Motivations

[Audio] Motivations Agricultural Knowledge Systems (AKS): defined as frameworks that facilitate the generation, sharing, and application of agricultural information among farmers, researchers, intermediate agents and policymakers. AKS increasingly depends on the ability to manage vast volumes of textual information in a structured and accessible manner. To enhance the functionality and impact of AKS, efficient information and data sharing is essential. Automated semantic annotation tools are becoming indispensable, as they help structure and interconnect data by linking it to knowledge graphs. Natural Language Processing (NLP) offers a solution by enabling the precise tagging and categorization of agricultural texts according to specific entities. By integrating fine-grained text annotation and classification into AKS, stakeholders can enhance knowledge retrieval, support machine learning applications, and enable more targeted decision-making across the agricultural value chain. الدوافع أنظمة المعرفة الزراعية (AKS): تُعرّف بأنها أطر عمل تُسهّل توليد المعلومات الزراعية وتبادلها وتطبيقها بين المزارعين والباحثين والوسطاء وصانعي السياسات. تعتمد أنظمة المعرفة الزراعية (AKS) بشكل متزايد على القدرة على إدارة كميات هائلة من المعلومات النصية بطريقة منظمة وسهلة الوصول. لتحسين وظائف وتأثير أنظمة المعرفة الزراعية (AKS)، يُعدّ تبادل المعلومات والبيانات بكفاءة أمرًا أساسيًا. أصبحت أدوات التعليق الدلالي الآلي لا غنى عنها، إذ تُساعد على هيكلة البيانات وربطها ببعضها من خلال ربطها برسوم بيانية معرفية. تُقدّم معالجة اللغة الطبيعية (NLP) حلاً من خلال تمكين وضع العلامات وتصنيف النصوص الزراعية بدقة وفقًا لجهات مُحددة. من خلال دمج التعليقات والتصنيفات النصية الدقيقة في أنظمة المعرفة الزراعية (AKS)، يُمكن لأصحاب المصلحة تحسين استرجاع المعرفة، ودعم تطبيقات التعلم الآلي، وتمكين اتخاذ قرارات أكثر استهدافًا عبر سلسلة القيمة الزراعية..

[Audio] NER attempts to identify and classify named entities within unstructured text into pre-defined categories. NER faces several challenges, including limited availability of annotated data, high manual annotation costs, expertise requirements, low accuracy of generic NER, and the impracticality of reusing domain-specific tools. Many approaches address coarse-grained annotation with broad entities. The coarse-to-fine classification problem occurs when annotations become too generic and fail to support efficient content discovery in large or rapidly evolving datasets. Texts about poultry housing, social housing, or piggeries are frequently annotated with the umbrella term housing in databases like AGRIS or PubMed. This practice, pose limitations to support real-world applications. Developing methods that can accurately identify and classify fine-grained entities is essential, although it poses a significant challenge: increasing number of named entity types, single entity may belong to multiple types, further complicates the task. Integrating KGs with LM scan play a crucial role in advancing AKS. This synergy can greatly enhance information accessibility and data sharing, thereby driving innovation and improving decision-making in the food and agriculture sectors..

[Audio] Dictionary-based systems often achieve high precision, as the identified entities closely align with curated concepts. However, they typically suffer from low recall because they cannot detect fine-grained or paraphrased entities that do not exactly match ontology entries. Coarse-to-fine grained classification was initially coined by Mekala et al. (2021). It aims to perform fine-grained classification on coarsely annotated data by leveraging a pre-defined fine-grained label hierarchy..

[Audio] LANGUAGE MODELS Deep learning (DL) has significantly advanced NLP, enabling the transformation of unstructured text into structured data. Large language models (LLMs) such as GPT-3 and GPT-4 have opened new possibilities for improving NER through large-scale training and transformer architectures. The development of generative language models has progressed through several distinct phases.: Phase 1: Early NLP relied on statistical and rule-based approaches. Phase 2: The introduction of RNNs and CNNs improved natural-language understanding and generation. Phase 3 (current): Transformer-based models using the attention mechanism dominate modern NLP. Recent research focuses on adapting LLMs to specific tasks and domains without requiring full retraining. Retrieval-augmented generation (RAG) approaches enable LLMs to integrate external knowledge and execute complex instructions. LLMs using the: Encoder–Decoder architecture (e.g., T5) encode the full input into a semantic representation and then decode it to generate output. They are well-suited for tasks needing bidirectional understanding, like translation and summarization. Decoder-Only (e.g., GPT): Each word is generated sequentially based on the words predicted so far, making this architecture ideal for text generation, dialog systems, and code synthesis. Encoder-Only (e.g., BERT): Processes the input to create contextual representations without generating new text, making it suitable for understanding tasks like text classification, named entity recognition, and sentiment analysis..

[Audio] Prompt engineering Prompt engineering is the strategic design and structuring of input text to guide models toward producing accurate and contextually appropriate outputs. Prompts act as lightweight instructions that can elicit task-specific behavior from a model, often without requiring additional fine-tuning or retraining. This approach is particularly valuable in scenarios where labeled data is scarce or rapid prototyping is needed. Basic prompting typically involves the following sequential steps: Step 1: Turning the source text into another text (called prompt) that fits the model, by means of a function called prompting template (Radford et al., 2019; Schick & Schütze, 2020). Step 2: Predicting the highest scoring answer by computing the probability of the filled prompt using a pre-trained language model P(y|x; θ), with θ representing the model parameters learned during training. Step 3: Mapping the highest scoring answer to the highest-scoring output. Foundational Prompting Approaches Use of clear and precise instructions is essential for guiding models toward the desired behavior. Role prompting assigns the model a specific role or persona to improve performance. This approach is particularly effective in domains where expertise and accuracy are critical. One-shot prompting provides the model with a single example of the desired task output. It is effective when a single example suffices to establish a clear pattern. Few-shot prompting offers multiple examples, which is helpful for more complex tasks. Zero-shot prompting involves performing a task without providing any explicit examples. This approach is particularly useful in scenarios where labeled data is scarce or nonexistent..

[Audio] PLMs & LLMs Pre-trained language models (PLMs): are trained on large amounts of text data and undergo initial training on a specific task before being fine-tuned for a particular application. like masked language modeling (MLM) or next-token prediction. LLMs, on the other hand, are trained on even larger amounts of data and are designed to generate coherent and contextually appropriate text. They are often used for tasks such as language translation, text summarization, and chatbots. While PLMs are typically smaller and more specialized, LLMs are larger and more general-purpose. Training task-specific models from scratch has largely become obsolete, significantly reducing the data and computational requirements for domain adaptation. LMs & KGs Although LMs have shown remarkable performance, they have not yet reached optimal performance. To address these limitations, researchers have turned to KGs as a means of enhancing LMs’ performance . KGs explicitly represent knowledge using symbols, such as entities and their relationships, facilitating symbolic reasoning and yielding interpretable outcomes. By integrating KGs with LMs, researchers aim to achieve better performance and interpretability of both technologies in various NLP tasks AGROVOC is a dataset. AGROVOC is widely used to index the AGRIS database..

[Audio] State-of-the-art solutions Semantic text classification is the process of assigning labels or categories to textual documents based on their semantic content. Inspired by how humans can classify text using only a few descriptive words, Meng et al. (2020) leverage PLMs to learn from unlabeled data using only the label names (1–3 words per class). Their approach associates related words with labels, identifies category-indicative terms, and applies self-training to generalize. It achieves around 90% accuracy on four benchmark datasets without using any labeled data. Wang et al. (2021) used extremely weak supervision for text classification, i.e., based only on class names and without seed words. The approach builds class representations by incrementally adding similar words and uses a specialized attention mechanism to create document embeddings. Documents are then clustered and aligned to classes, and the most confident ones are used to train a classifier. (Nentidis et al., 2023) addresses the semantic classification of biomedical literature using deep learning with weak supervision strategies, which enable the generation of pseudo-labeled training data. The results indicate that while weakly supervised learning can be effective, there are significant challenges regarding data precision and coverage. Mekala et al. (2021) propose an innovative approach to fine-grained text classification using coarsely labeled data. Their method, known as Coarse2Fine, combines PLMs with weak supervision strategies. This approach enables fine-grained classification without the need for annotations. Mekala et al. (2021) generate pseudo-labeled training data from coarse labels and train a label-conditioned language model to refine the classification. Their method is designed to produce fine-grained labels without requiring additional human annotation, relying on a pre-defined label hierarchy. The approach integrates PLMs within an iterative weak-supervision framework..

