Knowledge graph embedding for improving Persian question answering

This paper explores the task of answering Persian questions using embeddings derived from a Persian knowledge graph called Fars-Wiki-KG (Shirmardi et al., 2021). This graph contains more than 2 million nodes and 7.5 million triples, covering over 90% of the Persian Wikipedia infoboxes. Based on the nodes and edges of this knowledge graph, 1370 patterns have been created to build a question answering dataset. These patterns include both simple and multi-hop questions. For example, a simple pattern might be: "Which country is the producer of X?" In contrast, a multi-hop pattern could be: "What language do the people speak in the country of the producer of X?" The first one is straightforward, while the second question involves two hops: first finding the country of X, and then determining the language associated with that country. By considering these patterns, and extracting their corresponding answer from Fars-Wiki-KG, we provided a dataset consisting of questions and answers. Having a question like "Which country is the producer of 'Breaking Bad'?", the head entity is 'Breaking Bad,' the relation is the country of origin, and the tail entity is the 'USA' which is the answer.

Various knowledge graph embedding models, such as translational and semantic matching models, are trained and evaluated on subsets of the knowledge graph to embed its nodes. This process captures both explicit and latent information between the nodes in numerical vector form. At the same time, Persian questions are embedded using a multilingual version of the BERT language model, namely XLM-RoBERTa (Liu et al., 2019), to represent the relationships between each question and its corresponding answer. Each question starts with a head entity, followed by a relationship that describes how to reach the tail, which is the answer to the question. By utilizing the head entity embedding and the relation embedding (which contains information about the question), a score is assigned to all other nodes in the graph. This score indicates the likelihood of each node being the tail entity (question's possible answers). Therefore, the above-explained approach, inspired from the research by (Saxena et al., 2020), allows the system to answer questions even when the relevant triples are not explicitly included in the graph. Moreover, it enables responses to more complex questions that require traversing multiple edges within the graph called multi-hop questions.