A Deep Natural Language Inference Predictor Without Language-Specific Training Data

In this paper we present a technique of NLP to tackle the problem inference relation (NLI) between pairs sentences in target language choice without language-specific training dataset. We exploit generic translation dataset, manually translated, along with two instances same pre-trained model - first generate sentence embeddings for source language, and second fine-tuned over mimic first. This is known as Knowledge Distillation. The has been evaluated machine translated Stanford NLI test Multi-Genre RTE3-ITA also proposed architecture different tasks empirically demonstrate generality task. native Italian ABSITA on Sentiment Analysis, Aspect-Based Topic Recognition. emphasise exploitability Distillation that outperforms other methodologies based translation, even though former was not directly trained data it tested over.