[Audio] AGROVOC AGROVOC (is a dataset) is widely used to index the AGRIS database. AGROVOC is a dataset published as Linked Open Data, forming part of the Linked Open Data cloud, and is curated by the Food and Agriculture Organization of the United Nations (FAO), covering different areas such as agriculture, forestry, and food systems. AGROVOC’s terminology serves as a valuable resource for text mining in the agricultural sector, enabling the identification and tracking of concepts mentioned in text. The semantic annotation of AGROVOC concepts facilitates the integration of annotated datasets and supports advanced semantic analysis techniques. AGROVOC suffers from limited term coverage, and its vocabulary does not always align well with the varying language used in domain-specific text corpora. A key example of a domain-specific knowledge graph in agriculture is AGROVOC. used to enhance data annotation, semantic search, and information retrieval in agricultural information systems. Each concept within AGROVOC is assigned a unique Uniform Resource Identifier (URI). AGROVOC is implemented using the RDF/SKOS-XL model, which enables rich semantic interoperability and integration with other linked data vocabularies. The SKOS-XL model enables users to explore AGROVOC’s contents through various access methods, including a SPARQL endpoint, REST API, SOAP web services, or by downloading the dataset. يُستخدم AGROVOC (مجموعة بيانات) على نطاق واسع لفهرسة قاعدة بيانات AGRIS. AGROVOC هي مجموعة بيانات نُشرت كبيانات مفتوحة مرتبطة، وهي جزء من سحابة البيانات المفتوحة المرتبطة، وتُشرف عليها منظمة الأغذية والزراعة للأمم المتحدة (الفاو)، وتغطي مجالات مختلفة مثل الزراعة والغابات ونظم الأغذية. تُعدّ مصطلحات AGROVOC موردًا قيّمًا لاستخراج النصوص في القطاع الزراعي، مما يُمكّن من تحديد المفاهيم المذكورة في النصوص وتتبعها. يُسهّل الشرح الدلالي لمفاهيم AGROVOC دمج مجموعات البيانات المُشرحة، ويدعم تقنيات التحليل الدلالي المتقدمة. يعاني AGROVOC من محدودية نطاق المصطلحات، ولا تتوافق مفرداته دائمًا مع اللغة المُستخدمة في مجموعات النصوص الخاصة بمجالات مُحددة..

[Audio] AGROVOC is utilized in AgroPortal, a repository of agricultural ontologies and annotation tools, and is also employed to support NER and semantic indexing in agricultural literature. AgroPortal project offers a publicly accessible ontology-based annotation service for agricultural texts in English. AgroPortal integrates numerous agricultural ontologies and provides tools for identifying and linking terms in unstructured text to formal concepts. AgricultureBERT is a domain-adapted pre-trained language model (PLM) designed to enhance NLP applications in agriculture. Knowledge graphs (KGs) are structured representations of information in the form of entities (nodes) and the relationships (edges) between them. A KG typically follows a triple structure—subject–predicate–object—where entities are connected through meaningful relationships. This structure enables machines to interpret, reason over, and derive insights from the data. يُستخدم AGROVOC (مجموعة بيانات) على نطاق واسع لفهرسة قاعدة بيانات AGRIS. AGROVOC هي مجموعة بيانات نُشرت كبيانات مفتوحة مرتبطة، وهي جزء من سحابة البيانات المفتوحة المرتبطة، وتُشرف عليها منظمة الأغذية والزراعة للأمم المتحدة (الفاو)، وتغطي مجالات مختلفة مثل الزراعة والغابات ونظم الأغذية. تُعدّ مصطلحات AGROVOC موردًا قيّمًا لاستخراج النصوص في القطاع الزراعي، مما يُمكّن من تحديد المفاهيم المذكورة في النصوص وتتبعها. يُسهّل الشرح الدلالي لمفاهيم AGROVOC دمج مجموعات البيانات المُشرحة، ويدعم تقنيات التحليل الدلالي المتقدمة. يعاني AGROVOC من محدودية نطاق المصطلحات، ولا تتوافق مفرداته دائمًا مع اللغة المُستخدمة في مجموعات النصوص الخاصة بمجالات مُحددة..

[Audio] Expanding terminology in ontologies and knowledge graphs (KGs) Expanding an ontology with synonyms that better reflect real-world usage can significantly improve the effectiveness of NLP tasks involving ontology-based models . Query expansion techniques aim to enhance concept discovery by generating new synonym substitutions. Some approaches generate new synonyms by removing shared substrings between synonym pairs (e.g., from "kidney biopsy" and "renal biopsy," deducing that "kidney" and "renal" are synonyms). These methods face challenges : Often substitute terms without considering contextual semantics, leading to invalid or meaningless results. They tend to generate a large volume of low-quality synonyms, creating a combinatorial explosion. To mitigate this, subsequent research has proposed constraints on algorithm parameters such as the number of replacements per term, term length , term frequency in the ontology, and maximum synonyms per term..

[Audio] Taboada et al. (2017) propose leveraging the hierarchical structure of ontologies to guide term expansion. In their method, synonym replacement is limited to strings shared between a term and its ancestor concepts, effectively reducing homonymy. However, their approach still lacks contextual awareness, often generating synonyms that are irrelevant or misleading in specific text settings. These limitations highlight the need for context-aware query expansion techniques capable of predicting high-quality synonyms, especially in domains such as agriculture and biomedicine, where semantic precision is critical. The proposal in this PhD builds on the strategy of Taboada et al. (2017) but enhances it by confirming new synonyms using knowledge from language models (LMs). Word embeddings: vector representations of words that capture both semantic meaning and syntactic relationships..

[Audio] ALGORITHMS FOR EXTENDING TERMINOLOGIES IN ONTOLOGIES Various approaches have been proposed to extend the terminology of ontologies using the knowledge contained within the ontology itself. In these approaches, new synonyms are often generated from multi-word phrases by replacing one or more words with known synonyms. For example, if “kidney biopsy” is a synonym for “renal biopsy”, the common word “biopsy” can be removed, allowing synonymy to be inferred between “kidney” and “renal”. One limitation of this approach is the generation of millions of candidate synonyms, many of which are not meaningful or useful. Additionally, the approach faces challenges with homonyms, where a single word may have multiple unrelated meanings, potentially leading to incorrect synonym mappings. Hole and Srinivasan (2000) use uncommon strings of synonymous terms within the ontology itself to generate new synonyms. This approach faces two main issues: the excessive generation of incorrect synonyms and the challenge of homonymy. Bodenreider et al. (2002) extract noun phrases from PubMed to generate synonyms. The primary challenge of this approach is the incorrect identification of parts of speech and acronyms, which can result in errors in synonym generation. Huang et al. (2007) apply uncommon chains of synonymous terms within the ontology while introducing two constraints: a maximum number of substitutions per term and a maximum term length. This approach addresses the challenges of excessive generation of incorrect synonyms and homonymy. In this PhD work, we focus on leveraging LMs to enhance the accuracy of the synonym generation process.

[Audio] Ontology Matching Algorithms Ontology matching (OM) is a fundamental task that involves identifying semantic relationships between entities from different ontologies. This process is essential for achieving data interoperability and integrating heterogeneous knowledge sources. Ontology matching (OM) algorithms are designed to establish correspondences between terms generated by LMs and formal concepts defined in ontologies. Without effective OM, many relevant concepts extracted by the LM would remain unlinked to the reference ontology, reducing the quality and utility of annotations. OM algorithms enable the identification of these semantic equivalences, even in the presence of lexical variation, paraphrasing, or compound term generation. By leveraging OM, annotations can be accurately mapped to their corresponding AGROVOC concepts, ensuring consistency across datasets, improving interoperability between systems, and enhancing the overall performance of downstream NLP tasks such as indexing, retrieval, and semantic search. مطابقة الأنطولوجيا (OM) مهمة أساسية تتضمن تحديد العلاقات الدلالية بين الكيانات من أنطولوجيات مختلفة. هذه العملية ضرورية لتحقيق التوافق بين البيانات ودمج مصادر المعرفة غير المتجانسة. صُممت خوارزميات مطابقة الأنطولوجيا (OM) لإنشاء تطابقات بين المصطلحات التي تُولّدها خوارزميات التعلم والمفاهيم الرسمية المُعرّفة في الأنطولوجيات. بدون مطابقة أنطولوجية فعّالة، ستبقى العديد من المفاهيم ذات الصلة التي تستخرجها خوارزميات التعلم غير مرتبطة بالأنطولوجيا المرجعية، مما يُقلل من جودة وفائدة التعليقات التوضيحية. تُمكّن خوارزميات مطابقة الأنطولوجيا من تحديد هذه التكافؤات الدلالية، حتى في وجود اختلافات معجمية، أو إعادة صياغة، أو توليد مصطلحات مُركّبة. من خلال الاستفادة من مطابقة الأنطولوجيا، يُمكن ربط التعليقات التوضيحية بدقة بمفاهيم AGROVOC المُقابلة لها، مما يضمن الاتساق عبر مجموعات البيانات، ويُحسّن التوافق بين الأنظمة، ويُحسّن الأداء العام لمهام معالجة اللغة الطبيعية (NLP) اللاحقة، مثل الفهرسة والاسترجاع والبحث الدلالي..

