Biomedical Named entity recognition (BioNER) is a crucial initial step for biomedical information processing, serving as the cornerstone technology for identifying biomedical entities and their interactions. However, the inherent workings of most BioNER models remain opaque to users, posing a significant challenge in enhancing their interpretability. In this paper, we propose a novel interpretable learning method based on Talmudic Public announcement logic (TPK), which can learn human-readable knowledge from Bi-directional Long Short-Term Memory (BLSTM) models and generate effective explanations in the form of dichotomous TPK trees. Empirical evaluations conducted on the publicly available BioNER dataset, GENIA, demonstrate that our TPK-based deep learning method outperforms the vanilla BLSTM models in low-source settings. Furthermore, the logical reasoning based on TPK models shows how BLSTM handles BioNER tasks in real applications, providing users with rigorous and transparent logical justifications.
With the rapid growth of digitization and multimedia, multimodal argumentation in advertisements has become a key focus in pragmatics and argumentation theory. Print advertisements, with their concise and elliptical communication, often rely on images to build logical coherence and practical effectiveness, presenting a unique challenge for multimodal argumentation research. In response, this paper combines a pragmatic perspective with argumentation theory to analyze the visual and linguistic elements as well as argumentation schemes in advertisements. It proposes a method for analyzing and evaluating multimodal arguments, particularly double-mode arguments in advertisements. This method consists of three steps: (1) the pragmatic interpretation of visual and verbal information in advertisements, (2) the reconstruction of argumentation structures including the typical argumentation schemes and their connections in print advertisements, and (3) the evaluation of reconstructed arguments based on argumentation scheme theory. By applying this method to the case study of the well-known commercial advertisement for the Sauce-flavored Latte, it offers an analytical framework for the systematic analysis and evaluation of double-mode argumentation in advertisements, thereby providing theoretical support and practical guidance for optimizing advertising design, fostering critical thinking in consumers, and identifying false advertisements.
The modal μ -calculus significantly extends the expressive power of basic modal logic, by adding fixed-point operators. In this paper, we study the extension of public announcement logic with such fixed-point operators. Besides the general increased expressive power, this will in particular allow us to reason about self-referential announcements, as in “after this very announcement, φ will be true”. Such self-referential announcements have been of recent interest in the study of the so-called surprise exam paradox. However, a straightforward combination of the two logics would rule out formulas expressing such self-referential announcements due to the restriction of variables in fixed-point operators to be positive, which ensures that the corresponding function is monotonic which again ensures that it has a greatest fixed-point. We argue, however, that first, even without the positivity requirement the function might still be monotonic in a given model, and, second, that greatest fixed-points might exist even if the corresponding function is not monotonic. We propose an extension of public announcement logic with generalised fixed-point operators, without restricting variables to be positive, and take some first steps in analysing it. We show that the logic can be reduced to the logic without public announcement operators: to what we call “full” modal fixed-point logic. We also show that the logic is strictly more expressive than the standard modal fixed-point logic. The logic offers a simpler way of modeling self-referential announcements in the surprise exam paradox, than existing approaches.
We propose the combination of two known extensions of Dung's abstract Argumentation Frameworks (AFs): Incomplete Argumentation Frameworks (IAFs), where the existence of arguments and attacks may be uncertain, and Higher-Order Argumentation Frameworks (HOAFs), where attacks can be directed not only towards arguments but also towards attacks. As a result, we obtain Incomplete Higher-Order Argumentation Frameworks (IHOAFs), where attacks on attacks can also be uncertain. We take some important initial steps in the formal understanding of this novel framework by providing: (i) the definition of a conservative notion of completion, the hypothetical removal of uncertainty used to reason about argument acceptability in other formalisms for incomplete argumentation; (ii) complexity results for argument acceptability problems; (iii) an analysis of the previous two points in a well-behaved variant of dynamic logic.
