WiseKB is a RDF knowledge base where SPARQL queries are supposed to access. Those SPARQL queries are also to be translated from natural language queries.
WiseKB has been designed and developed for a Korean national project called “Exobrain” (2013-2023) with its pairing QA system called “WiseQA”. But WiseKB is open that could be used for other QA systems. WiseKB is basically ontology-based but not strongly constrained only to allow ontology-based triplet, e.g., for open IE. WiseKB is not limited to ontology-based KB but also accommodates a style of Open IE. WiseKB is based on the already usable linked open data (e.g., DBpedia) as well as the triplets extracted from Wikipedia with other web data, in this moment from Korean-language text and LOD.
In this talk, a NL query set called “NLQ400” is analyzed to test the WiseKB whether it is answerable against DBpedia and Wikipedia in its Korean version mainly. Three types of NL queries will be proposed depending on the locations of the supposed answers for each NL query. Then we will explain its architecture, strategy, and algorithms to solve those types of NLQ400. NLQ400 would be evolved to be a shared task in OKBQA and in other opportunity. WiseKB construction modules are also to be coined under OKBQA deployment.
LODQA is an open source project to implement a natural language interface to SPARQL endpoint. LODQA has an highly modularized architecture. For an input query in natural language, first, the graphication module of the system produces a corresponding graph representation. It then asks PubDictionaries (http://pubdictionaries.org) for URIs of elements in the graph representation. Finally, it calls GraphFinder to search a SPARQL endpoint for the graph representation considering possible variations. In the talk, each module of LODQA system is explained with demonstration. It also aims at introducing other relevant projects in comparison with the corresponding LODQA modules.
The starting point of this talk will be QALD, a series of evaluation campaigns on multilingual question answering over linked data. QALD provides a bechmark of around 400 questions over DBpedia, that are annotated with SPARQL queries and answers. We will look at some of these questions in order to highlight the main challenges involved in mapping natural language questions to SPARQL queries. After that, we will introduce the two-step, template-based approach to question answering that is at the core of OKBQA. The main idea behind it is to first build a dataset-independent representation of the semantic structure of the natural language question, a template, that can then in a later step be instantiated by means of dataset-specific entity identification and predicate matching methods. We will focus on the process of generating templates, and will show how this allows us to meet some of the challenges in the QALD benchmark.
The ability to interpret and approximate the meaning of words or expressions is a central element for building semantic and intelligent systems. However, building effective semantic approximation and interpretation depends on techniques which support the representation and processing of commonsense and domain knowledge at scale. In this talk we will describe techniques for supporting robust semantic approximation at scale in the context of Question Answering systems. We will start analysing the user-data semantic gap, following to the description of which principles, state-of-the-art techniques and tools have shown to be effective in bridging the semantic gap.
Knowledge extraction approaches play a central role in the question answering process. In particular, they are of central importance during the conversion of natural-language questions to SPARQL queries. In this talk, we will begin by giving a general overview of knowledge extraction from text. Thereafter, we will present current state-of-the-art approaches for entity recognition, relation extraction and entity linking. In particular, we will focus on the FOX (http://fox.aksw.org) for entity recognition and relation extraction, AGDISTIS (http://agdistis.aksw.org) for entity linking and GERBIL (http://gerbil.aksw.org) for evaluation. Finally, we will give a short demo of these tools and present how they are/will be used in OKBQA.
"How does the brain represent conceptual knowledge? and, How we can learn this from PubMed" (Kyoung-Whan Choe)
Many computing theories and algorithms are inspired by brain mechanisms. Especially how the brain stores and retrieves knowledge will be of particular interest to the developers of KB & QA technology. In this talk, I will introduce functional magnetic resonance imaging (fMRI), a widely-used neuroimaging technique to measure brain activity, and how fMRI can be used to investigate the neural representation of conceptual knowledge. Also to provide an end-user perspective of KB & QA system, I will share my experience with PubMed, a medical knowledge base, in exploring scientific literature.