Expectations

At KAI, we set clear expectations for students. We want to make sure students know what to expect from us as their supervisors. We have prepared a short document which touches upon some important points like meetings, planning and writing of your thesis.

Table of contents

• VU Theses
  • Information Extraction
  • Ontology Evolution
  • Hybrid Intelligence
  • Argument and Rule Mining
  • Robotics and Knowledge Representation
  • Explanations and Narratives
  • Question Answering
  • Multi-lingual problems
  • Semantics of Deep Learning Methods
  • Explanations for Ontologies
  • Other Topics on Ontologies and Description Logics
• Internships
  • Lareb
  • Accenture
  • Triply DB

KAI Theses

You can find the slides presented at the MSc AI Thesis event here. Most of the topics below can be investigated by either BSc or MSc AI students. We also welcome groups of students working on the same or similar topic.

If you are interested in one of the projects below, please contact the supervisor(s) listed to receive more information about the topics. Where available, have a look at the detailed description first. Also, keep in mind that all theses can be shaped to accomodate your interests.

Information Extraction

Supervisor: Benno Kruit (b.b.kruit@vu.nl), Ilaria Tiddi (i.tiddi@vu.nl), Lise Stork (l.stork@vu.nl), Ritten Roothaert (h.m.roothaert@vu.nl)

We offer multiple projects under the umbrella of information extraction with varying foci. Information extraction focuses on generating structured data from unstructured inputs in an automated manner. The input as well as the output can vary based on the application or end usage of the extracted data.

• Automated Processing of Scholarly Data: (Ilaria) The goal of this project is to support the automatisation of processing the CEUR-WS proceedings data. For a BSc thesis, the objective is to extract an ontology of CEUR knowledge. For a MSc thesis, this would be extended with analysing abstracts or creating an interface for data input and knowledge graph population.
• Making research papers machine interpretable: (Lise) The goal of this project is to see if we can partially automate or support the construction of knowledge graph on hypotheses from the literature. Here is a more detailed description.
• Information extraction from Structured lists (BSc, Benno): Many different kinds of documents contain lists because they are a simiple way of enumerating several related items. We want to investigate ways of extracting the information from the lists and retaining the inherent relationship between list items. More info here.
• Higher-arty Relation Extraction with Qualifiers (MSc, Benno): This projects aims to inverstigate new techniques for extracting complex statements with meta-information from text, in particular Wikipedia. The goal is to leverage ontology/schema-level information about types, relations, and meta-relations to overcome the incompleteness problem. More information can be found here.
• Multilingual Travel Knowledge Extraction (MSc, Benno): In this project, we want to explore approaches for linking geographical location and information together within the application of travel. More details are given here.
• Data Schema Induction for Shopping (MSc, Benno): The goal is to investigate approaches for inducing a ceherent ontology for product descriptions. More information is available here.
• Automated Data-provenance Extraction (BSc, Ritten): In a world where linked data and data-mining is omnipresent, determining which data was used for training a ML-model becomes increasingly more difficult and tedious. This project revolves around automating the process of extracting the provenance information of data used in a ML-pipeline. More details are given here.
• Creating knowledge graphs for olfactory dysfunction and mental disorders (MSc, Lise, Halima) Biomedical text mining has emerged as a powerful tool to build up knowledge graphs and guide us through the vast plethora of available scientific literature and medical records. In this MSc. project, we primarily aim at the construction of a knowledge graph for the understanding of the link between olfactory dysfunction and mental diseases. Project together with Halima Mouhib from Bioinformatics.

Ontology Evolution

Supervisor: Romana Pernisch (r.pernisch@vu.nl)

Ontologies model specific domains. As domains evolve over time, ontologies have to be changed as well. Not only are the ontologies themsevels affected but also applications using those ontologies for various purposes. We have multiple theses in this domain.

• Evolution Process: The process of ontology evolution is relatively long and complex. In your thesis you can investigate and compare tools for supporting this process or investigate automations options within this process. Multiple projects within this area are possible and some ideas and details are given here.
• ChImp 2.0: The ChImp Protégé plugin helps ontology engineers during this process by summarising and displaying changes and the effects of changes on the ontology as a whole. We have multiple possible projects with ChImp. More information is here.
• Materialisation/Reasoning: We have previously investigated the impact on the materialisation (making implicit knowledge explicit) and want to further the analysis by diving into more depth. This means that we want to investigate the types of changes in more detail but also the effect of the changes more localized in the materialisation, rather then looking at the materialisation as a whole. More details are given here.