[Audio] Existing OM systems have integrated techniques based on lexical, structural, and semantic matching , as well as mapping repair methods , leading to enhanced performance. OM systems still face two persistent challenges: First, distinguishing between semantically similar entities and those that frequently co-occur but are not conceptually related remains a difficult task . Second, scalability remains a major hurdle, especially when applying OM methods to large datasets. To tackle these issues, many OM systems require fine-tuning with large, domain-specific training datasets to achieve optimal performance. However, this approach is not always practical, particularly when such datasets are scarce. A promising alternative is the use of LLM-based methods for OM. These methods do not require fine-tuning and can directly identify semantic correspondences between ontologies. Modern LLM-based systems often adopt a retrieve-then-prompt approach: the system first retrieves a smaller set of relevant target entities (k) and then prompts the LLM to predict the correspondences between them. Although this approach offers a more efficient solution, it still faces several challenges: Its performance (e.g., F-measure) remains limited in certain scenarios. Scalability issues persist due to the time-intensive nature of querying LLMs, particularly when working with large-scale ontologies in real-world applications. دمجت أنظمة إدارة العمليات الحالية تقنياتٍ قائمة على المطابقة المعجمية والبنيوية والدلالية، بالإضافة إلى أساليب إصلاح التعيين، مما أدى إلى تحسين الأداء. لا تزال أنظمة إدارة العمليات تواجه تحديين مستمرين: أولاً، لا يزال التمييز بين الكيانات المتشابهة دلاليًا وتلك التي تتواجد بشكل متكرر ولكنها غير مرتبطة مفاهيميًا مهمةً صعبة. ثانيًا، لا تزال قابلية التوسع تُمثل عقبة رئيسية، خاصةً عند تطبيق أساليب إدارة العمليات على مجموعات بيانات كبيرة. ولمعالجة هذه المشكلات، تتطلب العديد من أنظمة إدارة العمليات ضبطًا دقيقًا باستخدام مجموعات بيانات تدريب كبيرة ومحددة المجال لتحقيق الأداء الأمثل. ومع ذلك، فإن هذا النهج ليس عمليًا دائمًا، خاصةً عندما تكون مجموعات البيانات هذه نادرة. يُعد استخدام أساليب قائمة على LLM لإدارة العمليات بديلًا واعدًا. لا تتطلب هذه الأساليب ضبطًا دقيقًا، ويمكنها تحديد التطابقات الدلالية بين الأنطولوجيات بشكل مباشر. غالبًا ما تعتمد الأنظمة الحديثة المستندة إلى LLM على نهج الاسترجاع ثم المطالبة: يسترجع النظام أولاً مجموعة أصغر من الكيانات المستهدفة ذات الصلة (k) ثم يطالب LLM بالتنبؤ بالمراسلات بينها. على الرغم من أن هذا النهج يوفر حلاً أكثر فعالية، إلا أنه لا يزال يواجه العديد من التحديات:يظل أداؤه (مثل مقياس F) محدودًا في بعض السيناريوهات.تستمر مشكلات قابلية التوسع نظرًا لطبيعة استعلام أنظمة إدارة التعلم التي تستغرق وقتًا طويلاً، خاصةً عند العمل مع الأنطولوجيات واسعة النطاق في التطبيقات العملية..

[Audio] RAG & Ontology Matching Algorithms To scale OM to large ontologies several LLM-based frameworks have adopted Retrieval-Augmented Generation (RAG) . which aims to mitigate hallucinations and content generation errors by combining retrieval mechanisms with generation. RAG systems typically employ a straightforward pipeline consisting of three stages: indexing the target ontology using a vector store, retrieving the most relevant candidate entities, and prompting the LLM with the candidate set . Although RAG-enhanced OM systems occasionally outperform traditional OM techniques in F1-score, their overall F-Measure still lags in consistency and robustness across domains. In addition, execution times remain high when applied to large-scale ontologies . This highlights the need for more efficient RAG strategies that improve F1 performance while minimizing the number of LLM invocations. Main Steps in LLM-based Ontology Matching 1. Vector KB Construction: Ontological data is extracted, preprocessed, and encoded into dense vector representations that are then indexed in a vector database. 2. Mapping Prediction: Semantic similarity is computed between source entity embeddings and target ontology vectors to retrieve alignment candidates. 3. Mapping Refinement: A Large Language Model (LLM) is prompted with each candidate to validate or discard it. 4. Candidate Filtering: Post-processing heuristics are applied to eliminate erroneous or redundant correspondences. Post-processing filters are applied to remove spurious or redundant correspondences. Common heuristics include: Confidence-based filtering, which discards alignments with a similarity score below a given threshold. Cardinality-based filtering, which ensures that mappings are consistent with one-to-one or one-to-many constraints . دمجت أنظمة إدارة العمليات الحالية تقنياتٍ قائمة على المطابقة المعجمية والبنيوية والدلالية، بالإضافة إلى أساليب إصلاح التعيين، مما أدى إلى تحسين الأداء. لا تزال أنظمة إدارة العمليات تواجه تحديين مستمرين: أولاً، لا يزال التمييز بين الكيانات المتشابهة دلاليًا وتلك التي تتواجد بشكل متكرر ولكنها غير مرتبطة مفاهيميًا مهمةً صعبة. ثانيًا، لا تزال قابلية التوسع تُمثل عقبة رئيسية، خاصةً عند تطبيق أساليب إدارة العمليات على مجموعات بيانات كبيرة. ولمعالجة هذه المشكلات، تتطلب العديد من أنظمة إدارة العمليات ضبطًا دقيقًا باستخدام مجموعات بيانات تدريب كبيرة ومحددة المجال لتحقيق الأداء الأمثل. ومع ذلك، فإن هذا النهج ليس عمليًا دائمًا، خاصةً عندما تكون مجموعات البيانات هذه نادرة. يُعد استخدام أساليب قائمة على LLM لإدارة العمليات بديلًا واعدًا. لا تتطلب هذه الأساليب ضبطًا دقيقًا، ويمكنها تحديد التطابقات الدلالية بين الأنطولوجيات بشكل مباشر. غالبًا ما تعتمد الأنظمة الحديثة المستندة إلى LLM على نهج الاسترجاع ثم المطالبة: يسترجع النظام أولاً مجموعة أصغر من الكيانات المستهدفة ذات الصلة (k) ثم يطالب LLM بالتنبؤ بالمراسلات بينها. على الرغم من أن هذا النهج يوفر حلاً أكثر فعالية، إلا أنه لا يزال يواجه العديد من التحديات:يظل أداؤه (مثل مقياس F) محدودًا في بعض السيناريوهات.تستمر مشكلات قابلية التوسع نظرًا لطبيعة استعلام أنظمة إدارة التعلم التي تستغرق وقتًا طويلاً، خاصةً عند العمل مع الأنطولوجيات واسعة النطاق في التطبيقات العملية..

[Audio] Hypothesis and Objectives Hypothesis : The integration of zero-shot prompting engineering with logical reasoning will enable more accurate identification and classification of complex, domain-specific entities without requiring large annotated datasets. Primary Objective: To integrate LLM-based techniques and semantic web technologies to enhance the recognition and classification of fine-grained entities within the agricultural domain. This main objective can be broken down into : O1 : Design and development of a framework for the recognition and classification of fine-grained entities in the agriculture domain. O2: Development and implementation of a context- aware query expansion algorithm to predict high- quality synonyms. O3: Design and implementation of an OM algorithm to predict high-quality ontology alignments. 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis Thesis Hypothesis and Objectives Hypothesis : The integration of zero-shot prompting engineering with logical reasoning will enable more accurate identification and classification of complex, domain-specific entities without requiring large annotated datasets. Primary Objective: To integrate LLM-based techniques and semantic web technologies to enhance the recognition and classification of fine-grained entities within the agricultural domain. This main objective can be broken down into the following sub-objectives: O1 : Design and development of a framework for the recognition and classification of fine-grained entities in the agriculture domain. O2: Development and implementation of a context-aware query expansion algorithm to predict high-quality synonyms. O3: Design and implementation of an OM algorithm to predict high-quality ontology alignments..

[Audio] PhD Proposal This research focuses on the implementation of an algorithm based on AgricultureBERT, a public PLM specifically pre-trained with the objective called Masked Language Modeling (MLM) on a very large training corpus in the agriculture domain. The objective is to evaluate its effectiveness in performing fine-grained classification of agricultural texts using AGROVOC concepts, leveraging the advanced capabilities of LMs to address this complex task..

