Περιγραφές θέσεων για τέσσερις (4) υποτροφίες για την εκπόνηση διδακτορικής διατριβής του ΙΠ&Τ σε συνεργασία με Πανεπιστήμια του εξωτερικού

Printer-friendly versionSend by email

1 θέση για University of Houston

Τίτλος: Αναγνώριση επιχειρημάτων σε κείμενα φυσικής γλώσσας, με έμφαση στην αναγνώριση των συστατικών στοιχείων ενός επιχειρήματος και των σχέσεων μεταξύ των στοιχείων αυτών αλλά και μεταξύ διαφορετικών επειχειρημάτων. (Identification of argument elements and argument relations in natural laguage texts.)

Περιγραφή: This thesis aims to tackle a relatively new and exciting challenge in corpus-based discourse analysis, argument mining, which tries to automatically identify human reasoning in unconstrained natural language texts. The PhD student will be able to explore theories developed in the humanities for describing argument and rhetoric and approaches based on machine learning and statistical techniques from computational linguistics, in order to address challenges similar to the following ones:
1) Define algorithms for identifying argument elements in natural language texts.
2) Propose algorithms for identifying intra-argument relations, organizing the identified elements that belong to a single argument in order to mine the internal structure of an argument.
3) Propose algorithms for identifying inter-argument relations, that determine the role of each argument in the narrative of a text passage.


2 θέσεις για University of Texas at Dallas

Τίτλος 1ης θέσης: Εξαγωγή επιχειρημάτων από αναρτήσεις χρηστών στα μέσα κοινωνικής δικτύωσης. (Argument mining from user generated content.)

Περιγραφή 1ης θέσης: Argument mining involves several subtasks, such as discrimination between argumentative and non-argumentative text units, the detection/identification of the elements that comprise an argument such as claims, premises, support or conclusions, the identification of relations among argumentative elements and among identified arguments. Argumentation, having an important role in human communication, can also play an important role in improving extraction accuracy in tasks like sentiment analysis or opinion mining, if it can be performed on user generated content, which still remains an open and understudied research topic. This thesis aims to tackle open issues in argument mining from unconstrained natural language texts and more specifically from user generated content.

Τίτλος 2ης θέσης: Εξαγωγή πληροφορίας από το web για την αναγνώριση και συντήρηση περιγραφών οντοτήτων. (Entity Description Maintenance via the Analysis of Unmoderated Online Information Sources.)

Περιγραφή 2ης θέσης: The description of a given entity under a certain conceptualization is subject to changes regarding the values of the various properties associated with the entity. When using the web as the primary source for extracting information pertaining to the described entities, the discovery and assessment of new property values is an ambiguous process, as it is generally difficult to (i) accurately isolate the possible new value of a specific property, (ii) determine if the discovered value still holds as the current value of the property, and (iii) assess the credibility of the online source(s) where the value was discovered. The goal of the present thesis is the specification and implementation of methodologies for dealing with the aforementioned challenges, focusing on the analysis of unmoderated textual online information. The multifaceted nature of web text in terms of language complexity, text size (e.g. the case of social media posts and microblogs), document structure (e.g. distribution of an article in multiple pages), etc. imposes further complexity on the relevant language processing tasks. The overall approach should also take into account temporal and contextual ambiguities, as well as the provenance of the retrieved information in order to assess its credibility and, therefore, the confidence on the proposed description updates.


1 θέση για Rice University

Τίτλος: Σχεδιασμός κίνησης ρομπότ με βάση την εμπειρία. (Experience-Based Robot Planning.)

Περιγραφή: Robot motion planning is a core area in robotics concerned with finding feasible paths for a given robot and environment that take the robot from a given start state to a desired goal state. Current state-of-the-art sampling-based algorithms solve each motion planning problem from scratch. Independently, learning algorithms have made great advances in a more data-driven approach to solving specific motion planning problems. While sampling-based algorithms provide strong completeness guarantees and, under certain conditions, optimality guarantees, they do not exploit similarities among different motion planning problems the way learning algorithms do. On the other hand, it is generally difficult to give performance guarantees about what learning algorithms have learnt. In particular, it is difficult to characterize how much the learnt behavior can generalize to different motion planning problem. The aim of this PhD dissertation is to bridge the gap between planning from scratch and learning so as to leverage their relative strengths. The result will be a motion planning framework that can quickly solve (potentially hard) problems that are similar to ones solved before, but preserves the completeness guarantees to solve problems that are very different from anything solved before. One of the challenges will be to define a suitable notion of similarity between abstractions of environments and motion planning problems, so that solutions to similar problems are indeed helpful with high probability. These problems can be in the context of either fully autonomous operation or assisted teleoperation. The framework will benefit greatly from having multiple robots of the same type share their data. We envision a cloud-based infrastructure that can maintain an “experience database” that can be queried by each robot. This means that representations of environments, feasible motions, and goal states (including both autonomously computed goals and user intention in teleoperation) need to be concise so that information can be exchanged efficiently.

© 2018 - Institute of Informatics and Telecommunications | National Centre for Scientific Research "Demokritos"

Terms of Service and Privacy Policy