Mateusz Klimaszewski
supervisor: Tomasz Gambin
In recent years, Knowledge Graphs have played an essential role in natural language generation. Primarily the work focused on the question answering and dialogue systems. Neural Machine Translation (NMT), the task of automatic translation between languages, can be viewed as another text generation task due to the same technical schema - sequence-to-sequence architectures and similar evaluation metrics. However, there are significant differences. First, translation models must understand two languages' syntactic and semantic nature instead of one. Also, the degree of freedom and the error margin is much lower - the translation must preserve the meaning and fluency.
Our work evaluates whether it is possible to use Knowledge Graphs to improve NMT. In the experiments, we studied the impact of the shallow representation of a KG, Knowledge Graph Embeddings, on a task which requires in-depth natural language understanding - Word Sense Disambiguation. Our preliminary study demonstrates improvements in out-of-domain datasets in English to German translation.