Linked Data has emerged as the most widely used and the most powerful knowledge source for Question Answering (QA). Although Question Answering using Linked Data (QALD) lls in many gaps in the traditional QA models, the answers are still presented as factoids. This research introduces an answer presentation model for QALD by employing Natural Language Generation (NLG) to generate natural language descriptions to present an informative answer. The proposed approach employs lexicalization, aggregation, and referring expression generation to build a human-like enriched answer utilizing the triples extracted from the entities mentioned in the question as well as the entities contained in the answer.
Question Answering over Linked Data (QALD) o ers new opportunities to
traditional Question Answering (QA) systems by utilizing the massive Linked Data
cloud as an information source. At its core, QALD transforms the natural
language question to a SPARQL query and then execute it on a Linked Data
resource to retrieve answers. These answers are then presented to the user as
factoid answers without any further enhancements [
The RealText framework1 described in this paper enhances the bare factoid answers by enriching them with more information and presenting them as natural text. An enriched answer is de ned as an answer which provides a description of each of the entities contained in the question as well as in the answer to the question. Therefore, the enriched answer supports and validates the retrieved answer by providing background information more akin to a human generated answer. The RealText framework generates the description by using the triples related to the entity and application of a series of Natural Language Generation (NLG) techniques. In high level overview, these techniques can be categorized into lexicalization, aggregation, and Referring Expression Generation (REG), however each of these categories contain its own set of multiple subtasks to ne tune the nal output. 1 A video demonstration is available at https://vimeo.com/173608898
The rest of the paper presents an overview description of the framework
features. Further details on some of the modules can be found in [
The objective of the demonstration will be to present the complete RealText
work ow from associating lexicalizing patterns to presenting an informative
answer as natural text. The demonstration will use the RealText standalone
application (for a screenshot see Fig. 1).
For the demonstration we use the factoid questions extracted from the QALD-2
test dataset2. Since we work on the answer presentation (last step in QA) the
input data comprised of question, SPARQL query, and as well as the extracted
answer.
The work ow comprises of three main modules; the lexicalization module which
transforms the triples to natural language sentences, Referring Expression
Generation (REG) module which assigns appropriate referring expressions to the
mentions of the main entity, and aggregation module which aggregates
individual sentences to form paragraphs. The nal output contains the paragraphs as
well as the answer in sentence form generated using our answer sentence
generation framework [
Occupational Metonym Patterns utilize the -er nominal based occupational metonyms to derive a prede ned set of lexicalization patterns. For instance, a triple with occupational metonym, director, as the predicate and a movie as a subject. This triple can be lexicalized using a pattern such as hS?, is directed by, O? iL. We have developed a database which contains 33 of such patterns. These patterns are used to lexicalize a triple by matching the predicate and the core ontology class of the subject.
Context Free Grammar (CFG) Patterns uses the language generation capability of CFG and lexicalize the triples with past tense verb as a predicate. To be able to use CFG pattern, the verb (in predicate) should be identi ed as a verb having the frame, NP$VP$NP.
Relational Patterns use the unstructured text to derive patterns. We rst
pre-process text to resolve co-references and then extract relations (harg1, rel,
arg2 iR) using OpenIE [
Property Patterns are prede ned set of patterns which can lexicalize a given
triple with speci c predicate. For example, a pattern such as hS?'s predicate, is,
O? iL will be used to lexicalize triples with predicates, population total, area
total, and postal code. There are ve such patterns de ned with their associated
predicates from DBpedia [
We also carry out a realization phase after applying lexicalization patterns. The realization step corrects the syntactical errors of patterns such as a pattern does not match with the grammatical gender of the triple subject.
Table 1 shows some results from lexicalization modules where each triple is associated with a lexicalization pattern.
Aggregation The aggregation module rst cluster the triples based on the subject. Then within each cluster we sub-cluster the triples based on rules. The triples within sub-clusters are then transformed to the natural language sentences using associated lexicalization patterns. However, at this level we do not substitute the subject expression (S?) of the sentence as it may need a referring expression in the generated paragraphs. Such referring expressions are resolved in the next phase.
Referring Expression Generation The referring expression generation module substitutes the subject expression with appropriate pronouns and entity names to emulate humans. In order to emulate this we change the referring expression after two consecutive usages. hRubens Barrichello, birth place, hS?, was born in, O? iL Sao PauloiT hRubens Barrichello, birth date, hS?, was born on, O? iL 1972-05-22iT hMount Everest, rst ascent per- hS?, was climbed by, O? iL son, Edmund HillaryiT hCaptain America, creator, Joe hS?, was created by, O? iL SimoniT hLyndon B. Johnson, successor, hO?, succeeded, S? iL Hubert HumphreyiT hLondon, population 8308369iT hCanada, largest city, TorontoiT hlargest city in S?, is, O? iL hSocrates, in uenced, hS?, in uenced, O? iL AntisthenesiT hIntel, founded by, Robert hS?, is founded by, O? iL NoyceiT Metonym Metonym Property CFG total, hS?'s population total, is, O? iL Property
This paper described the process of generating natural language descriptions for QALD. The approach is mainly inspired by the NLG where triple content is transformed to natural language paragraphs. In future we expect to extend the framework mainly focusing on the lexicalization pattern mining module. Furthermore, we will be looking at integration of this new approach to Intelligent Personal Assistant (IPA) to provide natural descriptions when presenting answers.