We present a CNL, which is a component of L4, a domain-specific programming language for drafting laws and contracts. Along with formal verification, L4’s core functionalities include natural language generation. We present the NLG pipeline and an interactive process for ambiguity resolution.
We present an initial reference architecture for traceability in SBVR-based systems. It facilitates rule-based development that lets end users trace interface behavior back to the human decision points that lead to it. This closes a feedback loop, which facilitates agile development. The architecture is based on Web standards to generalize comparison and implementation. It begins with the human process of linking document excerpts to the SBVR code that defines them. The next step transforms SBVR into computer code that implements it. Reasoners then form conclusions by applying the rules to data. They can also provide rudimentary explanations for these conclusions. The system then provides an end-user interface to all this rulederived information. The core challenge here is maintaining the data needed to trace back through these layers, so end-user feedback can improve the entire development proces.
Automatic generation of feedback messages in a natural-language based programming for video games is presented. The input sentences are processed in four stages. During each stage, context vectors are aggregated and any violation to a syntactic or semantic rule is reported to allow users to debug and fix the text. The results discuss a list of common errors detected by the proposed method.
Grammar-based domain-specific MT systems are a common use case for CNLs. High-quality translation lexica are a crucial part of such systems, but involve time consuming work and significant linguistic knowledge. With parallel example sentences available, statistical alignment tools can help automate part of the process, but they are not suitable for small datasets and do not always perform well with complex multiword expressions. In addition, the correspondences between word forms obtained in this way cannot be used directly. Addressing these problems, we propose a grammar-based approach to this task and put it to test in a simple translation pipeline.
Negation detection is a key feature to the processing of biomedical text, and it involves two steps: identifying a medical term of interest in text and identifying that that medical term is mentioned as absent. However, processing biomedical text is made complex by the presence of medical jargon that typically requires custom systems, and detecting negation is complicated further because the representation of negation in natural language varies according to the grammar used. We investigated the use of a CNL with a general-purpose semantic parser to detect negation. Our CNL was created by representing medical terms as their semantic types and restricting the definition of the expression of negation. Through this method, we identified three kinds of negation–explicit negation, implicit negation, and explicit implicit negation. A pilot evaluation of our method on a sample of radiology reports achieved an F1 score of 0.99 on the sentences that could be parsed.
Competency questions (CQs) are used in ontology development to demarcate the scope, provide insights into their content, and verification. Their use has been impeded by problems with authoring good CQs. This may be assisted by a controlled natural language (CNL), but its development is time-consuming when carried out manually. A recent study on data-driven CNL design to learn templates from a set of CQs, resulting in CLaRO , had somewhat better coverage and some noise due to grammar errors in the source CQs. In this paper, we aim to investigate such a bottom-up approach to CNL development for CQs regarding the effects of 1) improving the quality of the source data 2) whether more CQs from other domains induce more templates and 3) if the structure of knowledge in subject domains has a role to play in the matching of patterns to templates; therewith might indicate that possibly a structure of knowledge in a subject domain may continue to affect bottom-up CNL creation. The CQ cleaning increased the number of templates from 93 to 120 main templates and an additional 12 variants. The new CQ dataset of 92 CQs generated 27 new templates and 7 more variants. Thus, increasing the domain coverage had the most effect on the CNL. The CLaRO v2 with all generated templates has 147 templates and 59 variants thereof and showed 94.1% coverage.
In this paper we present the controlled language and the grammar of the PENG system and explain how the new version of the grammar has been implemented in a logic programming framework. The grammar is now bidirectional and can be used to translate a specification written in controlled language into an executable answer set program and vice versa. The grammar is highly configurable for different application scenarios and can be used for incremental text processing together with a predictive authoring tool.
We introduce a method for tailoring a controlled language out of a specialized language corpus, as well as for training the user to ensure a smooth transition between the specialized and the controlled language. Our method is based on the selection of maximal coverage syntax rules. The number of rules chosen is a naturalness vs. formality parameter of the controlled language. We introduce a training tool that displays segmentation into left-toright maximal parsed sentences and allows utterance modification by the user until a complete parse is achieved. We have applied our method to a French corpus of maintenance reports of boilers in a thermal power station and provide coverage and segmentation results.
RACE is a first-order reasoner for Attempto Controlled English (ACE). This paper introduces mathematical and functional extensions. It is the third system description of RACE, and also the final one since RACE now covers all ACE constructs that have a representation in first-order logic.
