Recent advances in ai have spurred interest in NeSy architectures that integrate neural and symbolic methods. In particular, combining a Constraint Programming (cp) model with a language model for constrained sequence generation tasks allows the neural component to capture domain knowledge while cp enforces structural constraints. In this paper we propose combining cp with a Masked Language Model (mlm) to perform Large Neighbourhood Search (lns). Unlike conventional left-to-right Large Language Models, mlm s can complete sequences with gaps in arbitrary positions, making them well-suited for this task. Meanwhile, lns provides a cp-based iterative framework to explore constrained subspaces whenever searching the whole space would be intractable. We evaluate NeSylns on tasks in constrained text generation and molecule discovery. Our experiments show that it can quickly generate many high-quality sentences and molecules, even for highly-constrained tasks.

Megh Thakkar, Quentin Fournier, Matthew Riemer, Pin-Yu Chen, Amal Zouaq, Payel Das, Sarath Chandar. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

A recent renewal in interest in long text understanding has sparked the emergence of high-quality long text benchmarks, as well as new models demonstrating significant performance improvements on these benchmarks. However, gauging the implication of these advancements based solely on the length of the input text offers limited insight. Such benchmarks may require models to parse long-range dependencies or merely to locate and comprehend the relevant paragraph within a longer text. This work introduces the Minimal Viable Phrase (MVP), a novel metric that determines, through perturbations to the input text, the shortest average text length that needs to be preserved to execute the task with limited performance degradation. Our evaluation of the popular SCROLLS benchmark reveals that only one of its seven tasks necessitates an MVP of over 512 tokens–the maximum text length manageable by the previous generation of pre-trained models. We highlight the limited need for understanding long-range dependencies in resolving these tasks, discuss the specific design decisions that seem to have led to the QuALITY task requiring reliance on long-range dependencies to be solved, and point out specific modeling choices that seem to outperform on the QuALITY task.

In order to achieve interoperability of information in the context of the Semantic Web, it is necessary to find effective ways to align different ontologies. As the number of ontologies grows for a given domain, and as overlap between ontologies grows proportionally, it is becoming more and more crucial to develop accurate and reliable techniques to perform this task automatically. While traditional approaches to address this challenge are based on string metrics and structure analysis, in this paper we present a methodology to align ontologies automatically using machine learning techniques. Specifically, we use convolutional neural networks to perform string matching between class labels using character embeddings. We also rely on the set of superclasses to perform the best alignment. Our results show that we obtain state-of-the-art performance on ontologies from the Ontology Alignment Evaluation Initiative (OAEI). Our model also maintains good performance when tested on a different domain, which could lead to potential cross-domain applications.

The emergence of linked data in the form of knowledge graphs in RDF has been one of the most recent evolutions of the Semantic Web. This led to the development of natural language question answering systems that automatically translates a question into SPARQL based on these RDF knowledge graphs. In particular, hybrid question answering, the task of question answering by combining both structured (RDF) and unstructured knowledge sources (text) has emerged as an important challenge. This paper tackles hybrid question answering based on natural language questions. We present HAWK_R, a question answering system that improves an open source system called HAWK. We identify its limitations and propose enhancements using heuristic-based methods based on RDF and text search. Our results show a clear improvement of the F-score.

This paper explores the possibility of using neural embedding features for enhancing the effectiveness of ad hoc document ranking based on learning to rank models. We have extensively introduced and investigated the effectiveness of features learnt based on word and document embeddings to represent both queries and documents. We employ several learning to rank methods for document ranking using embedding-based features, keyword-based features as well as the interpolation of the embedding-based features with keyword-based features. The results show that embedding features have a synergistic impact on keyword based features and are able to provide statistically significant improvement on harder queries.

