Friday seminar

Verteidigung Diplomarbeit.

We present our idea for a new method of parsing discontinuous constituency trees and its current implementation progress. It consists of a new method of representing discontinuous constituency trees as matrices, an algorithm each for conversion from trees to matrices and vice versa and a neural network architecture that can produce those matrix tree represetations in a fully differentiable manner.

Zwischenvortrag Diplomarbeit.

We present our idea for a new method of parsing discontinuous constituency trees and its current implementation progress. It consists of a new method of representing discontinuous constituency trees as matrices, an algorithm each for conversion from trees to matrices and vice versa and a neural network architecture that can produce those matrix tree represetations in a fully differentiable manner.

Der Vortrag findet online via BigBlueButton statt.

Kolloquium Forschungsprojekt Grundlagen

Zwischenvortrag Forschungsprojekt Grundlagen

Verteidigung Bachelorarbeit

Zweite Prüfungsleistung AuD

Analyse eines Forschungsthemas

Abschlussvortrag Praktikum

Probevortrag

Kolloquium Profilprojekt Grundlagenforschung

Kolloquium Profilprojekt Grundlagenforschung

Zwischenvortrag Praktikum

Zwischenvortrag Forschungsprojekt

Bachelorverteidigung

Zwischenvortrag Forschungsprojekt

Zwischenvortrag Bachelorarbeit

Forschungsprojekt Grundlagen

Kolloquium Masterarbeit

Zwischenvortrag Forschungsprojekt Grundlagen

Zwischenvortrag Forschungsprojekt Grundlagen

Statusvortrag zum Promotionsvorhaben

Several languages (and in consequence also syntactic treebanks for these languages) such as Dutch (Lassy), German (NeGra, TiGer), but also English (Penn Treebank) contain so-called discontinuous structures that cannot be adequately modelled by context-free grammars. In consequence, a class of more power grammar formalisms called mildly context-sensitive has been studied. Although the parsing complexity of these grammar formalisms is polymomial, it is usally of higher degree than that of CFG. Hence, also other methods based on graph reductions and oracle-driven transition systems with more favourable complexity have been proposed.

In this talk I give an overview on different methods proposed for discontinuous parsing. Furthermore, I present my own contributions to the area of grammar-based discontinuous parsing. In particular, I recall linear context-free rewriting systems and hybrid grammars and present a method to learn powerful probabilistic extensions of these grammars. I discuss empirical results when applying said models to data from multiple languages.

Kolloquium zum Großen Beleg

Zwischenvortrag Masterarbeit

Bericht über die WHK-Tätigkeit

[notes, paper]

[paper, slides]

[slides]

[slides, GMS04]

[slides, vCSB2016, vC2012]

I give an overview of the discontinuous parsing framework disco-dop. In particular, I discuss which high-level and low-level optimizations were chosen in order to achieve state-of-the-art LCFRS parsing.

[slides]

We talk about pushdowns, logarithmic probabilities, and double-ended priority queues (with example code in Rust).

[slides]

[slides]

[slides]

[script]

Role and Reference Grammar (RRG) is a theory of the grammar of natural language which has been developed as a descriptive tool for the analysis of typologically distinct languages, and which takes into account the interaction between syntax, semantics, and pragmatics. While RRG already contains a number of formal elements such as syntactic templates and the logical structures, it lacks a thorough formalization, which makes it difficult to study its formal properties or to develop a computational implementation.

In the talk, we present an outline of how a formal framework for RRG could look like. Our approach adopts and extends concepts from Lexicalized Tree Adjoining Grammars (LTAG) for the specification and composition of syntactic templates. We introduce two general operations, sister adjunction and wrapping substitution, for the composition of elementary syntactic trees. The resulting formalism, Tree Wrapping Grammar (TWG), is more powerful than LTAG. However, if the number of wrapping substitutions that can stretch across a node in the derived tree is limited to some constant k, it is possible to construct a weakly equivalent simple context-free tree grammars of rank k, which is in turn weakly equivalent to a well-nested linear context-free rewriting systems (LCFRS) of fan-out k + 1. Hence, with this linguistically plausible restriction, our formalization of RRG leads to a mildly context-sensitive grammar formalism.

We furthermore show how the operator projection of RRG, which reflects the scopal properties of functional operators, can be integrated into a single tree representation by adding appropriate node and edge features. This extends our formalism to a feature-based TWG.

The work presented in this talk is to a large extent joint work with Rainer Osswald.

[slides]

We present two parsing algorithms for weighted multiple context-free grammars (short: wMCFGs). The first one is based on an automata characterisation of wMCFGs and a coarse-to-fine parsing algorithm for weighted automata with data storage. The second one is based on a Chomsky-Schützenberger characterisation of wMCFGs.

[slides]

[slides]

[paper]

[paper, slides]

[book]

[slides, paper]

We develop a definition of automata with data storage by stripping down the classic definition. We observe that automata with data storage can be extended to weighted transducers elegantly, using the free semiring.

[slides]

[slides]

We have a close look at the constructions of C-S parsing and find strong similarities to the well-established chart parsing approach.

[slides]

[slides, abstract]

[slides, paper]

[slides]

[slides]

[slides]

[slides]

[slides]

[slides]

[slides]

[slides]

[slides]

I give a short introduction to word embeddings and, in particular, to the skip-gram model implemented in word2vec.

[slides]

[slides, reference, code]

The inside-outside algorithm computes so-called inside weights and outside weights for a weighted (intersection-)grammar. I illustrate that each outside weight can also be obtained as a certain partial derivative of inside weights. This property can be exploited to obtain an EM algorithm from automatic differentiation present in tools such as dynet or Tensorflow.

[slides]

We apply the idea of coarse-to-fine parsing to recognition with weighted automata with storage by introducing the concept, and implementing an algorithm in the framework rustomata.

[slides]

We use weighted deduction systems to describe three different types of parsers for weighted linear context-free rewriting systems. Furthermore, we introduce heuristic weights for these parsers to enable an informed best-first search among parsing results. We implemented all described formalisms in Haskell using the Vanda framework and tested the implementation with grammars extracted from a natural language corpus.

[slides, paper]

We present an n-best parsing algorithm based on the Chomsky-Schützenberger theorem for weighted multiple context-free languages. The algorithm will be implemented and evaluated in the future.