Hybrid Intelligence

Supervisor: Ilaria Tiddi (i.tiddi@vu.nl), Lise Stork (l.stork@vu.nl)

• Knowledge Engineering for Hybrid Intelligence (Bsc, MSc): Inspired by Software Design and Engineering, Knowledge Engineering deals with the formal design, maintainance and usage of knowledge-based systems. In this project, we will look at modelling Hybrid Intelligent systems using knowledge engineering techniques.
• Embodied Instructable Agents (MSc): The goal of this project is to integrate a Reinforment Learning model trained for trajectory/segmentation learning on an embodied Hybrid Intelligent agent (a ROS-operting robot). Knowledge of one between ROS, Reinforcement Learning and simulation environments is required.
• Researching human-in-the-loop workflows for research assistants using KGs: (MSc, Lise) The goal of this project is to research human-in-the-loop workflows for digital assistants for scientific discovery. Here is a more detailed description.

Argument and Rule Mining

Supervisor: Loan Ho (t.t.l.ho@vu.nl), Lise Stork (l.stork@vu.nl), Atefeh Keshavarzi(a.keshavarzi.zafarghandi@vu.nl)

There are multiple projects in the domain of argument mining with different objectives:

• Computing extensions for argumentation with collective attacks (BSc): In this project, we will build a tool to calculate extensions for argumentation with collective attacks. The inputs are a set of arguments and collective attacks. The output returns different (stable, preferred, complete, grounded) extensions. Here are more details.
• Explaining query answers in prioritized databases (BSc): This project aims to provide explanation of how the query answer was reached in consistent database. Here are more details.
• Handling inconsistencies in argumentation knowledge: In the project, we will address inconsistencies in argumentation data by using logic-based reasoning. Here are more details.
• Leveraging Argumentation Frameworks for Local Explainability of Black-Box Models (MSc, Atefeh): This project aims to leverage the formalisms of argumentation to provide local explainability for black-box models, enabling non-experts to understand the reasons behind the system’s decisions. Here are more details.
• Orange3 Argument Mining Widget Based on a Formalism of Argumentation (MSc, Atefeh): This project aims to bridge the gap between argumentation theory and data analysis and visualization by developing an argument mining widget for Orange3, an open-source data mining and machine learning software. Here are more details.

Rule mining, similarly to information extraction, aims at finding structures. In this case, we want to learn rules that describe the data best to help us understand it better. There are multiple projects that involve rule mining:

• Community Detection from a Species Interaction Network: The goal of this project is to extract interesting communities in species interaction networks and or detect interesting logic rules that relate to these communities. More details are given here

Robotics and Knowledge Representation

Supervisor: Ilaria Tiddi (i.tiddi@vu.nl), Mark Adamik (m.adamik@vu.nl)

These projects look at the intersection of robotics and knowledge graphs. Knowledge of Robotic Operating System (ROS) is a plus:

• Semantic Mapping with a mobile service robot (MSc): The goal of this project is to implement semantic mapping for an indoor mobile robot using knowledge graphs. More information about this project can be found here.
• Visual Scene Understanding for Indoor Mobile Robots (BSc): In this project, you will use the camera feed of the robot and off-the-shelf image recognition algorithms to understand the environment by generating knowledge graphs. More information about this project can be found here.
• Energy-efficient robots through knowledge-awareness. (BSc, MSc): We have an ontology representing the capabilities of the robots (picking objects, moving, scanning surroundings, etc.). The ontology needs to be expanded with energy budgets so the robot can choose the actions to performed based on its capabilities and energy-efficiency. Collaboration with the Software Engineering group.
• Embodied Instructable Agents (MSc): See above.
• Impact of ontology changes in robotic tasks. (MSc): Based on a robot operating with an ontology in the background, the goal here is to study how changes in such ontology can positively/negatively impact the tasks the robot has to perform. Knowledge of one between ROS, RDFlib and simulation environments is required.
• Explaining Yolo Predictions with Argumentations. (BSc,MSc): The goal of this project is to explain the predictions of an object recognition model in the form of argumentations. Knowledge of one between ROS, RDFlib and Yolo/Object recognition algorithms is required.

Explanations and Narratives

Supervisors: Lise Stork (l.stork@vu.nl), Ilaria Tiddi (i.tiddi@vu.nl)

The following topics are aimed at providing a more human-like AI, by creating explanations or creating narratives.

