Mateusz Klimaszewski
supervisor: Tomasz Gambin
The recent rise of deep learning changed the Machine Translation field, introducing the next-gen approach - Neural Machine Translation (NMT). The NMT methods require large-scale datasets with parallel sentences in source and target languages. However, the NMT quality suffers when the amount of data is limited or the translation is out-of-domain.
Following the successful work in question answering and text generation fields, we aim to include Knowledge Graphs (KG) to improve the mentioned shortcomings of NMT. Knowledge Graphs are specific knowledge bases intended to extract and structure human knowledge. Formed as a graph, Knowledge Graphs allow AI systems to perform complex reasoning leveraging organised data.
In our experiments, we studied the impact of the shallow representation of a KG, Knowledge Graph Embeddings. To verify whether the data in KG is robust enough for complex reasoning, we extended the previous research with out-of-domain evaluation and expansion beyond named entities. Our preliminary study demonstrates improvements in English to German translation on out-of-domain datasets.