[Audio] Contributions This thesis introduces five significant contributions : C1 C2 AgriAnnotator Framework – A framework for fine-grained text classification using coarsely labeled data. LexLog Method – A novel lexical-logical approach for mapping out-of-KG entities to Knowledge Graph (KG) concepts. C3 C4 C5 Development of MILA, an LLM-based OM method using a novel “retrieve-identify-prompt” pipeline. Design and implementation of a context-aware query expansion algorithm for high-quality synonym prediction. Annotated Agricultural NLP Dataset – A benchmark dataset of 4,054 sentences annotated with 116 AGROVOC concept URIs, focused on animal welfare. 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis Contributions This thesis introduces five significant contributions to fine-grained NLP in the agricultural domain: AgriAnnotator Framework – A framework for fine-grained text classification using coarsely labeled data. LexLog Method – A novel lexical-logical approach for mapping out-of-KG entities to Knowledge Graph (KG) concepts. Annotated Agricultural NLP Dataset – A benchmark dataset of 4,054 sentences linked to 116 AGROVOC concepts, focused on animal welfare. Context-Aware Query Expansion Algorithm – For high-quality synonym prediction. Ontology Matching Algorithm – Development of MILA, an LLM-based approach that employs a retrieve-identify-prompt pipeline and a depth-first search strategy, adaptable across domains..

[Audio] Elsevier logo Elsevier logo Elsevier logo with wordmark Research Publication Journals J1.Taboada, M., Martinez, D., Arideh , M., & Mosquera , R. (2025). Ontology Matching with Large Language Models and Prioritized Depth-First Search. Information Fusion (accepted for publication 19 April 2025, available online 7 May 2025 ). Accessible from: https://www.sciencedirect.com/science/article/pii/S1566253525003276. Journal Impact Factor (JCR 2023): 14.8 Computer Science, Artificial Intelligence (Q1, 4/197) Computer Science, Theory & Methods (Q1, 2/144). Role & Contribution: Third author contributed to some parts of implementation, analysis and writing. International conferences: IC1.Arideh, M., Taboada, M., & Martínez, D. (2023). A New Query Expansion Algorithm for Enriching the AGROVOC Vocabulary. In Proceedings of the Future Technologies Conference (pp. 594-600). Cham: Springer Nature Switzerland. Role & Contribution: First author implemented the algorithm, conducted experiments, design validation and led the writing of the manuscript. IC2.Arideh, M., & Taboada, M. (2023). Synonym-Substitution Algorithms for Enriching the Agrovoc Vocabulary. In International Conference on Disruptive Technologies, Tech Ethics and Artificial Intelligence (pp. 123-134). Cham: Springer Nature Switzerland. Role & Contribution: First author proposed the implemented the algorithm, validated results with AGROVOC data, and authored the full paper. Research Publications In journals, we published the paper "Ontology Matching with Large Language Models and Prioritized Depth-First Search" in Information Fusion (2025). In international conferences, we published "A New Query Expansion Algorithm for Enriching the AGROVOC Vocabulary" and "Synonym-Substitution Algorithms for Enriching the AGROVOC Vocabulary" in Springer (2023). Finally, "Zero-Shot Prompting Engineering for Fine-Grained Annotation in Agriculture" was published in SSRN (2024)..

[Audio] AGRIANNOTATOR. AGRIANNOTATOR. [image].

[Audio] Fine-grained entity prediction by adapting the three basic steps : Step 1: Turning the input text into a cloze prompt, i.e., a text containing gaps to be filled in. [T]. This text is about [MASK] [A]. Step 2: Predicting the top-N highest scoring words.(Feeding the cloze prompt into AgricultureBERT). Step 3: Align the predicted entity to the most suitable AGROVOC concept. Our approach applies three mapping strategies Strategy 1: The named entity is looked up by exact matching (preferred terms plus synonyms). Strategy 2: The named entity is matched to AGROVOC via a new algorithm developed, MILA. Strategy 3: The named entity is classified through the algorithm LexLog. Fine-grained entity selection. For each document, AgriAnnotator builds one or several AGROVOC subgraphs.The answer for each sub-prompt can result in multiple annotations for the same document. To avoid producing many closely related annotations for the same document, it is usual to apply some preference criterion to select the ones that best annotate the document. EXPERIMENTAL SETUP IN FINE-GRAINED CLASSIFICATION The data collected in two steps: Sentence Collection: Focused on animal welfare topics, including housing, feeding, production, and pollution. (Experiments focused on animal welfare, one of the main factors contributing to the sustainability of food production system). 2. Human Annotation: Experts selected the most suitable fine-grained entities using AGROVOC indexing concepts and author keywords. Similarly, PubMed documents were accessed using MeSH indexing terms and author keywords. Our experts identified 116 distinct fine-grained entities. This set represents a substantial proportion—approximately 37%—of all fine-grained entities defined in AGROVOC, demonstrating the dataset’s strong and comprehensive coverage..

[Audio] EXPERIMENTAL SETUP IN FINE-GRAINED CLASSIFICATION At the time of our study, AGROVOC provided a total of 311 fine-grained concepts with 451 different standardized fine-grained labels – named entities. Organized into AGROVOC hierarchies with an average depth of 4.5. The number of nested named entities was 40% of the total number of named entities. Coarse-grained concept Number of fine-grained concepts in the hierarchy Number of named entities in the hierarchy Hierarchy depth Number of nested named entities (%) Housing 42 66 4 28 (42%) Feeding 49 60 5 35 (58%) Production 171 250 5 85 (34%) Pollution 49 75 4 31 (48%) Total 311 451 4-5 179 (40%).

[Audio] . AGRIANNOTATOR. [image].

[Audio] Results and Analysis Perfect Annotation Precision Recall F-Measure 77 80 79 correct Annotation Precision Recall F-Measure 82 86 84 Results and discussion on fine-grained classification: Overall Results: Precision Production Housing Feeding Pollution 75 77 79 79 Recall Production Housing Feeding Pollution 84 74 83 82 F-Measure Production Housing Feeding Pollution 79 75 81 80 Precision Production Housing Feeding Pollution 77 88 83 81 Recall Production Housing Feeding Pollution 86 85 88 83 F-Measure Production Housing Feeding Pollution 81 86 86 82 Figure 2.8 : Precision, recall and F-measure of AgriAnnotator in annotating fine-grained entities on production, housing, feeding and pollution. 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis Menu As shown in Fig. 2.8(a), AgriAnnotator achieves an average F-measure of 79% for perfect annotation across the 113 fine-grained entities corresponding to the four major coarse-grained entities. This result is remarkably promising, considering that it is a zero-shot approach based on a small-scale PLM. Performance further increases to 84% when correct annotations are considered, indicating that the method can predict correct but slightly imprecise annotations. In deeper insights in Figs. 2.8(b) and 2.8(c), shows that the F-measure for correct annotations is consistently higher than for perfect annotations across all entities. Interestingly, the lowest F-measure for perfect annotations occurs in housing entities, whereas it becomes the highest F-measure when considering correct annotations. These results suggest that AgricultureBERT is not able to recognize very specific entities related to housing, although is able to recognize broader terms..

[Audio] Results and Analysis 2. Comparative performance analysis in Perfect Annotation Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 80 65 72 52 73 75 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 43 43 66 67 87 84 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 56 52 69 59 80 79 Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 72 57 75 66 71 77 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 26 22 71 68 66 74 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 38 31 73 67 69 75 Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 64 55 76 70 75 79 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 32 34 83 76 76 82 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 43 42 79 73 76 80 Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 96 68 71 52 70 79 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 21 28 71 66 74 83 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 34 40 71 58 72 81 Figure 2.9: Comparative analysis of precision, recall, and F-measure in perfect annotation for production, housing, feeding, and pollution 26 November 2025 This analysis compared different AgriAnnotator configurations with AgroPortal Annotator for "perfect" fine-grained annotation. The Holistic Prompt and AgroPortal configurations performed the lowest on perfect annotation tasks. This supports the strategy of breaking down a holistic prompt into smaller sub-prompts, each tailored to a specific relevant sentence, to enhance zero-shot performance. All zero-shot configurations, except the holistic approach, outperform the AgroPortal Annotator, showing that simple, entity-centered prompts effectively identify fine-grained entities for accurate text annotation. AgriAnnotator – Exact Mapping substantially improves results, particularly for feeding and production, revealing discrepancies in terminology granularity between AgricultureBERT and AGROVOC. The AgricultureBERT-based mapping configuration showed weaker performance, indicating that the prompt requires additional fine-tuning to classify entities effectively. AgriAnnotator – Masked Word Selection by PLM Score performed lower than other AgriAnnotator configurations, except in production. This suggests that fine-grained concepts are typically located on the leaves or near the leaves of AGROVOC, whereas the PLM tends to predict coarser-grained concepts..