• Identifying Formal Narratives from KGs: (Ilaria) The goal of this project is to extract as many narratives as possible (in terms of set of facts) from existing KGs such as DBpedia or Wikidata. In order collect these facts, we will use the narrative formal structure as presented in this paper.
• A Benchmark for understanding Narratives: (Ilaria) Language Models and KGs. Work on extending the three existing benchmarks (1, 2, 3) for understanding narratives.

• Semantic robustness of Language Models with causal inference. See below.

• SHACL-forms for publishing scientific findings: (MSc, Lise) The goal of this project is building a form from SHACL shapes for the publication of scientific findings. Here is a more detailed description.

Question Answering

Supervisors: Benno Kruit (b.b.kruit@vu.nl), Stefan Schlobach (k.s.schlobach@vu.nl), Lise Stork (l.stork@vu.nl)

QA is a very broad topic. We, however, focus on QA over structured data in various forms:

• Playing “20 Questions” with a KG (BSc): detailed description is located here.
• Multi-modal Question Answering (MSc): Description here.
• Graph Queries on Relation Databases (MSc CS): Description here.
• Answering Research Questions over Data Cubes as SQA (MSc., Lise) : Description here.

  Multi-lingual problems

  Supervisors: Benno Kruit (b.b.kruit@vu.nl)

Even though these topics would also fit under different topics already discribed above, we wanted to highlight them as they are both addressing the problem of multiple languages in different tasks:

• Multilingual Entity Linking (BSc): Many names can refer to several different entities. In this project, we want to look at the problem of disambiguation and linking of entities. More info here.
• Multilingual Travel Knowledge Extraction (MSc): This project is about integrating Wikivoyage location data with data from OpenStreetMap. The approach will leverage distant supervision, relation extraction, data integration and deep learning techniques. More details can be found here.

Semantics of Deep Learning Methods

Supervisor: Ilaria Tiddi (i.tiddi@vu.nl)

• Incorporating Semantics in Message Passing methods: Message passing models are neural network architectures that operate by propagating information along the structure of a graph over which they are trained end-to-end. Currently, these methods treat all relationships in the same way, while in knowledge graphs some edges carry more semantics than others (e.g. entity type or subclass hierarchies). Here, we will look at feeding such information in a message passing model such as R-GCN and test it in a node labelling or link prediction scenario.
• Extracting Semantics from neural co-activation graphs: Here, we will look at extracting taxonomical or ontological information using the analysis of the co-activation graph (CoAG) of a neural network architecture. We will use community analysis incrementally to analyse different knowledge granularities (CoAG alone, CoAG with rdfs:subClassOf, CoAG with other relationships). The taxonomy will help understand what the neural representation has learned, and it will be compared with a ground-truth KG to see how correct the neural representation is.
• Semantic robustness of Language Models with causal inference: We will adapt the approach in this paper to test how robut language models are toward different semantics (description logics, OWL, etc.) The idea is to build a causal model for graph2text tasks and compare its prections with the ones of a language model. Understanding of probability theory, statistics and graph theory is necessary.

Explanations for Ontologies

Supervisor: Patrick Koopmann (p.k.koopmann@vu.nl)

We offer a range of projects around the topic of explanations for ontologies. We focus on ontologies based on description logics or OWL. A main advantage of formulating knowledge in such a formalism is that one can use a reasoner to derive implicit information. However, not always is the result of this reasoning process easy to understand: users might wonder why something was derived (explain positive entailment), or why something was not derived (explain negative entailments). Motivated by this, different methods have been developed to provide explanations for positive and negative entailments.

• Explaining Entailments using New Inference Rules (Bsc): One way of explaining why something follows from an ontology is by providing a proof tree, that shows in small steps how the positive entailment is derived from the statements in the ontology. Such a proof is usually generated based on a set of rules. We have a tool that can process user-defined sets of inference rules to generate rules. Existing sets of rules are usually not optimized for human understanding. The aim of this project is to develop new sets of inference rules that lead to nicer proofs, and provide an evaluation and comparion with existing explanation methods based on realistic ontologies.

• Explaining Positive Entailments using Universal Models (Bsc,Msc): The topic of this project is to explain queries to data that is used together with an ontology. Specifically, the user asks for instances of a concept C, and he would like to understand why individual name “a” is an answer to this query. There are ontology languages where this can be explained using a special type of models called universal models. The aim of this project is to develop a method to generate explanations based on such models.

