Enhancing Spatial Information Extraction from Arabic Text: A Hybrid Approach with Ontology and Rule-Based

In recent years, the rapid growth of digital information has highlighted the need for efficient information processing systems, especially for languages that are rich in vocabulary and complex in structure, such as Arabic [1]. Extracting spatial information from raw text has become a crucial research topic in various fields such as Automatic Natural Language Processing (NLP), Information Extraction (IE), Information Retrieval (IR) and Geographic Information Systems (GIS).

Spatial information extraction offers many advantages and can be applied to a variety of domains. It improves the accuracy and relevance of search results, optimises geographic information systems (GIS) [2], enriches geospatial databases, and enhances location-based services (LBS) [3]. It also facilitates decision-making in areas such as urban planning, natural resource management and disaster response. The information extraction process transforms unstructured textual data into structured output by identifying and extracting entities, relationships, semantic roles and events [4].

Due to its complex semantic and morphological features, the Arabic language faces significant challenges in the field of information extraction and retrieval, particularly for spatial information. Existing methods often prove ineffective in meeting these challenges [5, 6]. It is therefore necessary to develop a new approach to improve the accuracy and efficiency of this extraction.

Ontologies appear as an important solution to building a shared and reusable body of knowledge that support human-machine interaction and understanding [7]. They play a crucial role in knowledge representation and contribute to the development of the Semantic Web [8]. However, the exploitation of ontologies for the automatic indexing, retrieval, extraction and annotation of Arabic texts is still little explored. In the context of information extraction and search, the ontology is exploited to: 1) index entities or concepts of a specific domain and their diversities; 2) a hierarchical design to generate retrieval rules by the IE system; 3) the properties of concepts and the relations between these properties that guide the information extraction process; 4) The relationships with the concepts and their properties. The result of the information extraction process should be an updated/improved semantic annotation, model or ontology [9]. In this context, our study aims to answer the question of how to satisfy the needs of users of an Arabic GIS and which techniques to adopt to extract relevant spatial information from Arabic texts. To address these challenges, we propose a hybrid approach combining rule-based and ontology-based methodologies. This approach draws on the strengths of both techniques to deal with the linguistic complexities of Arabic, such as its morphological richness and semantic ambiguities.

The objectives of this study are as follows:

• Develop a hybrid approach that integrates rule-based and ontology-based methodologies for extracting spatial information from Arabic text.
• To construct the Arabic Spatial Toponym Ontology (ASTO) to aid in accurate indexing, annotation, and extraction of spatial data.
• To employ JAPE rules for precise disambiguation and classification of spatial information.
• To evaluate the effectiveness of the proposed approach through rigorous testing and demonstrate its potential to improve GIS application performance.

The remainder of this paper is organized as follows: Section 2 reviews related work. Section 3 presents the hybrid approach and describes the main components of the proposed system. Section 4 discusses the experiments and results. Finally, Section 5 provides the conclusion and perspectives for future work.

In the scientific community, geographic information is defined as a composition of three concepts: spatial, temporal, and thematic information. The main idea is that the combination of these three types of information allows for describing an event that is occurring or has occurred at a specific location and time. Among these components, this paper focuses on the processing of spatial information.

Currently, the integration of ontologies and rules has garnered significant attention in research related to ontologies and the semantic web [5]. The rise of ontologies is noteworthy as they can represent knowledge across various domains. This knowledge aids in semantic search by enhancing accuracy through understanding the purpose and contextual meaning of terms as they appear within the data space. The proposed approach consists of two main parts: first, the construction of the Arabic Spatial Toponym Ontology (ASTO), which represents spatial entities and relations; and second, an automatic information extraction system that leverages the built ontology along with natural language processing techniques and JAPE rules in GATE [27].

The general architecture of the proposed approach (see Figure 1) comprises four phases. In the first phase (Ontology Creation) (Figure 2), concepts are collected, and classes (representing Arabic spatial entities and relations) (Figure 3) are created for use in the text-matching steps. The second phase (Text Processing) applies various modules to annotate named entities within the text. The third phase (Combination and Extraction) utilizes the ontology to match classes, subclasses, or instances with the text (Figure 4). The fourth phase (Disambiguation and Classification) employs JAPE rules, which consist of algorithms developed and implemented in Java (Figure 5).

