ACQuA 2.0: Answering Comparative Questions with Arguments
In the second phase of the ACQuA project in the DFG-SPP RATIO we continue our work on arguments in comparison scenarios. Comparisons are an interesting and valuable subject of research as people often need to compare different options and need argumentative explanations of pros and cons to come up with an informed own opinion.
Confronted with a question like 'Is X better than Y with respect to Z?', even the big commercial web search engines today often only show facts from a knowledge base that match 'X', 'Y', or 'Z', or they show some related questions mined from question answering websites as kind of a direct answer above the classic ten blue links. One of the main issues---besides the usual dilemma of short direct answers not being able to cover all angles of more complex information needs---is that mining question answering websites and extracting facts from knowledge bases can only help to answer a minority of comparative questions. By analyzing large Yandex query logs in our first project phase, we have shown that most real comparative questions need argumentative explanations that can be found on the web, but not necessarily on question-answering websites. We thus developed the comparative argumentation machine CAM that retrieves sentences from the web related to a comparative information need. We could show that CAM users are faster and more correct than users of a traditional keyword-based search. To identify the most argumentative sentences (and text passages), we developed the neural argument tagger TARGER, and we started to work on distinguishing comparative questions that can be answered by facts from questions that need argumentative answers. Some pilot studies in the first ACQuA phase showed promising preliminary results for extracting facts from a knowledge base to answer factual comparisons. Further, we have co-organized the Touché lab at CLEF 2020: the first shared task on argument retrieval; we will continue to co-organize it in the coming years.
In the second project phase, we plan to (1) strengthen the connection of knowledge bases and textual evidence as sources for argumentative answers on comparative questions, to (2) diversify arguments via axiomatically re-ranking search results from web-scale multilingual and multicultural sources for a broad coverage of different viewpoints (e.g., for "controversial" topics like programming languages or cuisine), and to (3) develop prototypes to embed summarized argumentative answers on natural language comparative questions in applications with interactive conversational interfaces, especially paying attention to argument provenance to ensure a high transparency for the user.