Semantic Quran

Abstract

Abstract

In this paper we describe the Semantic Quran dataset, a multilingual RDF representation of translations of the Quran. The dataset was created by integrating data from two different semi-structured sources and aligned to an ontology designed to represent multilingual data from sources with a hierarchical structure. The resulting RDF data encompasses 43 different languages which belong to the most under-represented languages in the Linked Data Cloud, including Arabic, Amharic and Amazigh. We designed the dataset to be easily usable in natural-language processing applications with the goal of facilitating the development of knowledge extraction tools for these languages. In particular, the Semantic Quran is compatible with the Natural-Language Interchange Format and contains explicit morpho-syntactic information on the utilized terms. We present the ontology devised for structuring the data. We also provide the transformation rules implemented in our extraction framework. Finally, we detail the link creation process as well as possible usage scenarios for the Semantic Quran dataset.

1 Introduction

Over the last years, the Linked Open Data (LOD) movement has gained significant momentum [1]. A large number of datasets was extracted from sources as different as Wikipedia infoboxes and curated bio-medical databases. Still, most of the datasets in the Linked Data Cloud contain only English labels and fail to represent the diversity of languages used across the Web.1

¹

From the 315 datasets analyzed by the LodStats framework (http://stats.lod2.eu), 128 datasets provide English labels. French, the second most popular language in the LOD Cloud, is used in only 15 (approximately 4.8%) of the datasets. Most other languages occur in at most one dataset.

Yet, a more multilingual Linked Data Cloud would represent a tremendous resource that can be used for novel knowledge extraction techniques and more broadly for novel natural-language processing (NLP) approaches. For example, novel NLP approaches for minority languages could be developed by reusing information available across the different languages [6]. Moreover, a structured representation of corpora would improve their use in applications such as the specification of templates for question answering [7] or the efficient merging with other resources [3].

In this paper, we present the Semantic Quran dataset. The Semantic Quran dataset consists of all chapters of the Quran in 43 different languages including rare languages such as Divehi, Amazigh and Amharic. The data included in our dataset was extracted from two semi-structured sources: the Tanzil project and the Quranic Arabic Corpus (cf. Section 4). We designed an ontology for representing this multilingual data and their position in the Quran (i.e., numbered chapters and verses). In addition to providing aligned translations for each verse, we provide morpho-syntactic information on each of the original Arabic terms utilized across the dataset. Moreover, we linked the dataset to three versions of Wiktionary as well as DBpedia and ensured therewith that our dataset abides by all Linked Data principles.2 ²

http://www.w3.org/DesignIssues/LinkedData.html

In the following, we present the data sources that we used for the extraction (Section 2). Thereafter, we give an overview of the ontology that underlies our dataset (Section 3). Section 4 depicts the extraction process that led to the population of our ontology. We present our approach to interlinking the Semantic Quran and Wiktionary in Section 5. Finally, we present several usage scenarios for the dataset at hand (Section 6).

2 Data sources

Two web resources were used as raw data sources for our dataset. The first web resource is the data generated by the Tanzil Project,3

³

http://tanzil.net/

which consists of the original verses in Arabic as well as 42 manual translations of the entire book. Our second web resource, the Quranic Arabic Corpus,4 ⁴

http://corpus.quran.com

was used to obtain morpho-syntactic information on each of the words contained in the Arabic version of the Quran.

2.1 Tanzil project

The Tanzil Project5

⁵

http://tanzil.net/wiki/Tanzil_Project

was motivated by inconsistencies across the different digital versions of the Quran. These were mainly due to missing/incorrect diacritics, Arabic text conversion problems, and missing encoding for some Arabic characters.

Tanzil was launched in early 2007 with the aim of producing a curated unicode version of the Arabic Quran text that can serve as a reliable standard text source on the web. To achieve this goal, then Tanzil team developed a three-step data quality assurance pipeline which consists of (1) an automatic text extraction of the Arabic text, (2) a rule-based verification of the extraction results and (3) a final manual verification by a group of experts.

