Evaluation of text simplification currently focuses on the difference of a source text to its simplified variant. Datasets for this evaluation base on a specific topic and group of readers for which is simplified. The broad applicability of text simplification and specifics that come with intended target audiences (e.g., children compared to adult non-experts) are disregarded. An explainable assessment of the overall simplicity of text is missing. This work is BenchmArking Text Simplicity (BATS): we provide an explainable method to assess practical and concrete rules from literature describing features of simplicity and complexity of text. Our experiments on 15 datasets for text simplification highlight differences in features that are important in different domains of text and for different intended target audiences.

The dissertation investigates the correlations and effects between biases in search queries and search query suggestions, search results, and users’ states of knowledge. Search engines are an important factor in opinion formation, while search queries determine the information a user is exposed to in information search. Search query suggestions play a crucial role in what users search for [22]. Biased query suggestions can be especially problematic if a user’s information need is not set and the interaction with query suggestions is likely. Only recently, research has started to investigate the general assumption that biased search queries lead to biased search results, focusing on political stance bias [17]. However, the correlation between biases in search queries and biases in search results has not been sufficiently investigated. Sparse context and limited data access pose challenges in detecting biases in search queries. This dissertation thus contributes datasets and methodological approaches that enable media bias research in the field of search queries and search query suggestions.

This publication describes the motivation and generation of Qbias, a large dataset of Google and Bing search queries, a scraping tool and dataset for biased news articles, as well as language models for the investigation of bias in online search. Web search engines are a major factor and trusted source in information search, especially in the political domain. However, biased information can influence opinion formation and lead to biased opinions. To interact with search engines, users formulate search queries and interact with search query suggestions provided by the search engines. A lack of datasets on search queries inhibits research on the subject. We use Qbias to evaluate different approaches to fine-tuning transformer-based language models with the goal of producing models capable of biasing text with left and right political stance. Additionally to this work we provided datasets and language models for biasing texts that allow further research on bias in online information search.

Emojis represent an essential means of expressing sentiments such as opinions and attitudes in computer-mediated communication, especially in chats and social media. To effectively capture these sentiments, the sentiments associated with the emojis used must be known. Previous approaches to determining the sentiments expressed with emojis require a large amount of manual annotation. For many emojis, especially less frequently used platform-specific emojis, studies on expressed sentiments do not yet exist. Therefore, these emojis cannot be considered in sentiment analyses so far. In this work, a method for effective and efficient determination of emojis? sentiments and their compilation in a sentiment lexicon was developed. The determined sentiments are compiled as a sentiment lexicon. For this purpose, software was created in Python to process collections of texts into a corpus. The software derives the emojis? sentiments as valence values based on the sentiments of the texts in which the emojis appear. The lexicons produced by the method can be used in lexicon-based sentiment analysis approaches. The method also derives other information on the emojis and their usage that can be used to assess the sentiment lexicon produced and the usage of the emojis. Using the developed method, two analyses were conducted with corpora of different text sources. The results and subsequent comparisons with existing sentiment lexicons have shown that the developed method is able to efficiently produce similar results as sentiment lexicons produced with manual annotation.

In this paper, we describe the TH Köln's submission for the Shared Task on the Identification of Toxic Comments at GermEval 2021. Toxicity is a severe and latent problem in comments in online discussions. Complex language model based methods have shown the most success in identifying toxicity. However, these approaches lack explainability and might be insensitive to domain-specific renditions of toxicity. In the scope of the GermEval 2021 toxic comment classification task (Risch et al., 2021), we employed a simple but promising combination of term-frequency-based classification and rule-based labeling to produce effective but to no lesser degree explainable toxicity predictions.