I'm Trying to Teach a Computer to Translate from English to Czech

I found a sample of about 7000 sentences that are in both languages. I’ll tell it to go through them 100 times and try to understand the words. I’ll make it easier for it by providing a Czech-Czech dictionary with 2,000,000 words from Wikipedia (word embedding 300d).

I create a model and start the training. The desktop indicates that it’s working – the fan is whirring. The temperature on the GPU rises from 36 to 40, 50, 52 degrees Celsius. My computer athlete is audible, but it’s only operating at 33% of its maximum. A warm-up. Learning English is, after all, more of a warm-up for it.

I calculate that learning the entire dataset and going through it 100 times will take about a quarter of an hour. During this time, I realise that it would probably benefit much more from an English-English dictionary when translating from English. Too late, it has already gone through it 77 times. Even so, the little rascal has learned the training data with 99% accuracy and is passing the test set with a still impressive result of 68% correct. Wow! I wish I had such a knack for languages. I test it with a few examples – see the image. The first sentences are nothing special, but then it really gets going! Considering it has only been learning English for 15 minutes, that’s quite decent.

I’ll try giving it the English-English dictionary with 100 dimensions as well. The result? The accuracy on the test data is about 1% better. It handles 3500 English and 9500 Czech word forms reasonably well. What do you think?