NeuralSPARQLMachine--Pytorch-Tranformer-BiLSTM

The architecture of this Neural SPARQL Machine is inspired by the one prensented in this paper :

1. Encode data (Generator)

<Generator input>
      question : What is the ISSF ID of Kim Rhode?
  sparql_query : SELECT DISCTINCT ?answer WHERE { wd:Q233759 wdt:P2730 ?answer}


<Generator output w/o relations learning>
      question : what is the issf id of kim rhode
  sparql_query : select distinct var1 where bkt_open wd_qxxx wdt_pxxx var1 bkt_close
  
<Generator output w/ relations learning>
      question : what is the issf id of kim rhode
  sparql_query : select distinct var1 where bkt_open wd_qxxx wdt_p2730 var1 bkt_close

2. Train & Test seq_to_seq model (Learner)

• LC-QuAD2.0 dataset : Results on test set after 100 epochs of training

Without rel learning	Accuracy	BLEU
Transformer	89.3	81.6
BiLSTM w/ attention	61.5	19.9

With rel learning	Accuracy	BLEU
Transformer	86.6	78.2

3. Decode data & Entity linking (Interpreter)