Framing in Argument Analysis & Generation. Detecting and generating perspectives

Valuable cooperation and democracy require listening to each other, expressing opinions, and debating with each other by tolerating each other. Since humans are different due to varying experiences, knowledge, assumed axioms and faith, reasoning strategies, and divergent intentions and goals, they have diverse perspectives on a topic and frame content differently. Detecting these frames is important to evaluate a debate and detect biases, opening the possibility of broadening the field for deliberation and, therefore, is one way of overcoming the “tyranny of the majority”. However, detecting the farming in texts and opening a debate is not trivial due to a level of latency in that task. Expressing a frame can be done by a bandwidth of words, naming for certain groups, or even more indirectly by concealing other valid viewpoints of problems or solutions to an issue. Therefore, frame identification and framed text generation is a field that requires a deep understanding of the text, which we tackle in this thesis.<br /><br /> Hence, we approach this field with deep learning, experimenting with a bandwidth of Neural Net models, from an effective and lightweight aggregation module that analyzes word embeddings, to Reccurent Neural Nets, and finally to frame-tailored Large Languge Models which are proven to be the most successful models in Natural Language Processing due to their contextualized word understanding. Employing this range of techniques, we approach the task of frame identification in newspapers and arguments and framed text generation in terms of framed conclusion generation and reframing. To refine frame identification, we present a trained multi-task setup between framing in newspaper and framing in arguments in a supervised manner and employ an ensemble approach combining three different modules. To set up framed text generation, we enhance LLMs by a novel frame-tailored regularization technique while fine-tuning and applying framed deciding and output reranking at inference, guiding the generated text into a particular (set of) frames. For reforming, we overcome the lack of training data released to that specific subtask in the field of computational argumentation by methods from the field of eXplainable Artifical Intelligence.<br /><br /> Our findings give insight into the generalizability of classifiers adjusted to a set of framed training data. We successfully show the possibility of broadening this ability with our multi-task setup combined with soft-parameter sharing. To address multilingualism, we demonstrate the suitable performance of translation services using artificial intelligence and retrieve, for some languages, more consistent results by combining different approaches and incorporating common-sense knowledge. Our findings for framed text generation introduce the wide possibilities by reranking generated texts with automatic metrics involving frame classifiers and deep text evaluation methods guided by the findings from XAI. With comprehensive manual studies, we also point to a trade-off between validity and novelty in the task of conclusion generation given the premise of an argument while ensuring the desired frame in up to 90% of the cases. Based on these findings, our main contributions include the following: 1. We develop several frameworks for the classification task of frame identification, involving a multi-task setup and a dynamic ensemble approach, and for the task of framed text generation, including framed decoding and several approaches for text reranking 2. For frame identification, we perform a broad bandwidth of experiments, evaluated with a set of automatic methods. We present a method mapping sparse issue-specific frames to denser generic framesets, enabling the training and comprehensive evaluation on the argumentative dataset only providing those sparse issue-specific frames. Furthermore, we introduce a lightweight neuronal model and an approach that incorporates common-sense knowledge to detect frames on a global level. 3. We apply our insights on frame identification to an open debate portal to automatically classify the user-generated argumentative opinions there. We show a correlation between frames, frame-related concepts, and human values. We perform a case study to automatically detect biased debates and debates that lack relevant perspectives. 4. For framed text generation, in addition to a comprehensive automatic evaluation, we present a detailed manual evaluation analyzing the frames and plausibility of the generated texts. For conclusion reranking, we develop an annotation method analyzing the validity and novelty of a conclusion, resulting in a dataset for measuring the quality of conclusions.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bielefeld

Year

2024

Author dc:creator

• Heinisch, Philipp

Repository record source_url

https://pub.uni-bielefeld.de/record/3000218

OAI identifier oai:identifier

oai:pub.uni-bielefeld.de:3000218