM Watson's Jeopardy game in 2011. This system used some pattern based rules, but also a plethora of other signals, and was trained on correct question-answer pairs.

There is an open source implementation in java of a similar system called YodaQA, and I would recommend to take a look:

https://github.com/brmson/yodaqa

http://ailao.eu/yodaqa/yodaqa-poster2015.pdf

best, Chris

chris Thank you for replying to my question , but unfortunately yodaqa is still experimental and there are no steps to follow to implement or use yodaqa is java project . i was hoping to get more precise answer to further start my research regarding implementing this in java

The question is also what learning could be

in your case. If a user enters a question that

isn't recognized, it's too late for him. Your

program could store all sentences entered in

a file or database, which over time accumulates

a corpus for learning, but it will stay rather

a small one. How could you get training examples?

Maybe by inventing questions instead of real

users entering them. Or you could spider the

web, but in that case you will harvest lots of

sentences that have not much to do with your

domain of interest.

Regards,

Joachim