Computational Argumentation

Computational argumentation deals with the computational analysis and synthesis of natural language arguments and argumentation, usually in an empirical data-driven manner. Computational argumentation research that members of the NLP Group have particularly contributed to includes the following. Many of the outcomes of the group are or will be demonstrated in the argument search engine args.me.

The NLP Group has have worked on argument summarization, starting with the extraction of the most important claim and reason (SIGIR 2020), the contrasting of different views (COMMA 2022), and the reduction of a long argument down to its conclusion (ACL Findings 2021a). We have also used contrastive learning to reconstructed implicit conclusions of arguments (ACL 2020). 

In various works, we tackled controlled argument generation. We developed methods to integrate knowledge in neural argumentation generation (ACL 2021a), to encode into beliefs into arguments (EACL 2021) and to adjust arguments to the morals of the audience (ACL 2022a). To generate counterarguments, we computationally undermined arguments by attacking their weakest premises (ACL Findings 2021b). Additionally, we show the importance of inferring the argument conclusion in order to successfully counter it (EACL 2023). 

In another line of research, we modeled the human synthesis process following rhetorical strategies (COLING 2018), and operationalized it in a computational model (INLG 2019).  Finally, we also have looked at how to use argument generation to assess argument quality (ArgMining 2021) and improve arguments through revision (INLG 2023).

Starting from a literature survey, we defined a taxonomy of argumentation quality (EACL 2017), followed by an empirical comparison of theory and practice (ACL 2017). Later, we developed computational approaches to assess all 15 dimensions covered in these studies (COLING 2020), 

With respect to the specific dimensions, we investigated the impact of argument mining on argumentation-related essay scoring (COLING 2016), adapted PageRank for argument relevance (EACL 2017), and modeled relevance also in retrieval contexts (SIGIR 2019). Further, we developed a new dialectical notion of argument quality (CoNLL 2018) and assessed it for different target audiences (ACL 2020, PEOPLES 2020).

We also explored the detection of ad-hominem fallacies (NAACL 2018). Recently, we have started to study what can be learnt about quality from argument revisions (EACL 2021) and how to employ the generation capabilities of large language models in quality assessment (ArgMining 2021) and automated revisioning (ACL 2023a). In our project OASiS, we model appropriate language in argumentation (ACL 2023b) and seek to generate improved versions of inappropriate arguments.

With args.me, we introduced the first search engine for arguments on the web (ArgMining 2017). We studied visualizations for argument search (EMNLP 2018), data acqisition paradigms (KI 2019, best paper award), the use of voice as input (CHIIR 2020), and the cognitive biases in argument search (CUI 2021). 

Moreover, we developed a computational approach to retrieve the best counterarguments to arguments based on their simultaneous similarity and dissimilarity (ACL 2018). The above-mentioned PageRank for argument relevance (EACL 2017) may be used for ranking found arguments. 

As part of the research on argumentation generation, we also explored  in two works how to use argument summaries for snippet generation (SIGIR 2020, COMMA 2022). 

We have studied argumentation strategies based on a news editorial corpus with fine-grained evidence annotations (COLING 2016), we trained a classifier and used it to find topic-specific evidence patterns (EMNLP 2017), and we analyzed deliberative strategies in dialogical argumentation on Wikipedia talk pages (ACL 2018). We also modeled the general idea of framing in argumentation (EMNLP-IJCNLP 2019).

Furthermore, we developed computational models of the sequential flow of review argumentation (COLING 2014) based on an annotated corpus (CICLing 2014). Mapping a text into the feature space of overall structures, sentiment flows predict review sentiment robust across domains (EMNLP 2015). Later, we generalized the flow model into a universal model for discourse-level argumentation analysis across text genres and prediction tasks (ACM TOIT 2017), and we presented a novel tree kernel-based approach to capture joint sequential and hierarchical structure (EMNLP 2017).

Other research topics include cross-domain argument mining (NAACL 2016, ArgMining 2017, ACL 2021b), argument reasoning comprehension (NAACL 2018, SemEval 2018), the computational modeling of knowledge in argumentation (AAAI 2020, ACL 2021a, TACL 2022), and the role of human values (ACL 2022b).