I have 1000 data sentences in Turkish like "a esittir b arti c". The example sentence means "a = b + c". I basically want to translate mathematical Turkish sentences into math equations.
For example, i have 6 sentence data.
- sentence ("a esittir b arti c") means "a = b + c"
- sentence ("b esittir a arti d") means "b = a + d"
- sentence ("a esittir c arti d") means "a = c + d"
- sentence ("c esittir b arti b") means "c = b + b"
- sentence ("d esittir b eksi c") means "d = b - c"
- sentence ("d esittir a arti c") means "d = a + c"
After I train my neural network according to data above, when I want the result of "d esittir a arti b", It doesn't give me "d = a + b" where it is supposed to give. so its more like memorizing.
My network is not big. I forced it to be small in order to make it unable to memorize. However, it didn't solve my problem.
My network (seq2seq RNN-LSTM Encoder Decoder type) is working good enough on equations which have 2 3 or 4 variable (like a = a , a = a + b , a = a + b + c). what I told you above is just an example smaller version of my problem.
I use Adam learner and CNTK library if it is important.
what do you suggest for me to do to be able to get the correct results?