• Explaining Missing Entailments using Counter-Examples (Bsc,Msc): One way to explain a missing entailment is by providing an counter example. For example, a counterexample for the statement “every pizza is vegetarian” from an ontology about pizzas would be a pizza with a salami topping, which would be model of the concept “Pizza”, but not of “VegetarianPizza”. The topic of this project is to develop and evaluate a method for computing such counterexamples. The Master version is towards developing a new method based on existing reasoning procedures. The Bachelor version will be about improving and extending an existing method.

• Explaining Missing Entailments using Connection-Minimal Abduction (Bsc,Msc): Abduction is an approach to explain missing entailments by stating “what is missing” - namely, suggesting statements that, when added to the ontology, would make the entailment positive. There are many different conditions one can give to such an explanation to make it “useful”. This project focusses on a recently discovered criterion called “connection-minimality”. Depending on the interests, there are different directions possible. A more practically interested student would get the task of developing a new method for performing connection-minimal abduction, and compare it to the state-of-the-art. For students that are less interested in implementations, the project would be to develop a new criterion based on connection-minimality, which overcomes some of the limitations of the existing one.

• Explaining Missing Entailments using Signature-Based Abduction (Bsc): Abduction is an approach to explain missing entailments by stating “what is missing” - namely, suggesting statements that, when added to the ontology, would make the entailment positive. There is an abduction tool called LETHE-abduction that computes such explanations based on a provided signature: a vocabulary of names that are allowed to be used in the explanation. Selecting such a vocabulary is currently up to the user. The goal of this project is to investigate and compare heuristics for selecting signatures for computing a nice sequence of explanations.

Other Topics on Ontologies and Description Logics

We offer a range of further theses topics from the areas of ontologies and ontology languages. These topics are particularly interesting for students who like ontologies and/or logics.

• Learning Concept Descriptions From Examples (Bsc,Msc): The aim of this project is to implement and evaluate a new method for learning concept descriptions from examples, based on recent advancements on non-classical reasoning. In this scenario, the ontology is used together with a set of positive and negative examples. The aim is then to generate a concept that
  describes all positive, but none of the negative examples. The learning procedure will be based on logical reasoning, that is, will use logical reasoning (with the help of existing tools) to compute the concept.

• Extracting Subontologies (Bsc) Existing ontologies are often very large and complex, while applications are often only need a fragment of the information provided by the ontology. A subontology is a smaller and ideally simpler ontology that covers all relevant information from the ontology for a user-provided set of terms of interest. There are different techniques (module extraction, uniform interpolation) that can be used to extract subontologies. The aim of this project is to investigate and evaluate heuristics to improve the performance of these methods, or develop a new approach that works by combining them, with the aim of obtaining simpler ontologies and/or shorter computation times.

• Optimizing Concept Expressions (Bsc,Msc): Ontologies often contain expressions that are more complex than necessary. This is even more a problem with ontology content that is automatically generated, which appears in many applications. The aim of this project is to develop a method to optimize concept expressions by replacing them by equivalent expressions that are of minimal size. We recently developed a method that can do this for a language of rather limited expressivity. The student will extend this framework, possibly using techniques from concept learning.

• Automated Hypothesis Generation using ABox abduction (Msc): This project looks at the following problem: we have an ontology, as well as some data in the form of a knowledge graph of ABox. This contains our background knowledge about some domain such as medicine, or a context from robotics. We are then given a set of facts that do not follow from what we know according to our background knowledge - an observation that is somehow unexpected, for instance a description of symptoms of a patient or of an unexpected situation encountered by a robot. We then want to generate a hypothesis in the form of a set of facts that would explain the observation if added to the background knowledge. To avoid trivial answers, we assume that there is also a special vocabulary for explanations provided. This means, we want to compute a hypothesis that uses only terms from that vocabulary, but may refer also to unknown objects. This problem is called signature-based ABox abduction. The aim of this project is to develop a new method for signature-based ABox abduction based on some recent theoretical results of this problem.

Internships and Erasmus Programmes

Make sure to first check out the official internship process!

We suggest you to pick a topic from the ones above, and adapt it to the company needs in agreement with your supervisors (both from the company and from the group). We do have contacts in several companies and research centres to do experiences abroad. For more information, contact Ilaria Tiddi (i.tiddi@vu.nl)*