We present the first dataset and evaluation results on a newly defined task: assigning trigger warnings. We introduce a labeled corpus of narrative fiction from Archive of Our Own (AO3), a popular fan fiction site, and define a document-level classification task to determine whether or not to assign a trigger warning to an English story. We focus on the most commonly assigned trigger type “violence’ using the warning labels provided by AO3 authors as ground-truth labels. We trained SVM, BERT, and Longfomer models on three datasets sampled from the corpus and achieve F1 scores between 0.8 and 0.9, indicating that assigning trigger warnings for violence is feasible.
A trigger warning is used to warn people about potentially disturbing content. We introduce trigger warning assignment as a multi-label classification task, create the Webis Trigger Warning Corpus 2022, and with it the first dataset of 1 million fanfiction works from Archive of our Own with up to 36 different warnings per document. To provide a reliable catalog of trigger warnings, we organized 41 million of free-form tags assigned by fanfiction authors into the first comprehensive taxonomy of trigger warnings by mapping them to the 36 institutionally recommended warnings. To determine the best operationalization of trigger warnings, we explore state-of-the-art multi-label models, examining the trade-off between assigning coarse- and fine-grained warnings, open- and closed-set classification, document length, and label confidence. Our models achieve micro-F1 scores of about 0.5, which reveals the difficulty of the task. Tailored representations, long input sequences, and a higher recall on rare warnings would help.
At least 5% of questions submitted to search engines ask about cause-effect relationships in some way. To support the development of tailored approaches that can answer such questions, we construct Webis-CausalQA-22, a benchmark corpus of 1.1 million causal questions with answers. We distinguish different types of causal questions using a novel typology derived from a data-driven, manual analysis of questions from ten large question answering (QA) datasets. Using high-precision lexical rules, we extract causal questions of each type from these datasets to create our corpus. As an initial baseline, the state-of-the-art QA model UnifiedQA achieves a ROUGE-L F1 score of 0.48 on our new benchmark.
The paper presents first results of an ongoing project on text simplification focusing on linguistic metaphors. Based on an analysis of a parallel corpus of news text professionally simplified for different grade levels, we identify six types of simplification choices falling into two broad categories: preserving metaphors or dropping them. An annotation study on almost 300 source sentences with metaphors (grade level 12) and their simplified counterparts (grade 4) is conducted. The results show that most metaphors are preserved and when they are dropped, the semantic content tends to be preserved rather than dropped, however, it is reworded without metaphorical language. In general, some of the expected tendencies in complexity reduction, measured with psycholinguistic variables linked to metaphor comprehension, are observed, suggesting good prospect for machine learning-based metaphor simplification.
Most work on segmenting text does so on the basis of topic changes, but it can be of interest to segment by other, stylistically expressed characteristics such as change of authorship or native language. We propose a Bayesian unsupervised text segmentation approach to the latter. While baseline models achieve essentially random segmentation on our task, indicating its difficulty, a Bayesian model that incorporates appropriately compact language models and alternating asymmetric priors can achieve scores on the standard metrics around halfway to perfect segmentation.
n this paper we investigate the potential of answer clustering for semi-automatic scoring of short answer questions for German as a foreign language. We use surface features like word and character n-grams to cluster answers to listening comprehension exercises per question and simulate having human graders only label one answer per cluster and then propagating this label to all other members of the cluster. We investigate various ways to select this single item to be labeled and find that choosing the item closest to the centroid of a cluster leads to improved (simulated) grading accuracy over random item selection. Averaged over all questions, we can reduce a teachers workload to labeling only 40% of all different answers for a question, while still maintaining a grading accuracy of more than 85%.
We present a new corpus of tutorial dialogs on mathematical theorem proving that was collected in a Wizard-of-Oz setup. Our study is a follow up on a previous experiment conducted in a similar simulated environment. A major difference between the current and the previous experimental setup was that in this study we varied the presentation of the study-material with which the subjects were provided. One sub-group of the subjects was presented with a highly formalized presentation consisting mainly of formulas, while the other with a presentation mainly in natural language. Our goal was to obtain more data on the kind of mixed-language that is characteristic of informal mathematical discourse. We hypothesized that the language style of the subjects' interaction with the simulated system will reflect the style of presentation of the study-material. In the paper we briefly present the experimental setup, the corpus, and a preliminary quantitative result of the corpus analysis.