The basic principle of the Semantic Web carried by the RDF data model is that many RDF statements coexist all together and are universally true. However, some case studies imply contextual relevancy and truth - this is well known in the Conceptual Graph community and handled through the notion of contexts. In this paper, we present an approach and a tool for semantic annotation of textual data using graph contexts. We rely on both Natural Language Processing and Semantic Web technologies and propose a model of RDF contexts inspired by the nested Conceptual Graphs. Sentences are primarily analysed and their grammatical constituents (subject, verb, object) are extracted and mapped to RDF triples. Links between these triples are then established within a semantic scope (i.e., context). The context definition allows us to validate the generated annotations by disambiguating the misleading RDF triples. We show how far our approach is applicable to texts in Engineering Design.
The semantic annotation of texts consists in extracting semantic relations
between domain relevant terms in texts. Several studies address the problem of
capturing complex relations from texts - more complex relations than subsumption
relations between terms identified as domain concepts. They combine statistical
and linguistic analyses. The main applications are in the biomedical domain [
The core issue of the methodology we propose stands in the mapping between
grammatical elements of each sentence in the analysed text and the
corresponding entities in the dedicated-domain ontology. We base upon the MeatAnnot
approach previously designed to support text mining and information retrieval
in the biological domain [
SPARQL is provided with query patterns on named graphs enabling to choose
the RDF dataset against which a query is executed. This is a first step to handle
contextual metadata. A named graph can be used to limit the scope of an RDF
statement to the context in which it is relevant to query it. Furthermore, by
naming contextualized RDF graphs, they can be themselves associated with RDF
metadata, enabling querying on several “levels” of (meta-)annotations. This is
close to the notion of nested graphs in the Conceptual Graphs model [
This paper is organised as follows. We give in section 2 the Natural Language Processing (NLP) technique we use to annotate a given text with RDF triples by relating terms occurring in the text. We introduce in section 3 the Corese design pattern we use to represent and handle nested contexts. We show how we use it to enrich our primary text annotations. We explain how these contextualized annotations provide further information retrieval capabilities when applied to Engineering Design domain. Related work is discussed in section 4. Finally, concluding remarks are provided in section 5. 3 http://www.inria.fr/acacia/soft/corese
Extraction of relations from texts We use the RASP [
S: The L1 luggage compartment contains 100cc.
Hence, we give a simplified RASP syntax tree in Table 1. The sentence S consists of: (1) noun phrase NP, on the left branch of the syntactical tree, which represent the subject subj: determiner and two modifiers; and (2) verbal phrase VP, on the right hand side, constituted of the main verb and the direct object dobj. pppppppppppp8 S hPPPPPPPPPPPPPP NP((det + mod) +mmmmomdm)mmmmmmm6
The L1
O hQQQQQQQQQQQQ
]<<<<<<<<<<<<<<<<< luggage compartment
O ^<<<
<<<<<<<<<<<<<< contain + s 100cc Mapping of grammatical constituents to RDF triples Let us show on the running example the correspondence between a sentence and its translation to an RDF graph triple. Provided that the domain ontology conveys the following knowledge (as it is the case of the ontology we have built for the SevenPro project): a Luggage compartment is part of a Car; a Luggage compartment is related to a Capacity; property contain has for rdfs:domain Car parts (i.e., Luggage compartment, Door, etc.); property contain has for rdfs:range a Capacity unit. We can state that the triple L1, contain, 100cc is a valid instance of property contain and we add it to the text annotation set. The RDF/XML syntax of this statement is given in Table 2 (the spro namespace identifies the SevenPro ontology).
From simple- to complex-sentence semantic annotation We showed above how we generate RDF annotations for simple sentences with grammatical patterns subject, verb, object, hereafter called S − V − O (some possible ambiguity conveyed by the textual material put aside). Here we discuss the handling of more complex sentences and the annotations which we generate. In addition to the S − V − O (sentence in active form) and O − V − S (sentence in passive form) grammatical patterns, we correctly parse and annotate sentences with subordinate phrases when these phrases are “independent” from the main sentence.
However, for other complex sentences, the semantics of the connection between the subordinate and the main sentence is not so simple and cannot be captured in RDF –which is limited to the representation of conjunctive knowledge. It is, for instance, the case of disjunctive sentences where alternative statements co-exist in implicit different contexts. It is also the case when rhetorical relations play a key role in the sentences to be annotated, like the following one including a conditional premise: “If the car C3 has part door D4, then the 100cc are contained in the L1 luggage compartment.”, or this other one containing a causal premise: “The L1 luggage compartment capacity contains 100cc because the car C3 has part door D4.”. In some applications, it constitutes a major problem and may lead to a deadlock issue when querying the RDF graph with SPARQL. Hence, we define the so-called RDF graph context, with recursive capability, in order to tackle the current expressiveness capability lack. 3
Extension of SPARQL to Handle Contextual Relations and Nested Contexts 3.1
The SPARQL query language [
In line 2 of this example, we can state that the pattern graph ?s1 {?x c :
prop ?y} is named as graph ?s1. It can provide a URI to select one graph or
use a variable which will range over the URIs of named graphs in the dataset. A
complementary feature is proposed in [
In RDF/XML syntax, the first triple in line 3 above can be written extensively as: <spro:Tailgate rdf:about="#T2" cos:graph="http://www.sevenpro.org/car/ctx1" > <spro:separate>
<spro:Luggage_compartment rdf:about="#L1"> </spro:separate> </spro:Tailgate>
We use the SPARQL GRAPH primitive to handle RDF named graphs representing different contexts within which alternative metadata can be described. Furtehrmore, we provide an extension of SPARQL to query for contextual relations holding between recursively nested contexts. Once contextual knowledge is represented into RDF named graphs identified by URIs and queried with GRAPH query patterns, these graphs can themselves be described into other separate named graphs. This process of meta-annotating named graphs identifying contexts leads to a recursive nesting of contexts –contexts nested one into another. This is of prime interest for use cases where context graphs are annotated with rhetorical or temporal relations. The unstacking of contexts should make explicit the progress in which nested graphs are involved.