1.png

Figure 1. Phases of proposed approach

3.1 Ontology creation (ASTO: Arabic spatial toponym ontology)

Ontologies serve as a crucial solution for developing a set of shared and reusable knowledge that facilitates interaction [28], enabling interoperability and integration of various systems. Automatic extraction of ontological relations from texts is vital for representing documents and their content in a computerized and machine-readable format [29]. Ontology, from a philosophical point of view, is a discipline that deals with the nature and organization of being. With the development of information technology, ontology receives a new definition an Ontology is a formal specification of a shared conceptualization of a domain, it represents concepts using a human-understandable and machine-readable format and it consists of entities, values, relations and axioms [30].

As a conceptual model used for description at the semantic and knowledge levels, ontology is applied across many information processing domains, including knowledge engineering, digital libraries, software reuse, information extraction, and the semantic web [19].

When constructing the ASTO ontology, we follow the methodology outlined by Noy and McGuinness [31], which involves eight stages. The first stage is to define the domain. The next three stages are for constructing the taxonomy and the ontology and providing a formal description. The subsequent four steps include creating classes, conducting the reasoning process, performing a consistency check, and generating the result set (see Figure 2). 

The development of an ontology involves iterative refinement: an initial version is constructed, evaluated through applications or expert review, and then refined until a functional ontology is achieved.

2.png

Figure 2. A general procedure of an ontology construction [31]

3.png

Figure 3. Conceptual Ontology (ASTO) model

4.png

Figure 4. Pseudo-code for algorithm extraction spatial entities

Initially, as detailed by Konys [25], the process of building an ontology begins with a thorough domain analysis. This involves selecting an appropriate combination of OBIE approaches. Based on this analysis, a final set of characteristics and sub-characteristics is determined, followed by the formation of a class hierarchy. This hierarchy forms the general foundation for constructing taxonomies in existing OBIE systems and tools. During this process, various criteria and sub-criteria are established, and elements such as concepts, relationships, and attributes are integrated from disparate sources into the classification model [32].

Therefore, the taxonomy is based on the construction of the ontology. To implement the ontology, the Web Ontology Language (OWL) is used. OWL provides a formal and structured method for collecting, organizing, and sharing data, along with numerous features for efficient ontology management. The consistency of the created ontology is verified by defining a set of classes and using a reasoning mechanism to validate them and the entire ontology. The accuracy of the results obtained from this process confirms the ontology's overall consistency and coherence. It is important to mention that we have tested this process using various validation queries.

5.png

Figure 5. Example of rules JAPE for extraction spatial location and spatial object

3.1.1 Geographic object (spatial entity)

In our model, geographic objects are defined as spatial entities, which can be categorized into multiple geographic types to reflect their physical and natural attributes. Each geographic feature can have multiple names. During the development of the ASTO ontology, several iterations were required for the following reasons: initially, it was unclear whether the collected terms were sufficient to fulfill the ontology's purpose. New terms were added as needed, while unnecessary terms were removed.

As shown in Figure 3, the conceptual model of the Arabic Spatial Toponym Ontology consists of various classes and subclasses, with each subclass containing a set of instances and properties (see Table 2).

3.1.2 Spatial relations

In the presented approach, we focus on spatial relationships. Identifying these relationships is crucial for various tasks, such as reducing ambiguity, enhancing geographic information extraction, optimizing navigation systems, managing traffic, spatial reasoning, and responding to queries.

The identification of spatial relationships is a key element in the development of a new spatial information system. In ontology studies, we define relationships between geographic entities considering topological, orientation, distance, and directional relations, as shown in Table 2. We have developed a spatial ontology tailored to the Arabic language to account for the extensive range of terms and meanings, which improves the efficiency of spatial information extraction and adds originality to the work presented in this research.

3.2 Data sources and preprocessing techniques

To build our corpus of Arabic text, we follow rigorous selection criteria to ensure diversity, relevance, and quality. We selected articles from Arabic newspapers based on the diversity of media sources, the relevance of the content to spatial information, and the publication date to ensure the temporal relevance of the data. The newspapers were also chosen from among the most readable, well-known, and widely followed.

