+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Freitag, 12. Juli 2019, 10.00 Uhr
Ort: Informatikzentrum, E3, Raum 9222
Referent: Dipl.-Inform. Malte Nuhn
Thema: Unsupervised Training with Applications in Natural Language
Processing//
Abstract:
The state-of-the-art algorithms for various natural language processing
tasks require large amounts of labeled training data. At the same time,
obtaining labeled data of high quality is often the most costly step in
setting up natural language processing systems.Opposed to this,
unlabeled data is much cheaper to obtain and available in larger
amounts.Currently, only few training algorithms make use of unlabeled
data. In practice, training with only unlabeled data is not performed at
all. In this thesis, we study how unlabeled data can be used to train a
variety of models used in natural language processing. In particular, we
study models applicable to solving substitution ciphers, spelling
correction, and machine translation. This thesis lays the groundwork for
unsupervised training by presenting and analyzing the corresponding
models and unsupervised training problems in a consistent manner.We show
that the unsupervised training problem that occurs when breaking
one-to-one substitution ciphers is equivalent to the quadratic
assignment problem (QAP) if a bigram language model is incorporated and
therefore NP-hard. Based on this analysis, we present an effective
algorithm for unsupervised training for deterministic substitutions. In
the case of English one-to-one substitution ciphers, we show that our
novel algorithm achieves results close to human performance, as
presented in [Shannon 49].
Also, with this algorithm, we present, to the best of our knowledge, the
first automatic decipherment of the second part of the Beale
ciphers.Further, for the task of spelling correction, we work out the
details of the EM algorithm [Dempster & Laird + 77] and experimentally
show that the error rates achieved using purely unsupervised training
reach those of supervised training.For handling large vocabularies, we
introduce a novel model initialization as well as multiple training
procedures that significantly speed up training without hurting the
performance of the resulting models significantly.By incorporating an
alignment model, we further extend this model such that it can be
applied to the task of machine translation. We show that the true
lexical and alignment model parameters can be learned without any
labeled data: We experimentally show that the corresponding likelihood
function attains its maximum for the true model parameters if a
sufficient amount of unlabeled data is available. Further, for the
problem of spelling correction with symbol substitutions and local
swaps, we also show experimentally that the performance achieved with
purely unsupervised EM training reaches that of supervised training.
Finally, using the methods developed in this thesis, we present results
on an unsupervised training task for machine translation with a ten
times larger vocabulary than that of tasks investigated in previous work.
Es laden ein: die Dozentinnen und Dozenten der Informatik
_______________________________________________
--
--
Stephanie Jansen
Faculty of Mathematics, Computer Science and Natural Sciences
HLTPR - Human Language Technology and Pattern Recognition
RWTH Aachen University
Ahornstraße 55
D-52074 Aachen
Tel. Frau Jansen: +49 241 80-216 06
Tel. Frau Andersen: +49 241 80-216 01
Fax: +49 241 80-22219
sek(a)i6.informatik.rwth-aachen.de
www.hltpr.rwth-aachen.de
Tel: +49 241 80-216 01/06
Fax: +49 241 80-22219
sek(a)i6.informatik.rwth-aachen.de
www.hltpr.rwth-aachen.de
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Dienstag, 11. Februar 2020, 14.00 Uhr
Ort: Ahornstraße 55, E3, Raum 118 (Seminar-Raum I8)
Referentin: Dipl.-Inform. Ellen Dekkers
Lehrstuhl Informatik 8
Thema: Feature-aware and feature-driven editing of 3D surface meshes
Abstract:
Feature-aware and feature-driven editing of three-dimensional surface
meshes is a very prominent task in Computer Graphics and Geometry
Processing applications. Existing methods can be roughly classified into
general elastic approaches, which aim at the preservation of a manifold
surface's differential properties and hence of local, low-level
geometric surface detail, and structure-aware editing techniques
focusing on the preservation of high-level surface structures such as
feature curves or regular patterns. In this talk, we review both
research fields and discuss the respective approaches with a particular
focus on their capabilities in preserving various types of surface
features.