[Audio] 3. Comparative performance analysis in Correct Annotation Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 86 80 63 54 76 77 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 48 55 71 69 90 86 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 62 65 67 60 82 81 Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 95 90 88 81 85 88 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 37 40 86 86 80 85 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 53 55 87 83 82 86 Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 100 94 82 62 85 83 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 21 42 82 79 89 88 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 35 58 82 69 87 86 Precision Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 95 88 77 73 80 81 Recall Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 53 65 84 78 81 83 F-Measure Agroportal AgriAnnotator- Holistic Prompt AgriAnnotator- Exact Mapping AgriAnnotator-AgricultureBert Based Mapping AgriAnnotator- Masked word selection by PLM score AgriAnnotator 68 75 80 75 80 82 Figure 2.10: Comparative analysis of precision, recall, and F-measure in correct annotation for production, housing, feeding, and pollution. For "Correct Annotation," all models showed improved F-measures. Zero-shot prompting configurations demonstrated consistent performance, indicating that AgricultureBERT effectively refines coarse-grained labels. These results validate that carefully designed prompts, combined with domain-specific PLMs and the AGROVOC ontology, can effectively support zero-shot fine-grained annotation..

[Audio] AGRIANNOTATOR The overall experimental results suggests that AgricultureBERT manages to refine, to some extent, coarse-grained labels. This is largely due to the specific way AgriAnnotator constructs its prompts. Consequently, AgricultureBERT captures nuanced distinctions within categories, thereby enhancing precision and reducing ambiguity. Analysis of errors Discrepancies in terminology granularity between AgricultureBERT and AGROVOC. Two primary issues are discerned: Incompleteness of AGROVOC for some entities predicted by AgricultureBERT. Limited prediction capacity of AgricultureBERT for long-tail entities present in AGROVOC. AgricultureBERT's inability to recognize some variants of fine-grained entities, predominantly attributed to regional idiosyncrasies in dialects. AgricultureBERT's prediction errors. AgricultureBERT sometimes mispredicts named entities due to the presence of high correlations Categorization errors in AGROVOC: This mistake leads to an important decrease in the F-measure of our method. 5. Limited effectiveness of answer mapping. Three primary issues are discerned: Selection of a single fine-grained annotation per sentence. so the opportunity to capture all possible accurate results per sentence is lost. Lower scores for more accurate predictions. AgricultureBERT predicts broader entities with higher scores than the more accurate fine-grained entities. Same labels with different meanings (depending upon underlying data sources), ambiguous label meanings led to misclassifications (e.g., "social housing" ambiguity). Future work : A first challenge of zero-shot prompting engineering is developing suitable strategies for generating effective prompts. A second challenge of zero-shot prompting engineering is effectively capture the long-tail entities, i.e., entities that appear in training texts rarely. A first solution is fine-tuning pretrained models via few-shot examples, which improves performance in low-resource settings. A second promising solution is based on retrieving information from external resources to feed the MLM with relevant textual contexts ..

[Audio] A Lexical-Logical Inference Algorithm Knowledge graphs often suffer from incompleteness due to: 1. Lacking terms , Missing entities, and lacking relationships between entities. To address this, innovative algorithms are needed to classify a new out-of-KG entities and relationships. We propose LexLog, is an algorithm developed to classify predicted entities that are not explicitly present in AGROVOC. By analyzing both the lexical features of concept names and the semantic hierarchy within AGROVOC. LexLog Algorithm KG.png KGs Suffer From Incompleteness due to : 1. Lacking terms 2. Missing entities 3. Lacking relationships between entities. C2 LexLog We proposed Lexlog, an algorithm that can classify new out-of-KG entities and relationships. To address this, we need an innovative algorithm Is-a Forage production swine Pigs piglets sows Livestock Livestock production Feed production Production Is-a Forage Feed silage roughage Is-a Piglet production 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis Menu Knowledge graphs often suffer from incompleteness due to: missing terms, absent entities, and lacking relationships. To address this, innovative algorithms are needed to classify a new out-of-KG entities and relationships. We propose LexLog, is an algorithm developed to classify predicted entities that are not explicitly present in AGROVOC. By analyzing both the lexical features of concept names and the semantic hierarchy within AGROVOC..

[Audio] LEXLOG: A LEXICAL-LOGICAL INFERENCE ALGORITHM LexLog implements a new algorithm for entity classification that leverages substitution of nested entities by broader entities. The method begins by identifying a nested entity within the predicted entity. Then, LexLog navigates upward through the AGROVOC hierarchy, replacing the nested entity with its broader counterpart until it reaches the out-of-AGROVOC entity. Therefore, by leveraging the potential of broader entity substitution, our algorithm can classify missing entities in entities already asserted in AGROVOC. LexLog implements a new algorithm for entity classification that leverages substitution of nested entities by broader entities. LexLog iteratively moves upward through the AGROVOC hierarchy, replacing the nested entity with its broader parent concept until it encounters an entity not covered by AGROVOC. #How It Works: 1. Identify Nested Terms: Like : “Forage production” → nested term: “Forage “ 2. Navigate the Hierarchy: Traverses AGROVOC's hierarchy to find broader, semantically related terms. “Forage” → “Feed” 3. Substitute and Classify: Substitutes the nested term with its broader counterpart until it matches an AGROVOC concept. “Forage production” → “Feed production” Even if a term like "Forage production" isn't in AGROVOC, LexLog infers that it belongs under "Feed production", enabling fine-grained recognition and better concept alignment..

[Audio] A new algorithm to expand ontology terminology The primary objective of this algorithm is to integrate recent advancements in NLP into synonym-substitution strategies. The aim is to enhance the recognition and normalization of named entities while mitigating the over-generation and noise typically associated with traditional synonym-expansion methods. Our algorithm performs synonym substitution by identifying shared lexical substrings between a given term and its hierarchical parent within the ontology, and replacing these substrings with semantically equivalent alternatives. The validation step uses the AgricultureBERT language model. Specifically, it computes the cosine similarity between the contextual embeddings of AGROVOC terms and their corresponding synonym candidates, filtering out those below a defined semantic similarity threshold. The proposed method operates in two primary stages: Synonym Candidate Generation : The proposed method begins by recursively identifying all AGROVOC terms. The algorithm then systematically replaces lexical overlaps in descendant terms with synonymous expressions from their corresponding parent terms, thereby generating synthetic synonym candidates. The synonym substitution process is not limited to direct descendants; it also propagates recursively through the entire hierarchical sub-tree rooted at the overlapping term. Synonym Validation: The method executes two sequential steps: 1. Filtering of Missing Candidates. An exact search is performed in the AGRIS repository to verify each candidate. Candidates with no matching results are discarded. 2. Prediction of Valid Synonyms. The method leverages the AgricultureBERT model to predict synonyms most similar to existing AGROVOC terms. First, Keytotext is used to automatically generate a sentence for each original AGROVOC term. A second sentence is then generated for each synonym candidate by substituting the original term with the candidate. The cosine similarity between the encoding vectors of the original term and each candidate is computed, and only candidates with high similarity are retained. For this application, we empirically determined a threshold of 0.98..

[Audio] Data preprocessing The AGROVOC ontology was extracted using the Skosmos API (Through a series of HTTP requests) , the ontology data was retrieved and converted into structured JSON objects. These objects were then stored in relational SQL tables(serve as the input format for the synonym expansion algorithm). Data collection We utilized the AGROVOC ontology dataset published on May 25, 2021. Comprised a total of 38,986 concepts in English. 31,818 were identified as leaf concepts, meaning they do not have any narrower terms within the hierarchical structure. The remaining 7,168 concepts represent non-leaf node. These interest for the synonym expansion algorithm, as they: Serve as structural anchors for logical and semantic relationships. Basis for identifying potential logical overlaps. The algorithm can analyze both the explicit hierarchical links and the implicit semantic relationships expressed through synonyms (entry terms). Evaluation procedure: To assess the performance, we employed precision as the primary evaluation metric. Precision, a standard measure in information retrieval and machine learning, is particularly suited for evaluating tasks that involve generating sets of candidate terms, such as synonyms. The expert reviewed the expanded synonym sets and categorizing them into relevant or irrelevant based on their alignment with AGROVOC's intended concepts. The precision metric was then calculated by dividing the number of relevant synonyms by the total number of synonyms generated..

