Automatically checking whether generated text is factually consistent with an input segment is still challenging. We present a method that goes against the recent trend of directly utilizing large language models to evaluate factuality, and instead propose a more linguistically grounded approach, based on Semantic Role Labels.

Relative to the English NLP community, we find that the quality of German summarization datasets (and models) is heavily lacking; oftentimes, not even basic filtering criteria are respected when training and evaluating systems.

We describe our winning submission to the shared task on lexical simplification. In principle, we extract structured predictions from GPT-3 generations, and introduce a novel way of aggregating predictions across multiple prompt templates to increase result coverage.

This work introduces a highly multilingual summarization corpus, available in all of the 24 official languages of the European Union. It is based on legal acts published by the EU, and consists of extremely long documents in the legal domain.

We present an interactive interface, unifying the access to several temporal annotation frameworks. Aside from a graphical interface, we allow users to programmatically access the various tools through a streamlined API interface.

We present a high-quality resource of full-text alignments between German Wikipedia, and a German children’s encyclopedia. This yields a dataset that we empirically show to be suited for both summarization and simplification tasks.

Following a series of prior experiments in English, we outline a generalized training procedure for training a non-English temporal tagger with weakly supervised data. For this purpose, we utilize existing rule-based taggers as a way to scale up existing training resources in low-resource settings by several orders of magnitude.

In this (German) article, we outline the challenges that are currently preventing mainstream adoption of recent NLP advancements in the legal industry. Primarily, this can be attributed to a lack of proper domain generalization, as well as limited interpretability and scalability of such models.

We experimented with various transformer-based architectures to see which ones would work best for extracting temporal annotations, such as ‘yesterday’ or ’every week’. However, we have since found a significant flaw in our evaluation setup for seq2seq-based models, so we decided to retract this article. Resulting tagging-based models are still valid, though, and are available online.

Utilizing existing segmentation tools, which primarily operate on sentence-level granularity, yields poor performance when segmenting long documents, which are prevalent in a legal context. In this work, we address the issue by proposing a weakly-supervised paragraph-based segmenter, which we empirically show on a novel dataset consisting of web Terms of Service documents.

We participated in the workshop’s shared task on extracting relevant paper sections in cited works. Interestingly, we show that our setup based on traditional search heuristics, coupled with improved pre-processing steps, outperforms our BERT-based retrieval setup. Overall, we placed third on the blind shared task test set.

This demonstration illustrates a time-centric approach to content exploration. Extracting and processing temporal mentions on large document collections allows a temporal expression of events, even when the documents themselves are not ordered chronologically.

Results from my Master’s thesis contributed the experimental section of this work. Primarly, we present a theoretical retrieval model based on hypergraphs, and demonstrate that these operations can be utilized to perform common Information Retrieval operations more efficiently on co-occurrence based word networks than traditional dyadic graphs.

Ordering events in a chronological fashion requires a accurate modeling of the temporal hierarchy, which previously was not well-defined for long-term event horizons spanning several decades. Here, we present a temporal model that is shown to work well, even without explicit temporal ordering in underlying document collections.

In a collaboration with molecular biosciences, we developed an automated Image Processing pipeline that was able to speed up the annotation process and accuracy for determining lengths of chromatosome strands in AFM images. In this work, it was ultimately shown how mutations in a particular gene can cause different winding patterns.