We then present a novel approach to feature-aware mesh editing that
combines elastic Laplacian deformation with discrete plastic topology
modifications by transferring the concept of seam carving from the image
retargeting to the mesh deformation scenario. During editing, a
precomputed set of triangle strips, or geometry seams, can be
dynamically deleted or inserted in low saliency mesh regions, thereby
distributing the deformation distortion non-homogeneously over the model
which yields a much better preservation of salient surface features
compared to standard elastic deformation.
Finally, we remove the manifold restriction and address feature curve
driven editing of non-manifold meshes. First, we propose a
semi-automatic approach to efficiently and robustly recover
characteristic feature curves from free-form surfaces. We then present
two practical applications of this technique, the first of which
exploits the curves' shape-defining properties and employs them as
intuitive modeling handles for editing non-manifold surfaces. In our
second application, we turn to a practical scenario in reverse
engineering and consider the problem of generating a statistical shape
model for car bodies. The crucial step of establishing proper feature
correspondences between a large number of input models that exhibit
significant shape variations is essentially guided by characteristic
feature curves. These curves furthermore serve as modeling metaphors for
intuitive exploration of the shape space spanned by the input models,
thereby enabling the generation of semantically meaningful, novel car
bodies.
Es laden ein: die Dozentinnen und Dozenten der Informatik
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Donnerstag, 13. Februar 2019, 13.00 Uhr
Ort: Informatikzentrum, E3, Raum 9222
Referent: Sebastian Junges, M.Sc.
Lehrstuhl für Informatik 2 (Software Modeling and Verification)
Thema: Parameter Synthesis in Markov Models
Abstract:
Markov models comprise states with probabilistic transitions.
The analysis of these models is ubiquitous and studied in,
among others, reliability engineering, artificial intelligence, systems biology, and formal methods.
Naturally, their analysis crucially depends on the transition probabilities.
Often, these probabilities are approximations based on data or reflect configurable parts of a modelled system.
To represent the uncertainty about the probabilities, we study parametric Markov models,
in which the probabilities are symbolic expressions rather than concrete values.
More precisely, we consider parametric Markov decision processes (pMDPs)
and parametric Markov chains (pMCs) as special case. Substitution of the parameters yields classical,
parameter-free Markov decision processes (MDPs) and Markov chains (MCs).
A pMDP thus induces uncountably many MDPs. Each MDP may satisfy reachability and reward properties,
such as "the maximal probability that the system reaches an `offline' state is less than 0.01%",
or "the maximal expected energy consumption is less than 20 kWh."
Lifting these properties to pMDPs yields fundamental problems asking:
- "Is there an induced MDP satisfying the property?" (feasibility), its dual
- "Do all induced MDPs satisfy the property?" (validity),
and advanced problems such as "What is a concise representation for all induced MCs satisfying the property?"
We study these problems on a conceptual level, and design and implement both improved and novel algorithms.
On the conceptual side, a thorough discussion of the feasibility problem yields new results, such as:
(1) that answering various variants of the feasibility problem is — in terms of complexity — as hard as finding roots of a multivariate polynomial, and
(2) that these problems are tightly connected to the analysis of memoryless strategies in partially observable MDPs, a famous model in artificial intelligence.
Additionally, we introduce family MCs (fMCs), a subclass of pMCs with finitely many induced MCs.
Among others, fMCs define fundamental problems underlying the quantitative analysis of software product lines and sketching of probabilistic programs.
On the algorithmic side,
(1) we present and analyse improved but previously known approaches, and combine them to meet practical needs.
Their analysis inspired (2) three new and orthogonal approaches, utilising advances in convex optimisation,
as well as adapting prominent ideas such as inductive synthesis and abstraction-refinement to the particular setting.
All methods efficiently exploit advances in the off-the-shelf analysis of MDPs.