Articles were manually collected from four major Arabic information sources (see Figure 6): Echourouk (https://www.echoroukonline.com) News, Elkhaber (https://www.elkhabar.com) News, Ennahar (https://www.ennaharonline.com) News, and Elmihwer (http://elmihwar.dz/ar) News. The final corpus covers various domains such as human disasters (e.g., accidents, fires), natural disasters (e.g., floods), and technological events. Overall, the corpus totals approximately 9,000 tokens, manually collected from the mentioned websites [33].

6.png

Figure 6. Percentage of spatial information extraction

Table 2. Example of class and instance of ASTO

The preprocessing techniques applied to our corpus of Arabic texts include several crucial steps. We began with data cleaning, removing extra spaces and formatting elements to obtain raw text. Next, the typographical errors were then corrected using both automated and manual tools, followed by text normalization, including removing diacritics, normalizing spaces, and punctuation. Finally, we performed manual tokenization, in which texts were manually divided into meaningful units (sentences or phrases) to ensure accuracy and appropriateness for subsequent analysis.

 These cleaning and normalization operations are vital to ensure the texts are properly prepared for thorough and effective analysis.

3.3 Text processing

General Architecture for Text Engineering (GATE) is known for its popularity in text analysis and automatic natural language processing. It has been widely utilized in various projects, including information extraction (IE) in different languages. The main advantage of GATE lies in its extensibility, as it allows the addition of new components or applications to its existing standard components, thereby enabling the expansion of analysis and text processing capabilities. The primary focus of GATE is document annotation, which can be achieved through three methods: fully automatic using applications, manual annotation by users, and semi-automatic annotation through a combination of corpus processing and manual correction or addition of new annotations [34].

GATE offers a range of specialized modules for textual analysis, accessible through a system of plugins. Some of the commonly used modules include tokenizers (segmenters) for breaking text into smaller units, Part of Speech Taggers (morpho-syntactic taggers) for categorizing words, Gazetteers (lexicons) for identifying specific terms, and transducers (JAPE) for pattern-based matching. The named entities extracted by GATE encompass various types such as places, organizations, people, and dates, among others. With GATE, users can create new annotations by loading additional resources like corpora, documents, and texts, as well as incorporating new plug-in. It allows for the combination and parameterization of these resources within the same processing chain [35]. As an open-source Java platform dedicated to text engineering, GATE is well-suited for the development of systems that require natural language processing and information extraction capabilities.

3.3.1 Sentence splitter

In Natural Language Processing (NLP), dividing a text into sentences is a crucial step to extract more meaningful information from it. Each sentence is delimited by identifiers, such as punctuation marks like periods, exclamation marks, or question marks. Initially, a sentence is considered a unit of discourse that is acceptable to the application or NLP system being used. Segmenting a text into sentences allows NLP systems to process and analyze smaller chunks of text at a time, making it easier to understand the context and extract relevant information. By breaking the text into sentences, NLP models can focus on the relationships between words within each sentence, which helps with tasks such as named entity recognition, part-of-speech tagging, sentiment analysis, and other language understanding tasks. Sentence segmentation is an essential preprocessing step in many NLP applications, as it lays the foundation for further analysis and enables the extraction of meaningful insights from a given text.

3.3.2 Arabic tokenizer

The tokenizer is an essential component used to divide text into words, numbers, symbols, spaces, and punctuation marks. Each of these units is referred to as a token, representing a distinct syntactic element, which can be a complete word, part of a word, a multi-word phrase, or a punctuation mark. The tokenizer relies on a predefined list of word delimiters, such as whitespace and punctuation marks, and considers the nuances of tokens within words, especially when dealing with word stems and clitics. In this research, complete words, including stems with or without clitics, as well as numbers, are referred to as main tokens. These main tokens are delimited by whitespace or punctuation marks. Additionally, full-form words can be divided into sub-tokens by separating clitics from stems [35].

3.3.3 POS tagging

POS tagging is an important step in the process before extraction and recognition of named entities. It identifies and adds tags to the tokenized text model, i.e., identifying nouns, verbs, adjectives and other parts of speech for each token. The POS tagger used in this study with GATE is the Hepple tagger [36].

3.3.4 Analysis morphological

Morphology [36], in linguistics, is the study of the internal structure and composition of words and their formation. A morpheme is defined as the shortest meaningful unit of language, which cannot be divided down into shorter parts. The morphological analyzer helps to group words that express similar concepts. The role of the morphological analyzer in Arabic is to identify the morphemes of a word (stem): the affixes (prefix, infix, and suffix) and the root.