Creating meaning from a wide variety of available information and being able to choose what to learn are highly relevant skills for learning in a connectivist setting. In this work, various approaches have been utilized to gain insights into learning processes occurring within a network of learners and understand the factors that shape learners' interests and the topics to which learners devote a significant attention. This study combines different methods to develop a scalable analytic approach for a comprehensive analysis of learners' discourse in a connectivist massive open online course (cMOOC). By linking techniques for semantic annotation and graph analysis with a qualitative analysis of learner-generated discourse, we examined how social media platforms (blogs, Twitter, and Facebook) and course recommendations influence content creation and topics discussed within a cMOOC. Our findings indicate that learners tend to focus on several prominent topics that emerge very quickly in the course. They maintain that focus, with some exceptions, throughout the course, regardless of readings suggested by the instructor. Moreover, the topics discussed across different social media differ, which can likely be attributed to the affordances of different media. Finally, our results indicate a relatively low level of cohesion in the topics discussed which might be an indicator of a diversity of the conceptual coverage discussed by the course participants.

One of the most important prerequisites for achieving the Semantic Web vision is semantic annotation of data/resources. Semantic annotation enriches unstructured and/or semistructured content with a context that is further linked to the structured domain-specific knowledge. In particular, ontologybased semantic annotators enable the selection of a specific ontology to annotate content. This paper presents results of an empirical study of recent ontology-based annotators, namely Stanbol, KIM, and SDArch. Specifically, we evaluated the robustness of these annotators with respect to specific features of ontology concepts such as the length of concepts? labels and their linguistic categories (e.g., prepositions and conjunctions). Our results show that although significantly correlated according to most of the conducted evaluations, tools still exhibit their unique features that could be a topic of new research.

The reusability in learning objects has always been a hot issue. However, we believe that current approaches to e-Learning failed to find a satisfying answer to this concern. This paper presents an approach that enables capitalization of existing learning resources by first creating "content metadatardquo through text mining and natural language processing and second by creating dynamically learning knowledge objects, i.e., active, adaptable, reusable, and independent learning objects. The proposed model also suggests integrating explicitly instructional theories in an on-the-fly composition process of learning objects. Semantic Web technologies are used to satisfy such an objective by creating an ontology-based organizational memory able to act as a knowledge base for multiple training environments.

One of the goals of the knowledge puzzle project is to automatically generate a domain ontology from plain text documents and use this ontology as the domain model in computer-based education. This paper describes the generation procedure followed by TEXCOMON, the knowledge puzzle ontology learning tool, to extract concept maps from texts. It also explains how these concept maps are exported into a domain ontology. Data sources and techniques deployed by TEXCOMON for ontology learning from texts are briefly described herein. Then, the paper focuses on evaluating the generated domain ontology and advocates the use of a three-dimensional evaluation: structural, semantic, and comparative. Based on a set of metrics, structural evaluations consider ontologies as graphs. Semantic evaluations rely on human expert judgment, and finally, comparative evaluations are based on comparisons between the outputs of state-of-the-art tools and those of new tools such as TEXCOMON, using the very same set of documents in order to highlight the improvements of new techniques. Comparative evaluations performed in this study use the same corpus to contrast results from TEXCOMON with those of one of the most advanced tools for ontology generation from text. Results generated by such experiments show that TEXCOMON yields superior performance, especially regarding conceptual relation learning.

Competence-based learning models have great importance for learning resources: they constitute a meaningful structure for just-in-time and just-enough learning. In this paper, we present an ontology-based competence model that allows the on-the-fly generation of learning knowledge objects (LKOs). The automatic aggregation process relies on knowledge objects and ontologies created through text mining and natural language processing. It is guided by instructional theories encoded declaratively through SWRL. Our framework offers a constructivist learning approach through the presentation of the LKO's context to the learner based on domain ontology. Finally, it allows the standardization of the generated learning objects in SCORM and IMS-LD.

In this paper, we present The Knowledge Puzzle, an ontology-based platform designed to facilitate domain knowledge acquisition for knowledge-based systems and especially for intelligent tutoring systems. We present a new content model, the Knowledge Puzzle Content Model, that aims to create Learning Knowledge Objects (LKOs) from annotated content. Annotations are performed semi-automatically using natural language processing algorithms. These LKOs are then aggregated in an Organizational memory (OM) which serves as a knowledge base for an intelligent tutoring system (ITS)