1. Aramis Rodríguez Blanco
Extraction of Linked Data from unstructured information applying PLN techniques and Ontologies
Authors:
Aramis Rodríguez Blanco
Alfredo Simón Cuevas
Wenny Hojas Mazo
Jose M. Perea Ortega

2. Introduction Rising of smart applications: Semantic search
Popularity of web technologies
Web content reorganized in semantic data
Web focused on documents
Web focused on interconnected data
Linked Data
Search, analysis, knowledge engineering, integration, information retrieval

3. Problematic Linked Data Extracted from:
Semi-structured sources: Wikipedia (DBpedia),
Structured sources: databases,
Unstructured sources: text,
English approaches,
Information extraction limitations,

4. Proposed approach
Extraction of Linked Data from unstructured text in English and Spanish, based on automatic extraction of conceptualization from text as Concept Map, transformed to RDF data model.
Concept Map
Ontology

5. Conceptualization Extraction
Proposed approach phases
Pre-processing (NLP techniques).
Extraction of conceptualization (NLP techniques and ontologies).
Generation of RDF Model (rules based on patterns).
HTML
Text Pre-processing
Plain text extraction
Language recognition
Segmentation
POS Tagging
Morpho-Syntactic Analysis
Conceptualization Extraction
OWL Ontology
Concepts Extraction
Links Extraction
Normalization
RDF triples
RDF Data Model Generation

6. Pre-processing Text extraction from documents (pdf,txt,doc,html).
Text segmentation (by sentences).
Language recognition (English or Spanish).
Syntactic analysis of text (using FreeLing parser).
Token segmentation (separate words).
Morphological syntactic analysis (extract lemma and POS tagging).
Shallow parsing analysis (group tokens in noun groups, adjetive groups, verb groups, etc.).
Deep parsing analysis (dependency tree).

7. Conceptualization extraction
Concept extraction
Text mapping of terms represented as classes or instances in OWL ontology.
Use of predefined patterns which are applied on noun or adjective groups (Patterns were defined from DBpedia ontology).
Patterns
Examples NN
processes
Z NN
Five files
NNP
Ernest Hemingway JJ
impressive
JJ NN
Informatics system
JJ JJ NN
Operative logical processor
NN NNP
Chief Alfred
Z JJ NN
Three efficient methods
Legend: NN: common noun; NNP: proper noun; Z: number or numeral; JJ: adjective

8. Conceptualization extraction
Relations extraction
Identification of explicit relations in dependency tree.
Identification of concept nodes and linking phrase.
Identification of different kinds of relations (mainly verbal and prepositional).
Patterns
Examples
SN0 ‘such as’ {SN1,SN2…,(and|or)} SNn
…animals such as dog and bird.
SN1 ‘is a kind of’ SN0
concept map is a kind of concept graph
SN1 ‘is a’ SN0
dolphins are smart animals
SN {,} ‘has’ {SN,}+{and|or} SN
…humans and monkeys have similar characteristics
Legend: SN: noun group; +: concatenation of elements; |: disjunction of elements; { }: optional elements; ( ): group of elements

9. Conceptualization extraction
Relations extraction
Identification of implicit relations through relations in OWL ontology.
Identification of taxonomic relation between concepts C0 y C1 if both are ontological classes.
Other semantic relations if: C0 and C1 are related by a Object property relation.
Identification of class-instance relations.

10. RDF Model Generation
Every relations is coded as a triplet in RDF data Model.
The source concept is coded as subject and destiny concept is coded as object, and the linking phrase is coded as predicate.
URIs in triplets are built from URLs of processed html file, the URL of the own file, as those from hyperlinks related to terms represented as concepts.

