Abstract
One way to solve the knowledge acquisition bottleneck is to have ways to translate natural language sentences and discourses to a formal knowledge representation language, especially ones that are appropriate to express domain knowledge in sciences, such as Biology. While there have been several proposals, including by Montague (1970), to give model theoretic semantics for natural language and to translate natural language sentences and discourses to classical logic, none of these approaches use knowledge representation languages that can express domain knowledge involving normative statements and exceptions. In this paper we take a first step to illustrate how one can automatically translate natural language sentences about normative statements and exceptions to representations in the knowledge representation language Answer Set Programming (ASP). To do this, we use λ-calculus representation of words and their composition as dictated by a CCG grammar.
Introduction
A significant hurdle in Artificial Intelligence is the "knowledge acquisition bottleneck"—the high cost and effort required to manually encode human knowledge into a machine-readable format. While natural language is the primary medium for human knowledge, translating it into formal logic is difficult, particularly when dealing with "normative" statements (rules that are generally true) and their "exceptions."
Traditional approaches often rely on First-Order Logic (FOL). However, FOL is monotonic: once a fact is established, it cannot be retracted. This makes it unsuitable for common-sense reasoning where new information (e.g., "Penguins are birds that do not fly") might contradict general rules (e.g., "Birds fly"). To address this, non-monotonic logic is required to allow for defeasible reasoning—rules that can be overridden by more specific evidence.
Methodology
The proposed framework bridges the gap between linguistics and non-monotonic logic by integrating three components:
- Answer Set Programming (ASP): A declarative logic programming language designed for complex search and common-sense reasoning. ASP natively supports "negation as failure," allowing the system to assume a property holds unless explicit evidence suggests otherwise.
- Combinatory Categorial Grammar (CCG): A compositional grammar formalism where words act as functions. This allows for a direct mapping from syntactic structure to semantic meaning.
- λ-ASP-expressions: The core innovation of this work. The authors extend standard λ-calculus to create expressions where the meaning of a word is a function that constructs an ASP rule.
The workflow involves parsing a sentence with CCG, assigning λ-ASP-expressions to words, and composing these expressions via β-reduction to generate a set of ASP rules.
The semantic mapping is defined at the lexical level. For instance, the word "most" is mapped to a template that generates a default rule in ASP, allowing for the representation of normative statements.
Results and Discussion
The system was evaluated using a targeted dataset, $L_{bird}$, containing sentences that test various logical relationships, such as subclassing ("Penguins are birds") and strong exceptions ("Penguins do not fly").
A key insight is the representation of defaults. The sentence "Most $c$ have property $p$" is translated into the ASP rule: p(X) :- c(X), not -p(X). This effectively instructs the reasoner to assume property $p$ for any member of class $c$ unless a contradiction is found. The system successfully distinguishes between "negation as failure" (lack of knowledge) and "classical negation" (explicit knowledge of the contrary), which is vital for resolving conflicts between general rules and specific exceptions.
Qualitative results demonstrate that the system correctly handles complex scenarios. For example, it can conclude that a generic bird flies while simultaneously maintaining that a specific penguin does not, without triggering a logical contradiction. This successfully automates the creation of non-monotonic theories from natural language, a task that previously required significant manual labor by logic experts.
Conclusion
This work provides a proof-of-concept for the automated translation of natural language into non-monotonic theories. By leveraging CCG and λ-ASP-expressions, the authors demonstrate a path toward reducing the manual labor involved in knowledge engineering. Future research directions include expanding the grammar to support more complex structures like conjunctions and auxiliary verbs, and exploring methods to automate the acquisition of λ-ASP templates for a broader vocabulary.
Summary updated