[Audio] Results and discussion on synonym expansion lexical overlaping.png AGROVOC 38,986 concepts Lexical Overlaps 3,627 textual overlaps (51.6% of non-leaf nodes), Generating 4,439 initial candidates. 1,902 New Terms By sentence Similarity , (92%) Precision AgricultureBERT Semantic filtering 79% (1496) appear in recent literature 21% might be old Check Against PubMed/AGRIS At the time of experimentation, the AGROVOC thesaurus comprised a total of 38,986 concepts, of which 31,818 were leaf nodes and 7,168 were non-leaf nodes. The algorithm was designed to operate only on narrower concepts (i.e., child nodes) associated with non-leaf nodes( identifying 3,627 overlaps (51.6% of non-leaf nodes) and generating 4,439 initial candidates. After semantic filtering with AgricultureBERT, the list was refined to 1,902 high-quality synthetic synonyms with 92% precision. A cross-check against PubMed and AGRIS showed that 79% (1,496) of the generated synonyms appear in recent literature, confirming their contemporary relevance, while 21%(406) may represent older yet potentially valuable terms. Discussion : The algorithm achieves : High rate of lexical overlaps between child and parent: Serves as a strong foundation for synonym prediction. Multiple synonym candidates from a single overlap: Enhances vocabulary enrichment. Semantic similarity with the original terms: Captures deeper conceptual associations beyond surface-level matches. Validation results: 79% of generated synonyms have been used in the past decade, reinforcing their relevance for modern information retrieval. Unique generated synonyms: Help identify gaps and enrich vocabulary. 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis Menu at the time of experimentation , the synonym expansion study targeted AGROVOC's 38,986 concepts, identifying 3,627 overlaps (51.6% of non-leaf nodes) and generating 4,439 initial candidates. After semantic filtering with AgricultureBERT, the list was refined to 1,902 high-quality synthetic synonyms with 92% precision. A cross-check against PubMed and AGRIS showed that 79% (1,496) of the generated synonyms appear in recent literature, confirming their contemporary relevance, while 21% may represent older yet potentially valuable terms. The algorithm achieves a: High rate of lexical overlaps between child and parent: Serves as a strong foundation for synonym prediction. Multiple synonym candidates from a single overlap: Enhances vocabulary enrichment. Semantic similarity with the original terms: Captures deeper conceptual associations beyond surface-level matches. Validation results: 79% of generated synonyms have been used in the past decade, reinforcing their relevance for modern information retrieval. Unique generated synonyms: Help identify gaps and enrich vocabulary..

[Audio] Results and discussion on synonym expansion Contributions : Builds on Taboada’s strategy by incorporating language model (LM) validation using AgricultureBERT. Our approach enforces two strict constraints: 1) Only allow substitutions among hierarchically related terms. 2) Semantic confirmation by a PLM required for all substitutions. These constraints ensure high precision and contextual relevance in the generated synonyms. Our proposal is to integrate modern LM techniques that offer both interpretability and scalability, while minimizing the risk of introducing noise into the ontology. Our approaches : Avoid overgeneration, semantic drift, and reliance on vague synonym chains. The synonym expansion method has three main limitations: Restricted to multi-word terms: Single-word synonym generation is not supported. Potential solution: Use prompt engineering with domain-specific LMs (e.g., AgricultureBERT, GPT, LLaMA) to infer synonyms for single-word terms via natural language prompts. Reduced recall due to strict embedding thresholds: Potential solution: Employ prompt-based synonym generation to capture semantically relevant terms that embeddings might miss. Limited to AGROVOC terminology: Potential solution: Integrate open-vocabulary resources, such as AgricultureBERT’s internal terminology or embeddings. Mohammad.I.Arideh , Doctoral Thesis Contributions : Builds on Taboada's strategy but adds language model (LM) validation using AgricultureBERT. Introduces strict constraints: Substitutions only among hierarchically related terms. Substitutions must be semantically confirmed by a PLM. Avoids common pitfalls like overgeneration, semantic drift, or reliance on vague synonym chains. The synonym expansion method has three main limitations: Restricted to multi-word terms: Single-word synonym generation is not supported. Potential solution: Use prompt engineering with domain-specific LMs (e.g., AgricultureBERT, GPT, LLaMA) to infer synonyms for single-word terms via natural language prompts. Reduced recall due to strict embedding thresholds (While the use of embedding-based techniques (e.g., SBERT) contributes to high precision in synonym generation, it comes at the cost of reduced recall.) Potential solution: Employ prompt-based synonym generation to capture semantically relevant terms that embeddings might miss. Limited to AGROVOC terminology: Potential solution: Integrate open-vocabulary resources, such as AgricultureBERT's internal terminology or embeddings..

[Audio] MILA (Ontology Matching Algorithms) OM refers to the process of identifying semantic correspondences between entities across multiple ontologies . OM can be defined by a function that takes as input a source ontology Os, a target ontology Ot, an input alignment A, a set of parameters p and resources r (such as external background knowledge), and return an alignment A’ . OM systems continue to have difficulty distinguishing between entities that are semantically similar and those that are merely frequently co-occurring . Large Language Model (LLM) emerged as a promising method for OM. They leverage pre-trained knowledge for finding correspondences across ontologies and do not require fine-tuning. OM technology requires the use of external background knowledge to work effectively, as most ontologies are designed in specific contexts that are not explicitly modeled. The State-of-the-art LLM-based systems apply a retrieve-then-prompt pipeline. In this method, target entities are first retrieved and then used to prompt an LLM to predict mapping correspondences. These systems still face challenges, with limited performance in terms of F-Measure. They face scalability issues due to the long execution times required to query the LLM, particularly when applied to large-scale ontologies. This underscores the need for al-ternative RAG strategies that improve F-Measure performance while minimizing the number of requests to the LLM ..

[Audio] MILA, a novel approach that embeds a retrieve-identify-prompt pipeline within a prioritized depth-first search (PDFS) strategy. (MInimizing LLM Prompts in Ontology MApping). MILA aims to find high-confidence, bidirectional correspondences between ontologies. This approach efficiently identifies a large number of semantic correspondences with high accuracy, limiting LLM requests to only the most borderline cases. MILA introduces a novel retrieve-identify-prompt pipeline, which adds an intermediate step to identify high-confidence bidirectional (HCB) correspondences with high precision. Prioritized Depth-First Search (PDFS) combines elements of both Depth-First Search (DFS) and Greedy Best-First Search (GBFS). MILA applies a prioritized depth-first search (PDFS) strategy, which iteratively queries the LLM until a definitive match is confirmed. This approach also minimizes the number of queries by ending early when a valid match is found. EXPERIMENTAL SETUP IN OM To evaluate MILA, we selected a domain beyond agriculture by using the biomedical benchmark introduced in the 2023 and 2024 editions of the Ontology Alignment Evaluation Initiative (OEAI). The biomedical setting was chosen for several reasons: it has high practical relevance, is well-suited for analysis by PLMs and LLMs, and provides extensive datasets for ontology matching. All OM systems were evaluated using the traditional information retrieval metrics of Precision (P), Recall (R) and F-Measure. Precision reflects the correctness of an alignment achieved by an OM system with respect to a reference alignment . It is usually defined as the ratio between the number of matches correctly detected by the OM system and the total number of matches identified by the OM system. Recall reflects the completeness of an alignment achieved by an OM system with respect to a reference alignment . It is usually defined as the ratio of the number of matches correctly detected by the OM system to the total number of matches identified in the reference alignment. F-Measure combines P and R in a unique measurement..

[Audio] Dataset and Tasks for Evaluating MILA The experiments used five biomedical ontology matching (OM) tasks, involved six ontologies : (OMIM, ORDO, SNOMED CT, FMA, DOID, NCIT). Two Test Sets: Unsupervised: Full reference mappings; no training data required. Semi-supervised: 70% of mappings used for evaluation, 30% held out for training. As an unsupervised system, MILA’s performance was tested on both the unsupervised and semi-supervised settings. For all OM tasks, MILA integrates the SBERT retriever model (multi-qa-distilbert-cos-v1. )with the LLaMA-3.1 (70B) language model to assess ontology matching performance. To ensure high-quality results, candidate mappings retrieved by SBERT with a confidence score below 0.75 were discarded. Finally, mapping refine-ment is mainly based on heuristic and logical reasoning, with the exception of Olala, LLMs4OM and MILA, which leverage LLMs for mapping refinement. Compared to other current approaches, MILA’s main distinction is its use of a PDFS strategy, supported by LLMs, to solve mapping refinement..

[Audio] MILA was evaluated using three challenges from the 2024 edition of the Ontology Alignment Evaluation Initiative. The results of our evaluation demonstrate the ability of MILA to outperform state-of-the-art OM systems in terms of task-agnostic and high performance in terms of F-measure. Our method achieved the highest F-measure in five of the seven unsupervised tasks, outperforming state-of-the-art OM systems by up to 17%. It also performed comparably to—or better than—the leading supervised OM systems. Furthermore, MILA demonstrated task-agnostic performance, remaining stable across all evaluation settings. For each source entities MILA generates the set of the most promising target entities, and the corresponding ordered sequence of elements. A bidirectional correspondence between entities exists if both entities must appear in each other’s candidate sets, allowing traversal in both directions. A high confidence bidirectional (HCB) correspondence between both entities exists if they are the most prioritized entities in each other’s correspondence candidate sets. MILA uses a simple structured prompt with minimal ontology context to maintain clarity and focus. The prompt includes only the ontology names and the preferred labels of the source and target entities involved in the correspondence..

