In the first phase, concepts related to Arab entities and spatial relationships are identified and collected. These concepts are then used to formulate specific JAPE rules [ 15]. which are used to annotate and extract relevant spatial information from Arabic texts. JAPE rules are advanced regular expressions developed in Java, enabling the detection of complex patterns in text.

JAPE rules ofer significant flexibility in natural language processing, particularly for extracting information from unstructured text. Their main strength lies in the ease of adding or modifying new rules. It is straightforward to integrate new words or expressions into an existing system without disrupting the functionality of previously defined rules. This ability to quickly update the rules based on domain evolution or analysis needs makes JAPE a particularly adaptable and eficient tool for tasks such as entity recognition and contextual information extraction.

The second phase consists of applying natural language processing modules to prepare the raw text. This process includes steps such as normalization, tokenization and annotation of the spatial entities present in the text. These modules are crucial to ensuring that JAPE (Java Annotation Patterns Engine) rules can be applied eficiently and accurately.

The third phase is based on the application of the JAPE rules created in the first phase. These rules are used to associate text segments with defined classes, subclasses or instances. This phase is essential for automatically extracting structured spatial information from unstructured text, taking into account the linguistic and contextual specificities of the Arabic language.

The fourth phase focuses on disambiguation and classification of the extracted spatial entities. This step ensures that each entity and relationship is correctly interpreted in its specific context. JAPE rules are also used here to refine the results, applying disambiguation and classification criteria to improve the accuracy of the extracted data.

Our JAPE rule-based system architecture consists of two main phases, each playing a crucial role in the extraction of geographic information from natural language text (Figure 2). The first phase uses advanced Natural Language Processing (NLP) techniques to prepare and normalize text data. This preparation includes text cleaning, sentence segmentation and initial annotation of linguistic elements, facilitating better rule application.

The second phase focuses on matching JAPE rules to extract specific information. This phase involves the definition and creation of rules, the matching of these rules with the text, disambiguation and the extraction of relevant information. Finally, post-processing is carried out to filter and structure the extracted data, making it ready for further analysis or integration into geospatial databases. Together, these phases ensure accurate and eficient information extraction, tailored to the needs of geographic analysis.

We have chosen to use the GATE environment, a linguistic engineering framework developed by the University of Shefield and widely adopted since its first release in 1996 for teaching and research. GATE ofers a suite of reusable processing resources in JAVA, integrated into an information extraction system called ANNIE (aNearly-New Information Extraction System) [16].

By default, ANNIE is configured for languages other than Arabic. To adapt this tool to our target language, we will use specialized components such as the Arabic tokenizer, sentence splitter, POS tagger and Arabic morphological analyzer. To avoid interference with previous executions, we’ll apply the “reset” option to remove all traces of previous processes. Annotations in GATE will be performed by selecting words in the text and creating new annotation categories, enabling precise extraction of geographical features and other relevant information.

The first phase of our architecture implements NLP techniques that are essential for processing and understanding natural language text. The aim of this phase is to prepare the textual data in such a way as to facilitate the extraction of relevant information, bearing in mind that we have used the same dataset or corpus discussed in [ 1 ].

The second phase focuses on the application of JAPE rules for the extraction of specific spatial information. This phase follows the well-defined steps between rules and spatial information, i.e. the classification of spatial entities and relationships according to the following table

The Table 1 below presents a classification of spatial entities, including natural entities, non-natural entities and entities corresponding to place names or locations.

The Table 2 presents a classification of spatial relationships, detailing the following categories: topological relationships, directional relationships, distance relationships and orientation relationships. These categories help us understand how spatial entities position, orient and relate to each other in a given space. Topological relations describe relationships of contiguity or inclusion, directional relations indicate relative orientations, distance relations measure distances between entities, and orientation

A JAPE grammar consists of a set of phases, each containing a series of pattern/action [15]. The phases execute sequentially, forming a cascade of finite-state transducers on the annotations. The left-hand side (LHS) of the rules comprises an annotation pattern description, while the right-hand side (RHS) contains instructions for manipulating the annotations. Matching annotations on the LHS of a rule can be referenced on the RHS using labels attached to pattern elements. Below is an example of a JAPE rule (Figure 5) for extracting named entities from the "Non-Natural Object" class containing specified

instances: ÉJ. k. (mountain), éJ. ’ë (plateau), ù£Aƒ (beach), éK. A« (forest), X@ð (valley), Qm '. (sea), Z@Qm • The defined rules (Figure 5) are applied to the text or data to identify segments that match the specified patterns. A rule could be designed to identify geographic entities by searching for phrases containing keywords such as "region," "city," or "country." In our method, the option control = appelt is used to specify that the rules should be executed sequentially. This ensures that each rule is applied in a precise order, thereby maximizing the accuracy of the extraction.

When matching segments are identified, relevant information is extracted (Figure 6). This may include capturing specific words or phrases or identifying relationships between diferent entities. For example, the rule "Extract Natural Object" in our JAPE script checks for text matching one of the specified instances in a list of natural objects, such as é K. A « (forest), X@ð (valley), or Q m '. (sea). When a token matches a word in the text, a named entity "Natural Object" is created, and specific attributes, such as class = "Natural Object" and type (containing the character string of the identified entity), are associated with this entity.

• Precision: Our approach demonstrates a high precision of 0.90, indicating that 90% of the spatial information extracted is relevant and correct. This is a significant advantage, especially in applications where accuracy is paramount, such as in disaster response scenarios or when processing sensitive geographic data. In comparison, the rule-based methods [ 4 ] and [19] show lower precision values of 0.80 and 0.85, respectively. This discrepancy suggests that while these methods can recall a broader range of information, they may also include a higher number of false positives. • Recall: In terms of recall, our approach achieves a score of 0.85, which indicates a solid ability to capture a significant proportion of the actual relevant spatial information present in the texts. Although this recall rate is lower than that of [ 4 ] (0.91), it remains competitive, especially when considering that higher recall often comes at the cost of lower precision. The approaches based on rules [19] and the hybrid method exhibit similar performance levels in recall, with scores of 0.88 and 0.95, respectively. This suggests that while our method may miss some relevant entities compared to the others, it does so while maintaining a high level of accuracy. • F-measure: The F-measure, which balances precision and recall, is another critical metric for assessing the overall performance of the approaches. Our method achieves an F-measure of 0.87, which reflects a strong performance overall. The hybrid approach [ 1 ] leads in this area with an impressive F-measure of 0.94, underscoring the efectiveness of combining diferent techniques to leverage their respective strengths. While our approach does not outperform this hybrid model, it still outshines the purely rule-based approaches, [ 4 ] and [19], which both yield lower F-measure scores of 0.85.

The analysis indicates that while our JAPE rule-based method excels in precision, making it a robust option for applications that require accuracy, it falls slightly behind in recall compared to some other methods. This presents a critical trade-of in the context of information extraction: achieving a high precision often limits the breadth of recall. The hybrid method, although potentially more complex to develop and implement, demonstrates the highest overall efectiveness, suggesting that a combined strategy could yield the best results in future applications.

Moving forward, our findings advocate for a nuanced approach to spatial information extraction that considers the specific requirements of each task. For instance, in scenarios where precision is paramount, our JAPE method stands out as an ideal choice. Conversely, for applications requiring extensive coverage of information, exploring hybrid methodologies could enhance performance significantly. Further research could involve developing a hybrid model that integrates the best features of our JAPE approach with the comprehensive capabilities of hybrid and machine learning methods, aiming to improve both precision and recall without sacrificing eficiency.