Getting gibberish predictions when using recurrent LSTM and arrays of strings as output training data

[chrisvel]

I have a bunch of sentences and I want to generate tags for each one. I have my data which are simple:

[{
 input: "Buy tickets for opera",
 output: ['errands', 'orders']
},
{
 input: "Clean garage",
 output: ['errands', 'home']
},
....

I am training the model simply with :

const network = new brain.recurrent.LSTM();

  network.train(trainingData, {
    (error) => console.log(error),
    iterations: 1000,
  });

When I run a:

network.run('Some random text');

sometimes I get an array with a correct tag, but other times it returns gibberish with random characters or the tags joined together in a string, for example the sentence "Service my XBox dvd drive" returns this output:

["sco comhermer fililys.AAouto afamily.shopping"] 

I read somewhere that LSTM cannot classify so I am ok with this but what do you suggest?

Will something like making a matching table with numbers and tag words work? Something like:

1: orders
2: errands
3: family 
4: personal 
.....

and then feeding the output of my training data with numbers ?

Is it something not expected to work as I am hoping to or is it a bug?

[nilooy]

same for me, i'm passing strings in both , input and output object, but when run it, it gives mixed/gibrish output.

[purplnay]

The gibberish is probably due to 2 reasons, the model learns to read and write character by character an not word by word, the second I would say is that 1,000 iterations seems quite low for training.
For now I would recommend splitting your input into words, with a space after each, and same for your output, add a space so they don't collapse like "errandsorders".

I've filled an issue for a feature request here #871 since I think this is a recurrent issue that is not so well documented.