Linked Open Data (LOD) is growing rapidly as a source of structured knowledge used in a variety of text processing applications. However, the applications using the LOD need to be able to mediate between the front end user interfaces and LOD. This often requires a natural language interpretation of this structured, linked data. We demonstrate a middle-tier framework that can generate patterns which can be used to transform LOD triples back into natural text. The framework utilizes preprocessed free text to extract a wide range of relations which are then aligned with triples to identify possible lexicalization patterns. These lexicalization patterns can then be used to transform a given triple into natural language sentence.
The goal of the demonstration will be to show the whole of the RealTextlex
work ow from the extraction and preprocessing of unstructured text to the
generation of lexicalization patterns. The demonstration will utilize a Java client
application built on top of the RealTextlex framework. In essence, the
demonstration will focus on how the lexicalization patterns can be extracted from the
unstructured text and the techniques of utilizing extracted patterns to transform
a given Linked Data triple into natural text.
For the purpose of this demonstration we focus on triples belonging to ve
randomly selected DBpedia ontology classes; Bridge, Actor, Publisher, River,
and Radio Host. These ve classes o er 132 unique predicates which can be
lexicalized. A detailed description on the datasets used to evaluate the framework
can be found in [
In the next phase, the module extract text related to the selected DBpedia RDF le (e.g., Wikipedia page for Brooklyn Bridge). The main source of free RealTextlex: A Lexicalization Framework text extraction are the Wikipedia pages for selected DBpedia RDF le. However, Wikipedia on its own is not su cient to create a rich enough natural language representation of RDF triples. In order to enrich the possible patterns, a web search module (based on Bing API2) was integrated to extract more text snippets related to the RDF le from websites. The extracted text was cleaned using the shallow text feature based boilerplate removal algorithm3 and the co-references were resolved using the Stanford CoreNLP.
The resulting sentence collection from above step was analysed to determine the candidate sentences that have RDF triple elements corresponding to subject, predicate, and object. The candidate sentences were then assigned to triples with a score proportional level of presence of the individual triple elements. Relation extraction and alignment In the next step we extracted the relations (arg1$ rel $ arg2) from the generated candidate sentences from the previous step. For this we employed the self-supervised Open Information Extraction (Open IE) module4 which enables us to identify a large number of relations and to associate them with triples. Our hypothesis is that a relation can be aligned with a triple structure to identify potential lexicalization of the triple under consideration. We aligned each relation with the triple and assigned a score based on the level of alignment. The aligned relations were then processed to extract the patterns which were used to lexicalize the triple as natural text.
Ensemble pattern processing and combination In this phase, we rst transformed the aligned relations to patterns by substituting subject and objects literal values with generic expressions (e.g., h arg1: Brooklyn Bridge, rel: is designed by, arg2: John Roebling i) s? is designed by o? ). The success of Open IE based pattern extraction is based on the availability of the textual representation of the required triple. Therefore, we consider alternatives to generating patterns which work with the main module. For this we utilized the verb frames from lexical semantic resources, mainly from WordNet and VerbNet. The pattern processing module generates patterns that can be extracted from relations followed by the attempt to lookup the predicate in the verb frame database5. If the module is able to nd a pattern of NP$VP$NP structure for the given verb, then a predetermined pattern of [subject]$VP$[object] is created. A predetermined set of properties of entities (e.g., eye colour, hair colour) are lexicalized with a pre-processed set of patterns.
Pattern enrichment and persistence Some features of triples can result in patterns that cannot be generalized to other triples with the same predicate. We
3 https://code.google.com/p/boilerpipe/
4 http://knowitall.github.io/ollie/
5 An embedded database which contains verb frames from WordNet and VerbNet.
have identi ed three such features in our investigations; grammatical gender,
ne ontology class, and object multiplicity. A detailed explanation of how these
features a ect patterns to be in a di erent form is given in [
Recent frameworks like Lemon[
The approach for LOD lexicalization presented in this paper o ers new insights
into generating lexicalization patterns for Linked Open Data cloud by applying
Open IE on automatically collected and processed free text. The lexicalization
module described here is part of a parent Natural Language Generation project
[