On the empirical side, the new methods improve the state-of-the-art considerably, handling hundreds of parameters and millions of states.
All the approaches we present have been implemented in open-source tools.
Es laden ein: die Dozentinnen und Dozenten der Informatik
Sehr geehrte Damen und Herren,
wir möchten Sie freundlich auf den Vortrag (s.u.) von Herr Dr. Markus Freitag heute Nachmittag um 14:30 hinweisen.
Bitte entschuldigen Sie die kurzfristige Ankündigung.
Mit freundlichen Grüßen,
Christian Herold
***********************************************************************
*
*
* Einladung zum Gastvortrag
*
*
*
************************************************************************
Zeit: Dienstag, 14. Januar 2020, 14:30
Ort: Ahornstraße 55, E2, Raum 5056
Referent: Dr. Markus Freitag
Titel: (Some) Research Happening at Google Translate
Abstract:
Machine Translation is one of the most appealing research topics in
Natural Language Processing and Machine Learning. In this talk, you will
be given an overview of some of the current research efforts happening
at Google Translate.
We will start with a project with the end-goal of training a single
model to translate between all languages supported by Google Translate.
Neural models can be trained to perform several tasks simultaneously as
exemplified by multilingual NMT using a single model to translate
between multiple languages. Apart from reducing operational costs,
multilingual models improve performance on low and zero-resource
language pairs due to joint training. We attempt to study multilingual
neural machine translation, using a massive open-domain dataset
containing over 25 billion parallel sentences in 103 languages.
In the second half of the talk, we focus on translationese, a term that
refers to artifacts present in text that was translated into a given
language that distinguish it from text originally written in that
language. These artifacts include lexical and word order choices that
are influenced by the source language as well as the use of more
explicit and simpler constructions. Machine translation has an
undesirable propensity to produce translationese artifacts, which can
lead to higher BLEU scores while being liked less by human raters.
First, we train an Automatic Post Editing (APE) model that convert the
translationese output into a more natural text. We use this model as a
tool to reveal systematic problems with reference translations. Second,
we model translationese and original (i.e. natural) text as separate
languages in a multilingual model, and pose the question: can we perform
zero-shot translation between original source text and original target
text? To sum up, we will discuss why a loss in BLEU score does not
always mean lower translation quality.
--
Christian Herold
Faculty of Mathematics, Computer Science and Natural Sciences
HLTPR - Human Language Technology and Pattern Recognition
RWTH Aachen University
Ahornstraße 55
D-52074 Aachen
Tel: +49 241 80 21613
Fax: +49 241 80 22219
herold(a)i6.informatik.rwth-aachen.de
www.hltpr.rwth-aachen.de
***********************************************************************
*
*
* Einladung
*
*
* Informatik-Kolloquium
*
*
***********************************************************************
Zeit: Mittwoch, 8. Januar 2020, 14:00 Uhr
Ort: Raum 9222, Gebäude E3, Informatikzentrum
Referent: Michael Schaub
University of Oxford, UK and MIT, USA
Thema: Data Science for Networks
Abstract:
Networks have become a widely adopted model for a range of systems,
cutting across Science and Engineering.
However, our theoretical understanding of many fundamental phenomena
that arise in complex networks and networked systems is still limited.
My vision is to develop a data science for networks and dynamical
systems that will contribute to addressing this challenge, by combining
data-driven and model-based approaches, using the language of graphs and
networks.
In this talk, we will give a brief overview of such a Data Science for
Networks.
We first discuss how networks appear naturally within models in
different research domains and illustrate the underlying scientific
questions via examples drawn from applications.
We then examine in some detail the problem of feature learning from
graphs with unobserved edges, in which we aim to learn certain aspects
of a graph solely from dynamical observations on its nodes, without
knowledge of the edge-set of the graph.
We conclude with a brief outlook on future challenges and open problems.
Es laden ein: die Dozentinnen und Dozenten der Informatik