[Audio] MILA algorithm: the algorithm developed to perform semantic mapping of predicted entities to AGROVOC concepts. For each source entity 𝑒𝑆, the algorithm first retrieves the ordered sequence of target candidates 𝐿𝑒𝑆. Then iterates over this sequence using a PDFS strategy, looking for the first target entity 𝑒𝑇∈𝐿𝑒𝑆 that is a valid match to 𝑒𝑆. (Specifically, for each pair (𝑒𝑆,𝑒𝑇), the algorithm checks whether the pair is a bidirectional correspondence). If so, it further verifies whether the pair is an HCB correspondence. If the pair is not an HCB correspondence, the algorithm queries an LLM to confirm the potential mapping. If the LLM confirms the mapping, it is added to the final mapping, and the search stops for that source entity. If the LLM does not confirm the potential mapping, the search continues with the next candidate in 𝐿𝑒𝑆, iterating until a valid mapping is identified or all candidates are exhausted. The prompt template is as follows:This step is performed exclusively on low-confidence bidirectional correspondences involving the initial node. An LLM is used to confirm or reject the correspondence. You are a helpful expert in ontology matching, which involves determining equivalence correspondences between concepts from different ontologies. The source ontology is called [𝑠𝑟𝑐_𝑜𝑛𝑡𝑜_𝑛𝑎𝑚𝑒]and the target ontology is called [𝑡𝑔𝑡_𝑜𝑛𝑡𝑜_𝑛𝑎𝑚𝑒]. Classify whether the following concepts are equivalent: Source concept: [𝑠𝑜𝑢𝑟𝑐𝑒_𝑒𝑛𝑡𝑖𝑡𝑦] Target concept: [𝑡𝑎𝑟𝑔𝑒𝑡_𝑒𝑛𝑡𝑖𝑡𝑦] If so, answer ’Yes’, without adding any type of explanation. Other-wise, answer ’No’..

[Audio] MILA The framework comprises the following steps: simple lexical pre-processing : Pre-processing includes the use of lowercase and any processing required to index the labels in the vector KB 2) building of two vector KBs, one per ontology: Using a embedding model EM, such as SBERT, the encoder maps each processed label w to a vector representation, we used full labels to encode the vector KBs. we do not encode contexts in these KBs. 3) prediction and refinement of bidirectional correspondences: Given a search graph, an initial state, and a goal test, the algorithm searches for a solution by generating a search tree. The search objective is to find the most promising bidirectional correspondences between two entities. To achieve this, for each source entity (initial node), MILA applies the ‘retrieve-identify-prompt’ pipeline at each step of the PDFS algorithm. The search goal is to find bidirectional correspondences in some path from any node by applying the following steps in each iteration: Retrieve: The algorithm uses two retrievers to extract prioritized lists of successor nodes by encoding the ontology terminologies into two separate vector databases—one for each ontology. selecting the top k highest-scoring nodes ranked according to cosine similarity. Identify: An objective function evaluates whether a bidirectional correspondence can be identified between the node's predecessor and the node itself, and whether high confidence. The PDFS strategy stops the iteration when the initial node is part of some identified high-confidence bidirectional mapping. Prompt: If a high-confidence bidirectional correspondence cannot be asserted between two nodes, then an LLM is prompted to confirm it..

[Audio] MILA A bidirectional correspondence is identified between two nodes ni and nj if the following two conditions are satisfied: Forward correspondence: When node ni is expanded, node nj is in the list of the generated nodes. Backward correspondence: When node nj is expanded, node ni is in the list of the generated nodes. A high-confidence bidirectional correspondence is identified between two nodes ni and nj if the following two stricter conditions are satisfied: First-ranked node for ni : When node ni is expanded, node nj is the first node generated (i.e., the node with the highest similarity value). First-ranked node for nj: When node nj is expanded, node ni is the first node generated. MILA allows mapping scores to be rounded to a specified number of decimal places, based on an input parameter. After rounding, multiple nodes may be assigned identical priority scores, causing them to be ranked with the same priority and evaluated uniformly. Moreover, two alternative configurations were implemented in MILA: Configuration 1: In this configuration, high-confidence bidirectional correspondences can be identified between any two nodes within the search tree. If a high-confidence bidirectional correspondence is identified between a pair of nodes that do not include the initial node, all bidirectional correspondences between either of these nodes and the initial node are automatically excluded. If a bidirectional correspondence is identified between the initial node and any LLM for further processing. Furthermore, if the reverse correspondence is not recovered, then it is explicitly added at the end of the successor list (see Appendix B). Configuration 2: In this configuration, bidirectional correspondences between pairs of nodes that do not involve the initial node are not evaluated. Therefore, if a bidirectional correspondence is identified between the initial node and any other node, and this correspondence is not identified as a high-confidence mapping, then it is forwarded to the LLM for additional processing. This approach is more efficient for tasks that require a single mapping, as it eliminates unnecessary exploration. Both configurations use a multi-path pruning strategy that, once a path to a node is found, discards all other paths to the same node..

[Audio] MILA (Example 1) MILA con1.png MILA confirms the HCB correspondence and adds it to the set of final mappings. Example 1: Figure 5.5 illustrates the complete application of the “retrieve-identify-prompt” pipeline embedded within the PDFS algorithm using Configuration 1. Step 1 — Retrieve: For source entity, MILA retrieves an ordered sequence of target candidates. Step 2 — Retrieve: It then retrieves an ordered sequence of source candidates for the highest-ranked target candidate. Step 3 — Identify bidirectional correspondence: MILA checks whether a bidirectional correspondence exists between the highest-ranked candidate (DOID:438) and the source entity. Step 4 — Identify HCB correspondence: Since a bidirectional correspondence is found, MILA verifies whether it qualifies as a high-confidence bidirectional (HCB) correspondence. As ncit:C99383 and DOID:438 are each other’s top-ranking candidates, MILA confirms the HCB correspondence and adds it to the set of final mappings..

[Audio] MILA (Example 2) Figure 5.6: Example of Managing Bidirectional Correspondences in Configuration 2: Step 1 — Retrieve: For source entity, MILA retrieves an ordered sequence of target candidates.𝐿𝑛𝑐𝑖𝑡∶𝐶3745. • Step 2 — Retrieve: MILA retrieves the ordered sequence of source candidates for the most promising target entity, 𝐷𝑂𝐼𝐷∶4880. •Step 3 — Identify bidirectional correspondence: Next, MILA checks whether there is a bidirectional correspondence between ncit:C3745 and DOID:4880. •Step 4 — Identify HCB correspondence: As there is a bidirectional correspondence, it checks whether this is an HCB correspondence, by verifying whether ncit:C3745 is the top-ranked candidate for DOID:4880. Step 5 — Prompt: Since there is no bidirectional correspondence between ncit:C3745 and DOID:4880, MILA prompts the LLM to confirm the match. Since the LLM does not confirm the match between ncit:C3745 and DOID:4880, MILA proceeds to the next most similar candidate, DOID:4233 , and repeats the process. •Steps 6 and 7 — Retrieve and Identify bidirectional correspondence: MILA retrieves the ordered sequence of source candidates for the second most promising target entity, DOID:4233, and then verifies that the pair (ncit:C3745, DOID:4233) is a bidirectional correspondence, but not an HCB correspondence. Step 8 — Prompt: MILA prompts the LLM to confirm the pair (ncit:C3745, DOID:4233). This time, the LLM confirms the match, allowing MILA to finalize the alignment and return the valid correspondence..

[Audio] MILA - Results This section evaluates MILA’s ontology mapping (OM) performance against 14 state-of-the-art systems using the OAEI BIO-ML 2023–2024 benchmark, highlighting both its accuracy and efficiency. The comparison serves two purposes: Benchmark MILA’s performance and design against leading systems. Highlight its strengths in mapping accuracy and computational efficiency. The compared systems fall into three categories: Textual Similarity-Based: Like MILA, using PLMs for vector comparisons. Structure-Aware: Incorporating hierarchical or logical constraints. Domain-Specific: Embedding specialized knowledge for specific domains. This benchmarking positions MILA among the leading OM approaches. Traditional systems predominantly employ heuristic filters or logical reasoning modules for refinement. However, more recent approaches, including OLaLa, LLMs4OM, and MILA, leverage the capabilities of LLMs to perform this refinement. These models enable a deeper understanding of semantic similarity by capturing complex contextual information that traditional rule-based or embedding-only methods often overlook. MILA distinguishes itself from existing approaches by introducing a novel retrieve–identify–prompt pipeline within a Retrieval-Augmented Generation framework. Using a Retrieval-Augmented Generation (RAG) approach, this strategy enables MILA to efficiently identify a large number of semantic correspondences with high accuracy, while limiting LLM calls to only the most borderline cases. 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis Menu This section evaluates MILA's ontology mapping (OM) performance against 14 state-of-the-art systems using the OAEI BIO-ML 2023–2024 benchmark, highlighting both its accuracy and efficiency. The comparison serves two purposes: Benchmark MILA's performance and design against leading systems. Highlight its strengths in mapping accuracy and computational efficiency. The compared systems fall into three categories: Textual Similarity-Based: Like MILA, using PLMs for vector comparisons. Structure-Aware: Incorporating hierarchical or logical constraints. Domain-Specific: Embedding specialized knowledge for specific domains. This benchmarking positions MILA among the leading OM approaches. Traditional systems predominantly employ heuristic filters or logical reasoning modules for refinement. However, more recent approaches, including OLaLa, LLMs4OM, and MILA, leverage the capabilities of LLMs to perform this refinement. These models enable a deeper understanding of semantic similarity by capturing complex contextual information that traditional rule-based or embedding-only methods often overlook. MILA distinguishes itself from existing approaches by introducing a novel retrieve–identify–prompt pipeline within a Retrieval-Augmented Generation framework. Using a Retrieval-Augmented Generation (RAG) approach, this strategy enables MILA to efficiently identify a large number of semantic correspondences with high accuracy, while limiting LLM calls to only the most borderline cases..