11. Characterization of test collections:
Results and discussion
Evaluation focused on the amount and quality of the information extracted from text (concepts and relations), using or not the ontology.
Abstracts from corpuses DBPedia_ES and DBPedia_EN: and IA:
Characterization of test collections:
Characteristics
DBpedia_ES
DBpedia_EN IA
Language
Spanish
English
Documents 50 6
Word average
75,9
75,2
1111,83
Sentence average
3,0
3,4
46,83

12. Results and discussion
Test method without and with the ontological model.
DBPedia_EN and IA were tested with available ontology.
Evaluation through human evaluators, considering aspects such as:
amount of information extracted,
Contribution level of use of ontology,
Accuracy in concept and relations extracted.
Right concept: nouns or adjectives with important meaning in text, names of entities or phrases with sense.
Right proposition: can be interpreted with own sense (ex. When both concepts are correct and link is well defined).

13. Results and discussion
From information provided by evaluators, accuracy was calculated as the reason between extracted correctly and total.
Aspects
DBpedia_ES
DBpedia_EN IA
NoOnt
YesOnt AC
7,88
10,06
10,2
112,4
120,2 AE
2,2
4,86
5,08
29,2
36,2 AR
6,10
8,38
8,48
103
114,8
ACOnt. -
3,72
37,2
AROnt.
8,6 PC
93,66
89,21
92,58
96,81
98,57 PR
53,92
51,26
53,97
66,67
83,10
Legend: AC: Average amount of concepts; AE: Average amount of entities; AR: Average amount of relations; CCOnt.: Prom. cant. conceptos obtenidosde la ontología; AROnt.: Average amount of relations identified in ontology; PC: Precision in concept identification; PR: Precision in relations identification; NoOnt: Not usig ontology; YesOnt: using ontology.

14. Case of use
To exemplify the method it is selected a text abstract from DBpedia_ES.
Then, it is built the map, and finally the RDF model
La bronquitis crónica es una enfermedad inflamatoria de los bronquios respiratorios asociada con exposición prolongada a irritantes respiratorios no específicos, incluyendo microorganismos y acompañado por hipersecreción de moco y ciertas alteraciones estructurales en el bronquio, tales como fibrosis, descamación celular, hiperplasia de la musculatura lisa, etc.

15. Case of use
Map created automatically from previous sentence:

16. Case of use RDF model obtained automatically by the method:
<rdf:RDF
xmlns:rdf="
xmlns:owl="
xmlns:rdfs="
xmlns:daml=" >
<rdf:Description rdf:about="owl:hiperplasia_de_musculatura_liso">
<rdfs:subClassOf rdf:resource="owl:alteración_estructural"/>
<rdf:type rdf:resource="
</rdf:Description>
<rdf:Description rdf:about="owl:irritante_respiratorio_específico">
<rdf:Description rdf:about="owl:bronquitis_crónico">
<rdfs:subClassOf rdf:resource="owl:enfermedad_inflamatorio_de_bronquio_respiratorio"/>
<rdf:Description rdf:about="owl:microorganismo">
<rdfs:subClassOf rdf:resource="owl:irritante_respiratorio_específico"/>
<rdf:Description rdf:about="owl:fibrosis">
<rdf:Description rdf:about="owl:descamación_celular">
<rdf:Description rdf:about="owl:enfermedad_inflamatorio_de_bronquio_respiratorio">
<rdf:Description rdf:about="owl:alteración_estructural">
</rdf:RDF>

17. Case of use
RDF model obtained automatically by the method:

18. Conclusions
It is presented a method to extract Linked Data from unstructured texts in English and Spanish, and not domain dependent.
It is combined NLP techniques for text processing and information extraction, allowing to obtain more information.
Use of linguistic patterns to identify concepts and semantic relations between them, and dependency parsing, which has let to gain in text coverage.
Concepts extraction, including entities, as the semantics links identified between them enrich Linked Data.

19. Conclusions
Tests show promising results with over 90% in concept identification and over 50% in relations identification.
It is shown the benefits of using an ontological reference model, based on the amount of information extracted, and the improve in accuracy.