Cornelia Ferner

Ms. Cornelia Ferner

orcid

Research Focus & Metrics

Academic Career Length

10Years (2017-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
25 July 2020 in Information
Verified
Journal Article
Open access
PDF file
CFCornelia Ferner
Cornelia Ferner
CHClemens Havas
Clemens Havas
EBElisabeth Birnbacher
Elisabeth Birnbacher
SWStefan Wegenkittl
Stefan Wegenkittl
Bernd ReschBernd Resch
Bernd Resch
In the event of a natural disaster, geo-tagged Tweets are an immediate source of information for locating casualties and damages, and for supporting disaster management. Topic modeling can help in detecting disaster-related Tweets in the noisy Twitter stream in an unsupervised manner. However, the results of topic models are difficult to interpret and require manual identification of one or more “disaster topics”. Immediate disaster response would benefit from a fully automated process for interpreting the modeled topics and extracting disaster relevant information. Initializing the topic model with a set of seed words already allows to directly identify the corresponding disaster topic. In order to enable an automated end-to-end process, we automatically generate seed words using older Tweets from the same geographic area. The results of two past events (Napa Valley earthquake 2014 and hurricane Harvey 2017) show that the geospatial distribution of Tweets identified as disaster related conforms with the officially released disaster footprints. The suggested approach is applicable when there is a single topic of interest and comparative data available.

Verified
Journal Article
Open access
PDF file
CHClemens Havas
Clemens Havas
LWLorenz Wendlinger
Lorenz Wendlinger
JSJulian Stier
Julian Stier
SJSahib Julka
Sahib Julka
VKVeronika Krieger
Veronika Krieger
Bernd ReschBernd Resch
Bernd Resch
In 2015, within the timespan of only a few months, more than a million people made their way from Turkey to Central Europe in the wake of the Syrian civil war. At the time, public authorities and relief organisations struggled with the admission, transfer, care, and accommodation of refugees due to the information gap about ongoing refugee movements. Therefore, we propose an approach utilising machine learning methods and publicly available data to provide more information about refugee movements. The approach combines methods to analyse the textual, temporal and spatial features of social media data and the number of arriving refugees of historical refugee movement statistics to provide relevant and up to date information about refugee movements and expected numbers. The results include spatial patterns and factual information about collective refugee movements extracted from social media data that match actual movement patterns. Furthermore, our approach enables us to forecast and simulate refugee movements to forecast an increase or decrease in the number of incoming refugees and to analyse potential future scenarios. We demonstrate that the approach proposed in this article benefits refugee management and vastly improves the status quo.

27 September 2019 in Energy Informatics
Verified
Journal Article
Open access
PDF file
CFCornelia Ferner
Cornelia Ferner
GEGünther Eibl
Günther Eibl
AUAndreas Unterweger
Andreas Unterweger
SBSebastian Burkhart
Sebastian Burkhart
SWStefan Wegenkittl
Stefan Wegenkittl
The nationwide rollout of smart meters in private households raises privacy concerns: Is it possible to extract privacy-sensitive information from a household’s power consumption? For a small sample of 869 Upper Austrian households, information about consumption-heavy amenities and household characteristics are available. This work studies the detection of households with swimming pools (the most common amenity in the dataset) using Convolutional Neural Networks (CNNs) applied on load heatmaps constructed from load profiles. Although only a small dataset is available, results show that by using CNNs, privacy can be broken automatically, i.e., without the time-consuming, manual feature generation. The method even slightly outperforms a previous approach that relies on a nearest neighbor classifier with engineered features.

Verified
Journal Article
Open access
PDF file
CFCornelia Ferner
Cornelia Ferner
SWStefan Wegenkittl
Stefan Wegenkittl
Representations from common pre-trained language models have been shown to suffer from the degeneration problem, i.e., they occupy a narrow cone in latent space. This problem can be addressed by enforcing isotropy in latent space. In analogy with variational autoencoders, we suggest applying a token-level variational loss to a Transformer architecture and optimizing the standard deviation of the prior distribution in the loss function as the model parameter to increase isotropy. The resulting latent space is complete and interpretable: any given point is a valid embedding and can be decoded into text again. This allows for text manipulations such as paraphrase generation directly in latent space. Surprisingly, features extracted at the sentence level also show competitive results on benchmark classification tasks.


1 July 2022
Verified
Conference Paper
Open access
PDF file
CFCornelia Ferner
Cornelia Ferner
The artworks by Dutch painter Hieronymus Bosch are well known for their incredible wealth of details. The popular BoschBot regularly posts small segments of the digitized paintings on Twitter, thus relieving their density and making them more accessible. CaptioningBoschBot, the Twitter bot presented in this demo, reverses the creative process of the artist: It uses the out-of-context painting segments as input for an encoder-decoder model to generate captions that interpret the painted objects. As the model was only trained on realistic, photographic images, curious interpretations of the otherworldly details can be observed. The generated captions are again posted on Twitter to encourage discussions about Bosch's masterpieces and the AI technology in general.