Knowledge Graph Construction

A Knowledge Graph (KG) stores knowledge in the form of a graph, in which a node represents an entity and an edge denotes the relationship between entities. Well-known examples of KGs are DBpedia , Freebase and YAGO.

Knowledge Graph construction is the process of generating a knowledge graph from natural language text. The process is often comprised of two distinct steps:

1. Knowledge extraction where the text is converted to tuples
2. Knowledge mapping where the extracted entities are mapped to existing entities in the knowledge graph.

Recent papers on this topic:

1. KERTKEIDKACHORN, Natthawut, and ICHISE, Ryutaro . "An Automatic Knowledge Graph Creation Framework from Natural Language Text." IEICE TRANSACTIONS on Information and Systems 101.1 (2018): 90-98.‏
2. Kríž, Vincent, et al. "Data extraction using NLP techniques and its transformation to linked data." Mexican International Conference on Artificial Intelligence. Springer, Cham, 2014.‏
3. Carlson, A., Betteridge, J., & Kisiel, B. (2010). Toward an Architecture for Never-Ending Language Learning. In Proceedings of the Conference on Artificial Intelligence (AAAI) (2010), 1306–1313. https://doi.org/10.1002/ajp.20927

Research questions

Question Answering Systems (QAS) are a universal way for humans to obtain knowledge from the web. The users ask questions and each question could lead to another set of related questions. Hence it becomes a dialog between systems and users. For example, using chat bots, personal assistants or question answering websites (e.g. Google Assistant, Amazon's Alexa, Microsoft's Cortana).

Our primary objective is that a user will be able to get answers to a sequence of related questions. To do so, we will have to answer the following:

1. How to spot a related question in case of a semantic sequence of questions?
2. How to retrieve the missing part of this question?
3. How to generate a query for the new question using the discovered context and retrieve the answer from our knowledge graph?

Within that task, we are going to use a deep learning sequence model to establish two main points:

1. Predict whether this question is a factoid or Related Question.
2. Predict which components of the past questions are still relevant to the tested question.