Multilingual Grammatical Framework (GF) domain grammars have been used in a variety of different applications, including question answering, where concrete syntaxes for parsing questions and generating answers are typically required for each supported language. In low-resourced settings, grammar engineering skills, appropriate knowledge of the use of supported languages in a domain, and appropriate domain data are scarce. This presents a challenge for developing domain specific concrete syntaxes for a GF application grammar, on the one hand, while on the other hand, machine learning techniques for performing questionanswering are hampered by a lack of sufficient data. This paper presents a method for overcoming the two challenges of scarce or costly grammar engineering skills and lack of data for machine learning. A Zulu resource grammar is leveraged to create sufficient data to train a neural network that approximates a Zulu concrete syntax for parsing questions in a proof-of-concept question-answering system.
The main goal of this research is to optimize an existing Airbus Cockpit Controlled Language in order to integrate it in future cockpit design. The current controlled language used aboard Airbus cockpit interfaces was carefully constructed to avoid ambiguity and complexity. In order to optimize the existing language, we set out to evaluate the appropriate levels of simplification that would achieve more accurate and faster comprehension with optimized pilot training time by using psycho-linguistic experimentation and cognitive science tools. We present in this paper a congruency task similar to traditional judgment tasks in behavioral experiments. It provides a firmly controlled environment to test linguistic hypotheses and CNL rules. Results shows that what we sometimes mistakenly label as superfluous or empty syntactical elements could go a long way in ensuring better comprehension and faster information processing from a psycho-linguistic point of view.
RegelSpraak is a CNL developed at the Dutch Tax Administration (DTA) over the last decade. Keeping up with frequently changing tax rules poses a formidable challenge to the DTA IT department.RegelSpraak is a central asset in ongoing efforts of the DTA to ttune their tax IT systems to automatic execution of tax law. RegelSpraak now is part of the operational process of rule specification and execution. In this practice-oriented paper, we present the history of RegelSpraak, its properties and the context of its use, emphasizing its double functionality as a language readable by non-technical tax experts but also directly interpretable in a software generating setup.
The Dutch Tax Administration has developed and exploited a CNL, RegelSpraak, to auto-mate law execution. This CNL is meant to be comprehensible for legal specialists, IT developers and computers. However, quality assessment of rule patterns and their ability to express all relevant tax laws is currently not possible. In this study we evaluate potential quality criteria that offer the capability to evaluate rule pattern quality based on semantic expressive power, cognitive usability and functional and structural correctness. We design a quality framework based on insights from literature review, interviews and observations, which has been qualitatively operationalized. Initial re-sults suggest that they touch on relevant variables
Daily experience teaches us that a situation remains unchanged unless somebody or something changes it. Leibniz called this experience the law of inertia. Early attempts to formalise the law of inertia failed because they offered no easy way to describe that after a partial change of a situation the unaffected rest remains unchanged. This so-called frame problem was efficiently solved by later approaches, specifically by the event calculus and the default logic. Focusing on default logic, I will show that it can express the law of inertia not only in first-order logic, but also quite naturally in Attempto Controlled English. Furthermore, I will use the Attempto reasoner RACE to efficiently reason with the law of inertia. 1 The Common Experience of Inertia When you return to your office in the morning you expect to find the items on your desk in exactly the same order as you left them in the evening before – unless somebody or something moved them. This common experience was expressed by the philosopher Leibniz as the law of inertia: "Everything is presumed to remain in the state in which it is." (Leibniz, 1679). At about the same time Newton published his three laws of motion, the first of which expresses the specific case of the law of inertia for moving physical bodies (Newton, 1687). I will show that Leibniz' law of inertia can be formalised in Attempto Controlled English (ACE)1and that this formalisation allows us to reason with the law. In section 2, I describe early attempts to formalise common sense, specifically 1 http://attempto.ifi.uzh.ch/ the law of inertia, the encountered frame problem, and a variant of the Yale Shooting Problem. Section 3 presents two solutions of the frame problem, the event calculus and the default logic. In section 4, I express the default logic in ACE. Section 5 describes reasoning with the law of inertia using the Attempto reasoner RACE2. Section 6 shows that incorporating the law of inertia into RACE facilitates the reasoning. Section 7 revisits the Yale Shooting Problem in ACE/RACE. Section 8 summarises the paper and briefly addresses the fact that the law of inertia and, specifically, its