[Audio] Results in the unsupervised setting MILA is the best performing algorithm in four of the five OM tasks. In the OMIM-ORDO mapping task, MILA’s outperforms the second best OM system in this task, LogMapBio, by 17%. Specifically, MILA achieves high recall. In the NCIT-DOID mapping task, MILA outperforms the second best baseline, HybridOM, in terms of F-Measure by 3%. In the SNOMED-FMA mapping task, It outperforms the second-best OM systems in this task, BERTMap and HybridOM, by 13%. Again, MILA achieves a high recall, compared to the rest of the approaches. MILA is the best performing algorithm in four of the five OM tasks. In the OMIM-ORDO mapping task, MILA's outperforms the second best OM system in this task, LogMapBio, by 17%. Specifically, MILA achieves high recall. In the NCIT-DOID mapping task, MILA outperforms the second best baseline, HybridOM, in terms of F-Measure by 3%. In the SNOMED-FMA mapping task, It outperforms the second-best OM systems in this task, BERTMap and HybridOM, by 13%. Again, MILA achieves a high recall, compared to the rest of the approaches. In the SNOMED-NCIT (Pharmacology) mapping task, MILA is the second-best approach with 4% below the outstanding approach, HybridOM, in terms of F-Measure. Note that MILA out-performs the third baseline (i.e. AMD) in terms of F-Measure by 9%. In the SNOMED-NCIT (Neoplasm) mapping task, It outperforms the second best OM system in this task, LogMapBio, by 17%. As in the other tasks, MILA achieves high recall, compared to the rest of the approaches. This strong recall performance highlights MILA's ability to correctly identify a large proportion of relevant instances, ensuring a more comprehensive and accurate mapping process. Additionally, MILA's remarkable precision ensures that it maintains a balance between high recall and avoiding false positives, which is key to its outstanding F1 score. This consistent level of high performance across various metrics underscores MILA's robustness and its ability to handle complex tasks..

[Audio] MILA - Results Results in the unsupervised setting In the SNOMED-NCIT (Pharmacology) mapping task, MILA is the second-best approach with 4% below the outstanding approach, HybridOM, in terms of F-Measure. Note that MILA out-performs the third baseline (i.e. AMD) in terms of F-Measure by 9%. In the SNOMED-NCIT (Neoplasm) mapping task, It outperforms the second best OM system in this task, LogMapBio, by 17%. As in the other tasks, MILA achieves high recall, compared to the rest of the approaches. This strong recall performance highlights MILA’s ability to correctly identify a large proportion of relevant instances, ensuring a more comprehensive and accurate mapping process. Additionally, MILA’s remarkable precision ensures that it maintains a balance between high recall and avoiding false positives, which is key to its outstanding F1 score. This consistent level of high performance across various metrics underscores MILA's robustness and its ability to handle complex tasks. MILA is the best performing algorithm in four of the five OM tasks. In the OMIM-ORDO mapping task, MILA's outperforms the second best OM system in this task, LogMapBio, by 17%. Specifically, MILA achieves high recall. In the NCIT-DOID mapping task, MILA outperforms the second best baseline, HybridOM, in terms of F-Measure by 3%. In the SNOMED-FMA mapping task, It outperforms the second-best OM systems in this task, BERTMap and HybridOM, by 13%. Again, MILA achieves a high recall, compared to the rest of the approaches. In the SNOMED-NCIT (Pharmacology) mapping task, MILA is the second-best approach with 4% below the outstanding approach, HybridOM, in terms of F-Measure. Note that MILA out-performs the third baseline (i.e. AMD) in terms of F-Measure by 9%. In the SNOMED-NCIT (Neoplasm) mapping task, It outperforms the second best OM system in this task, LogMapBio, by 17%. As in the other tasks, MILA achieves high recall, compared to the rest of the approaches. This strong recall performance highlights MILA's ability to correctly identify a large proportion of relevant instances, ensuring a more comprehensive and accurate mapping process. Additionally, MILA's remarkable precision ensures that it maintains a balance between high recall and avoiding false positives, which is key to its outstanding F1 score. This consistent level of high performance across various metrics underscores MILA's robustness and its ability to handle complex tasks..

[Audio] MILA - Results Results in the semi-supervised setting While MILA is primarily an unsupervised system, we also evaluate its performance in comparison to systems that utilize training data to enhance their results. In the semi-supervised setting, MILA is the best performing algorithm in two of the five ontology mapping tasks . In the NCIT-DOID task, MILA outperforms the second-best OM system in this task, BioGITOM, by 6%, achieving an F-Measure of 0.97. In the SNOMED-NCIT task, it outperforms the second-best OM system, Matcha-DL, by 18%..

[Audio] MILA - Results Results in the semi-supervised setting MILA is the second or third best approach with an F-Measure between 3% and 4% below the leading approaches. Please note that no information is available for LLMs4OM, and therefore it does not appear in the tables. MILA, although developed for unsupervised tasks, frequently outperforms or closely matches the performance of supervised models that depend on training data. Experimental results on the anatomy and biodiversity evaluation bench-marks MILA achieved the second-best performance, with 2% below the outstanding approach, Matcha, in terms of the F-Measure. The results, indicate that MILA outperforms all other approaches in this benchmark. In particular, it outperforms LogMap, the second best system, by 17%..

[Audio] Results And Analysis Results and discussion on Ontology Mapping Evaluation of Computational Efficiency To assess the computational efficiency of MILA compared to baseline RAG-based OM systems, Table presents the number of LLM requests required by each approach. MILA adopts a more efficient strategy by querying the LLM only in ambiguous cases, as determined by the search algorithm, significantly reducing the number of requests. As shown, MILA achieves remarkable computational savings, with LLM requests reduced by more than 92% in all cases. This reduction in LLM queries not only improves the computational efficiency of MILA but also helps mitigate excessive computational overhead. Task Number of LLM runs in baseline RAG Number of LLM runs in MILA Percentage of decrease in MILA OMIM-ORDO 18,605 1,015 94.5% NCIT-DOID 23,430 375 98.4% SNOMED-FMA 36,280 1,014 97.2% SNOMED-NCIT (Pharm) 29,015 2,135 92.6% SNOMED-NCIT (Neoplasm) 19,020 570 97% 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis To assess the computational efficiency of MILA compared to baseline RAG-based OM systems, Table presents the number of LLM requests required by each approach. MILA adopts a more efficient strategy by querying the LLM only in ambiguous cases, as determined by the search algorithm, significantly reducing the number of requests. As shown, MILA achieves remarkable computational savings, with LLM requests reduced by more than 92% in all cases. This ability to minimize LLM queries without sacrificing performance makes MILA a more scalable and cost-effective solution for complex OM tasks. This reduction in LLM queries not only improves the computational efficiency of MILA but also helps mitigate excessive computational overhead..

[Audio] MILA - Results Discussion: OM plays a key role in enabling seamless communication between diverse and heterogeneous applications, ensuring data interoperability and effective integration across various domains. Our approach, MILA, demonstrates substantial improvements over state-of-the-art OM systems. First, our study reveals that MILA achieved the highest F-Measure score in four of the five unsupervised ontology matching tasks evaluated. Second, Even though MILA is designed as an unsupervised system, it still emerged as the best-performing algorithm in two of the five ontology mapping tasks in the semi-supervised setting. Third, MILA’s approach is task-agnostic, exhibiting stable performance across all tasks and settings (unsupervised or semi-supervised). Fourth :Efficient: MILA uses LLMs only when necessary, reducing LLM calls by over 92% and addressing scalability concerns. MILA effectively combines RAG techniques with advanced search strategies, resulting in superior performance. MILA introduces an innovative RAG-based approach that optimizes the search space, improving both precision and recall, while minimizing unnecessary LLM prompts. 26 November 2025 Mohammad.I.Arideh , Doctoral Thesis.