Question answering systems provide easy access to structured data, in particular RDF data. However, the user experience is often negatively a ected by questions that are not interpreted correctly. To remedy this, we present a new guided approach to QA that ensures that all questions that can be entered into the system also return a corresponding answer. For this, a template-based approach is used to generate all possible questions from a given RDF dataset using a number of templates. The question/answer pairs can then be indexed to provide autocompletion functionality at querying time. We describe the architecture and approach and present preliminary evaluation results.
QALD
Knowledge Graph
Question answering systems over linked data (QALD) often su er from their brittleness, that is the relatively high probability of not being able to parse and execute a question correctly. This is problematic as it negatively a ects the user experience and trust in a system. We hypothesize that user experience would be signi cantly increased if all questions that can be entered into a query interface would also produce an answer. Building on this hypothesis, we have developed a QALD system that provides a controlled interface such that every question entered is interpretable and returns and appropriate answer. This is accomplished by a template-based approach that matches basic graph patterns over the data and, using a lexicon, generates di erent variants of asking the same question. The question and corresponding answer are stored in an index and used to provide query writing support using auto-completion to propose possible continuations of a query. In this paper we present our approach, which has been implemented in Python and relies on Hbase as database for indexing question/answer pairs. We illustrate the workings of the system using DBpedia as knowledge graph. A live demo of the system will be shown during the demonstration session, which is available at:
Our overall approach is visualized in gure 1. In an o ine process, a set of prede ned templates are used to generate natural language questions and SPARQL queries automatically. The SPARQL queries are evaluated and the resulting question/answer pairs are stored and indexed in HBase3. On the basis of this index, questions are retrieved while the user is typing a question and possible completions are proposed in an auto-completion functionality.
The retrieval of the data and the generation of the questions is displayed in more detail in gure 2. The templates currently supported by the system are given in the table below: Template Example question Basic Triple (Noun/Verb) Who is the wife of Barack Obama? Verb with Prepositional Phrase Which team Walter Payton played for? Participle Construction Basic In which programming language is gimp written? Numeric Question What is the frequency of BBC Radio She eld? MaxCount Who has the most alma maters? Numeric Aggregation Which Educational Institution has the lowest faculty size? Numeric Filter Give me all ships where the draft is less than <number>? Ordering Give me all american football players ordered by birth place? Geographic Give me all west german movies?
In order to illustrate the behaviour of the system, we focus on the Basic Triple (Noun/Verb) and the Verb with Prepositional Phrase pattern. We will assume that for the property dbo:spouse the following lexical variants are in the lexicon: spouse, marry, wife, husband. In these settings, the template-based question generation would generate the following questions for the property dbo:spouse:
The variance in the verbalizations for each template depends on the quality and coverage of the lemon lexicon. To extend our approach a template with the 3 https://hbase.apache.org/
Title Suppressed Due to Excessive Length Who is the spouse of Barack Obama? Who is the wife of Barack Obama? spouse of Barack Obama? husband of Barack Obama?
Who is married to Barack Obama?
Who is the husband Barack Obama?
married to Barack Obama?
wife Barack Obama?
corresponding graph pattern has to be de ned. In the demo session we will show
how straight forward it is to do so.
For the evaluation the training data of QALD-5[
In addition to the automatic evaluation, we performed a manual evaluation in which we were able to identify types of questions that can not be answered given the current system. In particular, this analysis revealed that our system can not answer questions involving two conditions, such as Give me all actors who were born in Paris after 1950. One feasible solution for this example would be to generate a template which combines two kinds of questions, calculating the intersection between the answers, and using this data to generate new possible questions.
In this paper we have proposed a novel approach to guided QA that relies on a
template-based approach to generate pairs of questions / answers o ine and
indexes them to support real-time auto-completion. Our system is very similar to
a system presented earlier by Rico et al. [