The result of this process was a set of datasets that were made available in several versions and formats.6 ⁶

For more details on available formats and datasets, please see http://tanzil.net/download/.

In addition to the original Arabic sources, Tanzil provides sentence-parallel translations of the Quran in 42 different languages by different translators.7 ⁷

http://tanzil.net/trans/

We manually selected one translation per language for the extraction process.8 ⁸

The list of translations used can be found at http://goo.gl/s5RuI.

Note that all Tanzil datasets are distributed under the terms of Creative Commons Attribution 3.0 License.9 ⁹

http://creativecommons.org/licenses/by/3.0/

2.2 The Quranic Arabic Corpus project

The Quranic Arabic Corpus is an open-source project, which provides Arabic annotated linguistic resources which shows the Arabic grammar, syntax and morphology for each word in the Quran. This is a valuable resources for the development of NLP tools for the Arabic language, in which a single word can encompass the semantics of entire English sentences. For instance the Arabic word “faja’alnāhum” can be translated into the entire English sentence “and we made them”. The compact syntax of Arabic leads to that a single word being separable into distinct morphological segments. For example, “faja’alnāhum” can be subdivided into:

fa – a prefixed conjunction (engl. “and”),

ja’al – the stem, a perfect past tense verb (engl. “made”) inflected as first person masculine plural,

nā – a suffixed subject pronoun (engl. “we”) and

hum – a suffixed object pronoun (engl. “them”).

A Resource Description Framework (RDF) and Natural Language Processing Interchange Format (NIF) [4] representation of this rich morphology promises to further the development of integrated NLP pipelines for processing Arabic. In addition, given that this corpus was curated manually by experts, it promises to improve the evaluation of integrated NLP frameworks. We thus decided to integrate this data with the translation data available in the Tanzil datasets. Here, we used the Quranic Arabic Corpus Version 0.410

¹⁰

http://corpus.quran.com/download/

in its delimited text file version under the “GNU General Public License”.11 ¹¹

http://www.gnu.org/licenses/gpl.html

3 Ontology

To represent the data as RDF, we developed a general-purpose linguistic vocabulary. The vocabulary12

¹²

http://mlode.nlp2rdf.org/datasets/qvoc.owl.ttl

was specified with the aim of supporting datasets which display a hierarchical structure. It includes four basic classes: Chapter, Verse, Word and LexicalItem.

The Chapter class provides the name of chapters in different languages and localization data such as chapter index and order. Additionally, the chapter class provides metadata such as the number of verses in a chapter and provenance information. Finally, the chapter class provides properties that allow referencing the verses it contains. For example each chapter provides a dcterms:tableOfContents for each of its verses in the form qrn:quran<chapter>-<verse>.

The Verse class contains the verse text in different languages as well as numerous localization data such as verse index and related chapter index. Additionally, this class provides related verse data such as different verse descriptions and provenance information. Finally, it contains referencing properties similar to those of chapters.

The Word class encompasses the next level of granularity and contains the words in the verse text in different languages as well as numerous localization data such as related verse and chapter verse indexes. Additionally, the word class provides word provenance information and some referencing properties.

Currently, the LexicalItem class provides morphological data on the Arabic words only. Several ontologies can be used to represent such information. In our dataset, we relied on the RDF representation of the GOLD linguistic ontology13 ¹³

http://linguistics-ontology.org/

 [2] to provide linguistic properties of lexical items such as acoustic, root, part of speech, gender, number, and person. We chose to use GOLD in contrast to other ontologies because it belongs to the most exhaustive ontologies for modeling linguistic properties. Thus, it will allow us to easily extend this dataset in future work. All the objects of the previously mentioned properties are URIs from the OLIA Arabic Linguistic ontology.14 ¹⁴

http://nachhalt.sfb632.uni-potsdam.de/owl/

Analogously to the other classes, LexicalItem provides provenance information and referencing properties. A UML class diagram of the four basic ontology classes of the Semantic Quran Dataset with inter-class internal relations is shown in Fig. 1.

