This paper describes a long-term research goal which aims at creating a middleware interface between Argumentation Schemes and natural language. This idea comes from the need to face some challenges related to the automatic extraction of Argumentation Schemes from Natural Language: for example the ability to extract Argumentation Schemes at di erent level of granularity. In the paper we describe how this process can be designed and how the structures of Argumentation Schemes can be modeled to this aim.
Argumentation Schemes are stereotypical patterns of argumentative inferences
[
On the one side, Argumentation Schemes are important source of information
for Argument Mining. On the other side, however, the automatic extraction of
Argumentation Schemes (or their inner components) has been attempted only
in few studies [
Section 2 will describe the motivations behind this study. Section 3 will describe the main idea of this study: the combination of a top-down and a bottomup approach to exploit Argumentation Schemes' potential. Section 4 will describe some related works. Section 5 will conclude the paper.
Argumentation Schemes are stereotypical patterns of inference [
On the one side, a major obstacle is the fact that every-day natural language (like the one employed on internet comments or posts) has a very complex and variable ontological dimension; furthermore, it is often inferentially incomplete (information is often implicit or even incoherent). These two elements, i.e. ontological complexity and inferential incompleteness, can make it di cult to understand what schemes are actually employed in natural language. On the other side, Argumentation Schemes can o er a valuable interface between natural language and reasoning. In fact, they involve just a restricted ontological dimension (which is, however, too simple to catch all the possible expressions of the same scheme within human language). Moreover, they convey enough (although sometimes incomplete) inferential information to perform a formal evaluation on the argument. Argumentation Schemes o er, thus, a di erent scenario compared to every-day natural language: they are ontologically too simple and, potentially, they are inferentially incomplete.
These di erences between natural language and Argumentation Schemes regarding the ontological and the inferential dimensions, are the reason why it is di cult to leverage Argumentation Schemes knowledge directly from textual data. The long-term goal described in this paper is to create middleware inferential-ontological interfaces (called Argumentation Scheme Templates) where natural language complexity can be safely compressed, while Argumentation Scheme simplicity can be safely extended. We argue that the solution might be that of combining a top-down approach (from the layer of abstraction of Argumentation Schemes towards natural language) with a bottom-up approach (from natural language toward the layer of abstraction of Argumentation Schemes). The former aims at creating Argumentation Schemes templates (which can be designed to represent clusters of schemes with a variable degree of granularity). The latter aims at mapping pieces of natural language to the inner components of these templates (i.e., mapping natural language premises and conclusions to the templates' components).
This approach is described in Figure 1 and can be summarized in four points, two related to the top-down approach and two related to the bottom-up approach. Starting from the top-down, the rst aspect to consider is that the middleware templates should preserve all the ontological information of the original Argumentation Scheme. Usually schemes use stereotypical semantic-ontological expression, for example the rst premise of the Argument from Negative Consequences says that \if A is brought about, bad consequences will plausibly occur": this causal relation should be somehow represented in the template, as well as the entities which are pragmatically crucial for the scheme, i.e. an entity action (\A") and an entity outcome (\bad consequences"). A second aspect to consider is that, since Argumentation Schemes often represent stereotypical and incomplete ways of reasoning, the crucial inferential steps that are missing or implicit should be added. For example, taking in consideration the previous premise from the Argument from Negative Consequences, and its relative conclusion \Therefore A should not be brought about", one should notice that there is a missing inferential step: the warrant is missing (namely, the fact that whenever an action has negative consequences such action should not be brought about); the nal template must have a component for this missing inferential step. Table 1 provides a potential resulting template for the Negative Consequences scheme.
Regarding the bottom-up approach, the rst aspect to consider is that we need to build classi ers able to map (i.e., reduce) the complexity of natural language into the components of the Argumentation Schemes Template. For example, we can consider a Negative Consequences scheme like the sentence \Sending troops would provoke a war, so I think we should absolutely avoid it", our classi ers should be capable of mapping the piece \Sending troops would provoke a war" into the rst component of the Template of the scheme from Negative Consequences described in Table 1: doing(Action(\Sending troops")) =c=au=s)es OutcomeN egative(\war")). Similarly, the conclusion \so I would absolutely avoid it" might be designed as :isW anted(Action(\Sending troops")). Finally, a fourth element to consider is the fact that the ontological complexity of the original sentence will be compressed into the standard representative expressions of the chosen language. This point is crucial because Argumentation Schemes Templates should use a language which is carefully designed following pragmatics: for example, we chose functions like \doing", \isWanted" and entities like \Action" and \OutcomeNegative" because the pragmatic sphere of this Argumentation Scheme reasons about the goodness of doing or not an action. In this sense, pragmatics can guide us into the design of the language.
We designed :isW anted as the negation of a Defeasible Modus Tollens (DMT) in which the inferential negation (:) is intertwined with the semanticalontological sphere (isW anted): the DMT makes it possible that the negation :isW anted goes from the consequent OutcomeN egative(G) to the antecedent doing(Action(A)). This is how the ontological and the inferential dimension are intertwined, and why we envisage a language which can express this overlapping.
While some limitations of employing First-Order Logic languages to model
Argumentation Schemes have been rightly remarked in [
For the bottom-up approach, we envisage a combination of text classi cation and sequence labelling tasks. Text classi cation tasks can be used to cluster schemes while sequence labelling tasks can be used to select the spans of text that correspond to portions of template. However, the speci cations of the bottom-up approach ( lling templates) will necessarily depend on the speci cations of the top-down procedure (creating templates). 4
The studies which approached the task of extracting Argumentation Schemes
automatically resorted to highly engineered methodologies [
While a crucial e ort towards a high-level ontology has been provided by AIF
[
In this work, we described a long-term research direction aiming at facilitating the automatic extraction of Argumentation Schemes from textual data and the application of arti cial reasoners to natural language (using Argumentation Schemes knowledge). Being able to reason directly from textual data is an extremely challenging objective which is often made complicated by the fact that natural language has a huge ontological complexity and is often inferentially incomplete. In this regard, Argumentation Schemes are an appealing solution to this problem because they are an interface between natural language (ontologically complex) and the inferential dimension (ontologically very simple).
We shortly introduced a feasible direction to achieve this long-term goal, which envisages the combination of a bottom-up approach with a top-down approach. The former is an e ort to design Argumentation Schemes as quasi-logical templates (Argumentation Scheme Templates) composed of a logical language able to preserve the basic inferential and ontological information of schemes while following the pragmatic criteria of the scheme itself. The latter is an e ort to create classi ers able to map pieces of natural language into corresponding pieces of an Argumentation Scheme Template. Although this is a long-term goal, we believe that this direction can be valuable, and capable to leverage and maximize the argumentative knowledge conveyed by Argumentation Schemes.