Attempto Controlled English (ACE) is a controlled natural language designed for unambiguous knowledge representation and reasoning. It allows users to write formal specifications, queries, and rules in a subset of English that can be automatically translated into first-order logic (FOL) and other formal languages. ACE was developed at the University of Zurich, with its origins in the late 1990s, and has since been used in various domains including software specification, ontology engineering, and question answering. The primary goal of ACE is to bridge the gap between human-readable text and machine-interpretable logic, making formal methods accessible to non-logicians while preserving precision.
ACE is not a free-form language; it is a strictly defined subset of English with a fixed grammar and vocabulary. Every sentence in ACE must conform to a predefined set of syntactic rules, which ensures that each sentence has exactly one logical interpretation. This eliminates the ambiguity inherent in natural language, such as scope ambiguities, anaphora resolution, and lexical polysemy. For example, the sentence "Every customer who owns a car likes it" is unambiguous in ACE, with the pronoun "it" clearly referring to the car owned by the customer. The language supports constructs like universal and existential quantification, negation, conditional statements, and modal verbs, all expressed in a way that mirrors everyday English.
History and Development
The development of ACE began in the late 1990s at the Institute of Computational Linguistics at the University of Zurich, led by Norbert E. Fuchs and his team. The first version, ACE 1.0, was released in 1997, and it was designed as a tool for knowledge representation in the context of the Attempto project, which aimed to create a natural language interface for logic programming. Over the years, ACE evolved through several versions, with ACE 6.0 being a notable milestone in 2007 that introduced a more comprehensive grammar and improved translation to first-order logic. The language was also extended to support discourse representation structures, allowing for multi-sentence texts. In 2010, the Attempto project released ACE 6.1, which included a parser and a translator that could generate FOL, and later versions added support for OWL (Web Ontology Language) and SWRL (Semantic Web Rule Language). The development was active until around 2015, after which the project was largely maintained as open-source software.
Syntax and Semantics
ACE's syntax is defined by a context-free grammar that specifies the allowed sentence structures. The grammar covers declarative sentences, questions, and imperative forms, but with strict restrictions on word order and phrase types. For instance, adjectives must precede nouns, and relative clauses must follow the noun they modify. The vocabulary is also constrained; only certain common English words are allowed, and all terms must be defined in a lexicon. The semantics of ACE are compositional, meaning that the meaning of a sentence is built from the meanings of its parts. The translation to FOL is deterministic, using a set of rules that map syntactic structures to logical formulas. For example, the sentence "No dog barks" translates to ¬∃x(dog(x) ∧ bark(x)). ACE also supports anaphoric references, but with a rule that a pronoun must refer to the most recent compatible antecedent, which is a simplification of natural language discourse.
Applications and Use Cases
ACE has been applied in several practical domains. In software engineering, it has been used to write formal specifications for system requirements, allowing stakeholders to validate requirements without needing to understand formal logic. For example, the specification "If a user is logged in and clicks the delete button, then the system deletes the user's account" can be written in ACE and automatically checked for consistency. In the Semantic Web, ACE has been used to author OWL ontologies, where domain experts can express axioms in a readable form. The ACE parser can generate OWL files, which are then used in reasoning systems. Additionally, ACE has been employed in question answering systems, where natural language queries are translated into logical queries that can be executed against a knowledge base. A notable example is the use of ACE in the OWL verbalization project, which aims to make ontologies understandable to non-experts.
Relationship to Other Formal Languages
ACE is often compared to other controlled natural languages and formal notations. Unlike programming languages like Prolog or Lisp, ACE is designed to be read and written by humans without special training. It is more expressive than typical query languages like SQL, as it can represent complex logical statements involving quantifiers and implications. However, ACE is less expressive than full first-order logic, as it lacks certain features like higher-order quantification and equality. In the context of Artificial intelligence, ACE serves as a bridge between natural language processing and formal reasoning, complementing approaches like Large language models, which are powerful but often produce ambiguous or inconsistent outputs. While Large language models can generate text that resembles ACE, they do not guarantee logical consistency, whereas ACE ensures a one-to-one mapping to logic.
Limitations and Future Directions
Despite its strengths, ACE has limitations. The strict grammar can be cumbersome for users, as they must learn the allowed constructs and avoid common English expressions that are not permitted. The vocabulary is limited, and adding new terms requires updating the lexicon. Furthermore, the translation to FOL can produce complex formulas that are difficult for humans to read, even if the original ACE sentence is simple. As of the mid-2010s, the development of ACE has slowed, with no major releases after 2015. However, the ideas behind ACE have influenced other controlled natural language projects, such as PENG and CLCE. In the era of Generative AI, there is renewed interest in using controlled languages to constrain the output of Neural network models, ensuring that generated text is logically sound. Future work may integrate ACE with Machine learning techniques to automatically convert natural language into ACE, but as of the current state, such systems are not widely deployed.