Argumentative Frame Retrieval in the Climate Change Debate

Disentangling and understanding the main statements from an argumentative piece of text is an endeavour that is still far removed from the abilities of state-of-the-art AI systems. In fact, it even remains a hurdle for many undergraduate students, therefore resorting to reading comprehension courses. Today’s off-the-shelf NLP methods rely on small-window statistical co-occurrence counts or local information-theoretic measures, and remain therefore incapable of capturing an author’s claims and arguments. We refer to such existing techniques as distant reading tools, and contrast them to so-called close reading tools that go beyond statistical distributions in the training corpus but instead make use of the syntactic structure and the semantic frames underlying a specific sentence or paragraph.

In our talk, we propose a close reading method that makes use of Fluid Construction Grammar, a parsing and production engine that chunks together form-meaning mappings (i.e. constructions) into rich feature structures. It uses the exact same engine and constructions for building up an interpretation of a sentence as well as for composing an utterance based on a conceptualized meaning. We specifically show how such an engine can be used with a set of English constructions that are designed to extract argumentation frames from sentences (or paragraphs), through the theory of Frame Semantics. For instance, the Causation frame, which has frame elements such as an Actor and an Affected, could be detected by our system in sentences such as “Oxygen levels in oceans have fallen 2% in 50 years due to climate change.” (The Guardian, 20 Feb. 2017). Once these frames have been retrieved, and their frame elements have been filled in by pieces of the input text, there are two further steps that can be taken. In the first one, which we are currently exploring, the populated frames are returned to the human reader and can be used as a reading aid to grasp the main arguments and their relationship from a text. In a future second step, the AI system itself should be able to interpret the information present in the frames, based on their instantiation in Linked Data.