This paper is a report on the first Text Mining and Generation Workshop (TMG), which was a one-day virtual event hosted at the German Conference on Artificial Intelligence (KI 2022) in Trier, Germany. In addition to four accepted original papers, there were three invited talks by speakers who presented their works already published at high-ranked conferences as well as one keynote by a pioneer in the two research fields relevant to the workshop.
Digital text data is available in large amounts and diferent granularities. Typical sources of this data are social media posts, books, news articles, web pages, company reports, etc. A major challenge this text data imposes is that it is unstructured and has to first be processed to make further analysis possible. At the same time, there are also many situations in which only structured data is available that is to be verbally explained, for instance, by Explainable AI. These contrasting scenarios lead to two complementary application areas: text mining and text generation. The aim of text mining is to analyze the content of unstructured text and extract (useful) structured information. In contrast, text generation attempts to (automatically) create text from structured information or knowledge that is for example stored in large language models. The goal of the TMG workshop is to bring these two perspectives together by eliciting research paper submissions that aim for bridging the gap between knowledge extraction and text generation. Since recent approaches to text mining and text generation are predominantly based on artificial intelligence (AI) methodologies, KI 2022 has been a great venue to bring together AI researchers working on these two tasks.
In total, we accepted four papers for publication. In the following, we provide a brief overview of their main contributions. The paper “German to English: Fake News Detection with Machine Translation” applies translation (essentially a text generation task) to mitigate the fact that languages other than English mostly have worse ML models available. Their evaluation shows that translating such texts beforehand and then using the better English models is a valuable processing step. “IRT2: Inductive Linking and Ranking in Knowledge Graphs of Varying Scale” is concerned with knowledge graph completion based on natural language text and thus covers the text mining task. The authors propose two models for predicting links in knowledge graphs and provide initial results based on an experimental evaluation. The third paper “Explaining Hate Speech Classification with Model-Agnostic Methods” proposes a pipeline to create interpretable explanations for black-box models like BERT classifiers. By leveraging text generation, they aim to assist users in detecting hate speech in a model-agnostic way. Finally, “Comparing Unsupervised Algorithms to Construct Argument Graphs” is concerned with generating relations between argumentative statements that are extracted from plain texts.
A keynote talk titled “Detect—Verify—Communicate: Combating Misinformation with More Realistic NLP” was given by Prof. Dr. Iryna Gurevych (Technical University (TU) of Darmstadt, Germany), who addressed the omnipresent problem of misinformation in our society, and elaborated on how to identify and debunk misinformation. In particular, she addressed the fact that current NLP systems cannot be used for fact-checking in real-life scenarios and presented her work where they are exploring solutions for this. While the main problems reside with resources, their research is dedicated to the most harmful, novel examples. Besides two corpora constructed for this purpose, they compare the capabilities of automated NLP-based approaches to the methods people use for fact-checking. In order to include diferent perspectives, they collaborate with cognitive scientists and psychologists.
At the workshop, we also had three invited talks to spark a discussion and exchange of
knowledge and ideas. We selected three previously published papers at top-tier conferences (EACL,
SIGIR, and WWW) that cover the topics of text mining and generation [
In the first Text Mining and Generation Workshop, a total of four papers were accepted for publication that addressed various topics such as fake news, knowledge graphs, explanations, and relation generation in the context of text mining and generation. In addition, our keynote and invited talks allowed students and young scientists to connect with experienced researchers, ask questions, and participate in discussions.
We expect the topics of this workshop will be becoming increasingly important in the future,
particularly with the emergence of new large language models (LLMs). The significant
advancements in (retrieval-augmented) text generation introduced by LLMs like GPT-3, ChatGPT,
and BLOOM make it crucial that we extensively study their impact on downstream tasks and
benchmark key findings. On the one hand, for example, GPT-3 has been found to outperform
all supervised models on the news summarization task [
This work has been funded by the DFG within the projects “ReCAP-II” (No. 375342983) and “ACQuA 2.0: Answering Comparative Questions with Arguments” (No. 376430233) as part of the priority program RATIO (SPP-1999) as well as the “Studienstiftung”.