3.4 Combination and extraction 

Our hybrid spatial OBIE approach is a contribution to the field of information extraction based on the OBIE ontology. Additionally, ontologies allow formal and declarative descriptions of common terms and support automatic or semi-automatic reasoning about shared data in a domain. The main idea is related to the search and retrieval of spatial information.

The process starts with loading the ontology into GATE as an OWL ontology resource and provides facilities for loading corpora from a URL or file to the Pipeline application. The system in this step matches concepts, instances and relations between them of the ontology with an input Arabic text for extraction and annotation of spatial information. Consequently, if the system detects an identical word in the input text that is a class or instance in the ontology, it annotates it. The result obtained at this stage is an annotated corpus containing the set of annotated words (spatial information) with an ambiguity (see Figure 7), we cannot determine the spatial entities, spatial relations, or classes of each entity.

3.5 Disambiguation and classification

JAPE rules [37], consist of a sequence of phases, each consisting of a sequence of model/action rules. Each phase executes sequentially and is a series of state transducers finished as annotations. The rules consist of two sections: the left-hand section (LHS), which describes an annotation model, and the right-hand section (RHS), which consists of annotation manipulation instructions. The corresponding annotations on the left side of a rule can be identified on the right side by means of labels linked to the model elements.

The main functions of the JAPE transducer [37, 38] in this step are to combine and extract the spatial entities, relations and their classes of the ASTO ontology with input corpus. Then, the system searches into the Ontology rules whether the annotated word is contained (as subclass or instance) in the class matched the LHS part of the JAPE rule and if it found, the system classifies the annotated word in this class.

The following algorithm (Figure 4) shows the extraction of the spatial entities (Object and Location):

3.5.1 Detailed algorithm for algorithm extraction spatial entities

Inputs: T: Text row to analyze, Ont: Ontology containing classes and subclasses, S: Sentence extracted from the text, W: Word extracted from the sentence.

Output: CorpusT: Annotated and classified text corpus

Detailed Explanation of Steps

Initialization:

• The text T is parsed to extract individual words W. This is a preprocessing step that prepares the text for further analysis.

7.png

Figure 7. Extraction based Ontology without the use of JAPE rules

Sentence Processing:

• Each sentence S in the text is processed individually.

Word Processing: Each word W in the sentence is examined to determine if it matches a class in the ontology Ont.

Spatial Object Classes:

• If the current class in the ontology is an object (spatial-Object), the algorithm checks each subclass to see if the word W is an instance of that subclass.
• If a match is found, the word W is annotated and extracted as a spatial object in the output corpus CorpusT. The subclass index is then incremented to continue checking other subclasses.

Spatial Location Classes:

• Similarly, if the current class is a location (spatial-Location), the algorithm checks each class to see if the word W is an instance of that class.
• If a match is found, the word W is annotated and extracted as a spatial location in the output corpus CorpusT. The class index is then incremented to continue checking other classes.

End of Algorithm:

• The algorithm continues until all the words in all the sentences have been processed and extracted according to the ontology classes Ont. We employed the same algorithm for annotating and extracting each spatial relation (topological, distance, orientation, direction).

For example, the spatial entities transducer employs an explicitly defined JAPE rule that is able to match the object entity (Natural/Building) and location entity models which are listed in the following (Figure 5).

We use JAPE rules for disambiguation and classification of spatial entities and spatial relations in our text corpus. Figure 5 illustrates two distinct phases, each aimed at optimizing the annotation and classification of spatial data.

OntoMatching_Spatial_Object Phase: This first phase focuses on the annotation of spatial objects. The applied JAPE rules check if an annotation matches the spatial_object class defined in our ontology. When this match is found, the annotation is enriched with detailed information, including the class and instance of the object. This process ensures not only precise identification of spatial objects but also effective disambiguation by associating each entity with a clear ontological definition.

 OntoMatching_Spatial_Location Phase: The second phase follows a similar process for spatial locations. The JAPE rules check for correspondence with the spatial_location class and enrich the relevant annotations accordingly. This treatment improves the accuracy of spatial locations and ensures coherent and informed annotation.

These two phases enable automated identification and enriched classification of spatial entities, based on a well-defined ontology. We used the same JAPE rules for the classification and disambiguation of each spatial relation (topological, distance, orientation, direction).