formalisations are asymmetric with respect to time. 2 Formalising Common Sense and the Frame Problem Beginning in the 1960s researchers began to formalise common sense, predominantly in first-order logic. When trying the express the law of inertia they encountered the so-called frame problem, that is how to effectively and efficiently describe the unaffected part of a situation after a partial change. Initial attempts to solve the frame problem failed, mostly because only the changed parameters were taken into account and the unchanged parameters ignored. The shortcomings were strikingly demonstrated by the impossibility to adequately solve the so-called "Yale Shooting Problem" (Hanks and McDermott, 1987). In this paper I replace the original problem by a less violent, but "problemidentical" one. 2 http://attempto.ifi.uzh.ch/race/ The Law of Inertia and the Frame Problem in Attempto Controlled English Norbert E. Fuchs Department of Computational Linguistics University of Zurich fuchs@ifi.uzh.ch http://attempto.ifi.uzh.ch Initially, a wine glass is empty and a wine bottle is not open. Opening the bottle, waiting a moment to read the label, and then pouring the wine should fill the glass. If this situation is formalised in first-order logic by only taking the changed parameters into account and ignoring the unchanged ones, it cannot be uniquely proved that the wine glass is finally full. In one logical solution the wine glass is actually full; in another logical solution the wine bottle is again not open and the wine glass remains empty. Here is a simple formalisation3 of my version of the Yale Shooting Problem using four time points 0, 1, 2, 3 and the two fluents – conditions that can change their truth value over time – empty and open expressed by the following first-order formulas:
HESIP is a hybrid machine learning system in which a sub-symbolic machine learning component makes a prediction for an image classification and afterwards a symbolic machine learning component learns probabilistic rules that are used to explain that prediction. In this paper, we present an extension to HESIP that generates human-understandable and machineprocessable explanations in a controlled natural language for the learned probabilistic rules. In order to achieve this, the literals of the probabilistic rules are first reordered, and then aggregated and disambiguated according to linguistic principles so that the rules can be verbalised with a bi-directional grammar. A human-in-the-loop can modify incorrect explanations and the same bi-directional grammar can be used to process these explanations to improve the decision process of the machine.
Understanding texts in Attempto Controlled English (ACE) is considered undemanding, nonetheless hides some problems. To deal with these problems I propose an experiment based on Kuhn's ontographs that tests the understanding of simple ACE texts. Furthermore, I suggest to compare the relation between authors and readers with human verbal conversations. My conclusion is that the correct understanding of an ACE text is possible, but requires contributions from both authors and readers, quasi their cooperation.
While machine processable Controlled Natural Languages (CNLs) as a natural language interface have proven a popular, unambiguous and user friendly method for non experts to engineer formal knowledge-bases, human-oriented CNLs however remain under-researched despite having found favor within industry over many years. Whether such human orientated CNLs like the machine processable counterparts can be captured automatically as formal knowledge remains an open question. In addition, rewriting all or most of a human-oriented CNL into a machine-oriented CNL could unlock significant silos of general purpose domain knowledge, contained within existing human-oriented CNL content for exploitation by knowledge based systems. This paper explores the feasibility of rewriting a human-orientated CNL represented in Simplified English into a well know machine-oriented CNL represented in ACE CNL and describes preliminary results.
The correct modelling of negation in a computational grammar is essential in order for the grammar to be useful in natural language processing, and controlled language development and applications. An important and unique feature of Afrikaans is how it deals with negation. We present an exposition of a substantial fragment of negation in Afrikaans and discuss our implementation thereof in GF. Examples are given as illustration of the relevant issues. The paper is concluded with a discussion of results and plans for future work.
We present an editor for controlled languages which is a combination of a syntax editor and a predictive editor. It shows a bird's-eye view which lets the user to explore what is possible in the language. Still, unlike the syntax editors the user is not expected to understand the underlying abstract syntax or ontology behind the language. It also lets the user to enter arbitrary phrases from which the editor finds the phrases which are the closest match.
Controlled Natural Languages (CNLs) have many applications including document authoring, automatic reasoning on texts and reliable machine translation, but their application is not limited to these areas. We explore a new application area of CNLs, the use of CNLs in computer-assisted language learning. In this paper we present a a web application for language learning using CNLs as well as a detailed description of the properties of the family of CNLs it uses.