Fig. 1.

UML class diagram of the Semantic Quran Ontology.

4 Extraction process

The original Tanzil Arabic Quran data and translations are published in various formats. For the sake of effectiveness, delimited text files were selected as the basis for the RDF extraction. The format of the delimited file is chapterIndex|verse|verseText. For example, the first verse of the first chapter of the English translation of the Quran is 1|1|In the Name of Allah, the Most Beneficent, the Most Merciful. On the other hand, the Quranic Arabic corpus is available as tab-separated text file of the form ”LOCATION FORM TAG FEATURES“:

• The LOCATION field consists of 4-part numbering scheme of the form (Chapter : Verse : Word : Segment). For example, the first segment of the first word of the first verse of the first chapter has the form (1:1:1:1).

• The FORM field contains the text of the current segment in the Extended Buckwalter transliteration.15

  ¹⁵

  The Buckwalter transliteration uses ASCII characters to represent the orthography of the Arabic language. For the conversion table, see http://www.qamus.org/transliteration.htm.

  For example the corresponding FORM to (1:1:1:1) is bi (engl. “In”).

  The TAG field contains the part-of-speech tag for the current segment. For example the corresponding TAG to (1:1:1:1) is p which stands for preposition.

  The FEATURES field contains a complete morphological analysis of the current segment such as root, case and person-number-gender properties. For example the corresponding FEATURES to (1:1:1:1) is PREFIX|bi+ which stands for preposition prefix (“by”, “with”, “in”) with acoustic property “bi”.

  Given the regular syntax used in the text file corpus at hand, we were able to carry out a one-to-one mapping of each fragment of the input text file to resources, properties or data types as explicated in the ontology shown in Fig. 1. We relied on the Apache Jena Framework16 ¹⁶

  http://jena.apache.org/

  for the conversion. The part-of-speech information and morphological characteristics of each segment of the Arabic Quranic Corpus were extracted and integrated with the words found in the Tanzil dataset. The merged data is now available in the RDF format. In order to simplify the interoperability of the generated dataset, we followed the specifications of the NIF. Currently, the original Arabic and four different translations of the Quran (Arabic, English, German, French and Russian) abide by the NIF formalization. Details of the Semantic Quran dataset CKAN entry, its SPARQL endpoint, version and license are listed in Table 1.

  Table 1

  Technical details of the Quran RDF dataset

  Name	SemanticQuran
  Example Resource	http://mlode.nlp2rdf.org/resource/semanticquran/quran1-1
  Dataset dump	http://mlode.nlp2rdf.org/datasets/semanticquran.nt.gz
  Sparql Endpoint	http://mlode.nlp2rdf.org/sparql
  Dataset graph	http://thedatahub.org/dataset/semanticquran
  Ontology	http://mlode.nlp2rdf.org/datasets/qvoc.owl.ttl
  Ver. Date	29.11.2012
  Ver. No	1.0
  Licence	Attribution-NonCommercial-ShareAlike 3.0 Unported (CC BY-NC-SA 3.0)
  CKAN	SemanticQuran

  5 Linking

  We aimed to link our dataset with as many data sources as possible to ensure maximal reusability and integrability in existing platforms. We have generated links to 3 versions of the RDF representation of Wiktionary as well as to DBpedia. All links were generated by using the LIMES framework [5]. The link specification used was essentially governed by fragments similar to that shown in Listing 1. The basic intuition behind this specification is to link words that are in a given language in our dataset to words in the same language with exactly the same label. We provide 7617 links to the English version of DBpedia, which in turn is linked to non-English versions of DBpedia. In addition, we generated 7809 links to the English, 9856 to the French and 1453 to the German Wiktionary. Links to further versions of DBpedia and Wiktionary will be added in the future.

  
  Listing 1.

  Fragment of the link specification to the English Wiktionary.