We propose an extension of SPARQL with a REC GRAPH keyword whose grammar rule is similar to the standard SPARQL GRAPH one. The following query enables to retrieve the triples from nested graphs related to a given contextual relation c_Rel. Moreover, all sub-properties of c_Rel –following rdfs:subPropertyOf subsumption relations having c_Rel as value in the RDFS ontology– are matched with the SPARQL query.
In addition, when the property is not specified, e.g., a variable ?p replacing
c_Rel, Corese retrieves the RDF triples having any property (cf. details in [
We have used Corese Graph context capabilities within Sevenpro textual corpus in Engineering Design and the subsequent spro ontology. We show the practical use of the contexts for giving additional metadata with a sentence of the form: If [C1] then [C2], unless [C3]. Then, we show how to improve the SPARQL triple set results with corresponding context-augmented SPARQL queries. We comment the RDF graph context representation, we justify the SPARQL queries, followed by a presentation of the possible RDF triple results. Moreover, in the sentence depicted in Table 3, we show the use of nested contexts. In the second column of Table 3, we describe the corresponding RDF triples for the sentence augmented with RDF graph contexts g1 to g3. The third column describes how these graphs are defined as URI resources (with rdf:Description syntax) and nested within nesting graphs c1 and c2 through the domain relations spro:then and spro:unless. In so doing, we are able to query, with context-augmented SPARQL language using the keyword REC GRAPH. Then, Corese matches the triples in the RDF graph corresponding to triples matching the contextual relations spro:then and spro:unless. We extensively obtain the triples shown in column three of Table 3, (lines 3 to 5 in the result part), alongside with the contextual relations spro:then and spro:unless (first two lines in the result part). We show the context-augmented triple results compared to the mere results which we query with standard SPARQL without contexts.
The named graphs in the sentence of Table 3 are nested as it is shown in Fig. 1. They are organised in the hierarchy of contexts: [c1] : [g1]then[g2]; [c2] : [c1]unless[g3]. Hence, we can relate the RDF triple “a p b” to “c q d” by traversing the hierarchy of Fig. 1. In so doing, the semantics of the example sentence is fully captured with annotation capability of nested graph contexts. g1 a p b then c1
unless g2 c q d c2 g3 The mechanism introduced by RDF graph contexts is powerful enough to represent a variety of NL expressions. First, with the RDF context expressiveness, we can represent the logical disjunction or, the negation not as RDF graph contexts. Moreover, we can describe the modal primitives can, may, as in: The headliner may be projected beyond the vertical of the external surface. There are a number of other relations which we can model: temporal (i.e., after, meanwhile, etc), spatial (i.e., below, behind, etc.), comparative (i.e., more... than, etc.). Presently, we fail to model the correct annotations of sentences having an ambiguous subject/object constituents. Moreover, a variant in the example sentence raises the still-open problem of anaphora resolution in NLP. The inlet headliner H1 should be lifted by metal bar B2 [. . . ] unless it is in position P5; where the pronoun it represents H1.
In the Semantic Web domain, the work of [
The objective of this paper is twofold: (i) to show how we generate accurate RDF triples from texts using NLP techniques, and (ii) to augment the semantic annotation generation with RDF graph context metadata in order to catch the semantics of the analysed texts, and consequently to enhance the retrieval capabilities. Linguistic analysis is used to suggest appropriate annotations to the text. The text analysis process strongly depends on the background knowledge (i.e. ontologies, terminology, etc.) of the analysed domain. The more precise ontologies and related terminology - list of domain terms, e.g. car manufacturer names, etc. -, the more significant the extracted annotations are. We have started to generate RDF annotation triples from simple (S − V − O) sentences. Then, a number of features were designed to generate more complex annotations, e.g., sentences containing subordinate phrases. Based upon the context graph capability, we have shown new capabilities of high usefulness in the query of the graph by using named graphs and nested contexts. The RDF graph context paradigm can be used recursively. Hence, the text annotation allows us to produce the accurate corresponding semantic annotation. Finally, our approach is domain independent. The analysis process remain the same provided that ontologies have been adapted according to the text domain.
In the future, we aim at developing more complex sentence analysis
following the rhetorical relations studied in RST [