  We evaluated the quality of the links generated by manually checking 100 randomly selected links from each of the three languages. The manual check was carried out by the two authors. A link was set to be correct if both authors agreed on it being correct. Overall, the linking achieve a precision of 100% for the English version, 96% for the French and 87% for the German. The error in the French links were due homonymy errors. For example, “Est” (engl. East) was linked to “est” (engl. to be) in some cases. Similarly in the German, “Stütze” (engl. support) was linked to “stütze” (engl. imperative singular form the verb “to support”). In the next version of the dataset, we will add context-based disambiguation techniques to improve the quality of the links. Especially, we will consider the type of the expression to link while carrying out the linking to ensure that verbs cannot be matched with nouns for example. Still, the accuracies we achieve in these three languages are sufficient to make the dataset useful for NLP applications. The recall could not be computed manually. While these values are satisfactory, they can be improved further by devising a disambiguation scheme based on the context within which the words occured. To achieve this goal, we aim to combine the results of LIMES with the AGDISTIS disambiguation framework17 ¹⁷

  http://github.com/AKSW/AGDISTIS

in future work.

6 Use cases

The availability of a multilingual parallel corpus in RDF promises to facilitate a large number of NLP applications. In this section, we outline selected application scenarios and use cases for our dataset.

Listing 2.

Verses that contains moses in (i) Arabic (ii) English and (iii) German.

Data retrieval

The Quran contains a significant number of instances of places, people and events. Thus, multilingual sentences concerning such information can be easily retrieved from our dataset, for example for the purpose of training NLP tools. Moreover, the aligned multilingual representation allows searching for the same entity across different languages. For example, Listing 2 shows a SPARQL query which allows retrieving Arabic, English and German translations of verses which contain “Moses”.

Arabic linguistics

The RDF representation of Arabic morphology and syntax promises to facilitate the retrieval of relevant sub-corpora for researchers in linguistics. For example, Listing 3 provides an example of a SPARQL query witch retrieves all Arabic prepositions as well as an example statement for each of them.

Listing 3.

List all the Arabic prepositions with example statement for each.

Another example is provided by Listing 4, which shows a list of different part-of-speech variations of one Arabic root of the word read “ktb” (engl. “write”); note that in this example we use the Arabic root “ktb” witten in The Buckwalter transliteration.

Listing 4.

List of different part of speech variations of one Arabic root of the word read “ktb”.

Interoperability using NIF

Using the interoperability capabilities provided by NIF, it is easy to query all occurrences of a certain text segment without using the verse, chapter, word, or lexical item indexes. For instance, Listing 5 lists all the occurrences of “Moses” with no need to have an extra index.

Information aggregation

The interlinking of the Quran dataset with other RDF data sources provides a considerable amount of added value to the dataset. For example, the interlinking with Wiktionary can be used as in Listing 6 to get the different senses for each of the English words contained in the first verse of the first chapter “qrn:quran1-1”.

7 Conclusion and future work

In this work, we presented the Semantic Quran, an integrated parallel RDF dataset in 42 languages. This multilingual dataset aims to increase the availability of multilingual data in LOD and to further the development of NLP tools for languages that are still under represented, if not absent, from the LOD cloud. Thanks to its RDF representation, our dataset ensures a high degree of interoperability with other datasets. For example, it provides 26735 links overall to Wiktionary and DBpedia. As demonstrated by our use cases, the dataset and the links it contains promise to facilitate research on multilingual applications. Moreover, the availability of such a large number of languages in the dataset provides opportunities for linking across the monolingual datasets on the LOD Cloud and thus perform various types of large-scale analyses.

Listing 5.

List of all occurrences of “Moses” using NIF.

Listing 6.

List of all senses of all English words of the first verse of the first chapter “qrn:quran1-1”.

To improve the ease of access to our dataset, we aim to extend the TBSL framework [7] to allow even lay users to gather sensible information from the dataset. Moreover, we aim to provide links to the upcoming versions of Wiktionary. Additionally, we will link the Semantic Quran dataset with many of the publicly available multilingual Wordnets. We already provided NIF for the five languages Arabic, English, French, German and Russian. We will extend the NIF content of the dataset to the remaining 38 languages.