Helmut Horacek

Dialog control in a natural language system

In this paper a method for controlling the dialog in a natural language (NL) system is presented. It provides a deep modeling of information processing based on time dependent propositional attitudes of the interacting agents. Knowledge about the state of the dialog is represented in a dedicated language and changes of this state are described by a compact set of rules. An appropriate organization of rule application is introduced including the initiation of an adequate system reaction. Finally the application of the method in an NL consultation system is outlined.

Probabilities in Game-Playing: Possible Meanings and Applications

In the domain of game-playing the notion of ‘probability (to win)’ is used in several methodological approaches of the field, but without a precise specification in many instances. This paper aims at an improvement of that situation by accomplishing two joint tasks. Firstly, a clarification of the meaning of such a probability according to the circumstances of its use is performed. Secondly, several devices and methods used in game-playing programs are examined. Their benefits and shortcomings are discussed, with emphasis on a comparison of the intended meaning and the application of the probability.

Choice of words in the generation process of a natural language interface

Abstract In this paper we present our approach to language generation focusing on the choice of words. We demonstrate the requirements for a generator that is successfully integrated with a complete natural language system. A basic word-mapping method for primitive relations of the semantic representation language is outlined that can be relied on if no specific realization is available for the system. Information chunking and tailoring of the word choice according to the specific circumstances are presented, stressing the extension to the most general coverage for each specific instance. In particular, the influence of the scope of quantifiers on the choice of words and syntactic relations is pointed out.

Decision Making in Unclear Situations

Large domains often contain problem states which cannot be evaluated accurately enough (neither statically nor dynamically). However, it is necessary to make decisions in such an environment, too. Here we discuss some of the usual methods of decision making in two-person games and give an analysis of the problems induced by unclear situations. Two classes of such situations are distinguished and a model for improved decision making based on the point-valued minimax approach (for reasons of practicability) is presented. Asymmetric evaluations (depending on the side to move) help to avoid unclear situations the program does not understand. Moreover, the inclusion of the global state of the game helps to avoid risks in favourable situations and to play for counter-chances in disadvantageous ones. Practical results have been gained from the implementation in a chess program, showing that problems can be solved this way, which have been intractable up to now.

How to Say WHAT - IT or SOMETHING?

This paper deals with the nearly classical dichotomy whether or not to partition the entire generation process into ‘What to say’ and ‘How to say it’. Evidence from practical systems is collected in order to detect those properties which are inconsistent with the chosen approach of the system with respect to the dichotomy. The reason for these failures is examined which, in our view, lies in various aspects of limitations of the representation and availability of knowledge. An attempt is made to formulate necessary conditions that enable a strict partition into ‘What’ and ‘How’ and the potential benefit of such a design is sketched out.

An Outline of a New Algorithm for Game Tree Search

In this paper the basic ideas of a new, flexible and knowledge-intensive algorithm for game tree search are outlined. A quick selection is made by filtering out the most promising part of a usually large game tree (in the discovery phase). This partial tree can be examined afterwards in a more intensive way (in the decision phase). Notable features of the algorithm are a method for the suitable expansion of the search tree, means for a finegrained evaluation of uncertainties and a basis for well-founded decisions.

Facets of Knowledge About Natural Language Syntax Representation and Use in Parsing and Generation

This paper aims at a concise representation of purely static knowledge which expresses the syntactic well-formedness of a subset of German. A method is proposed for automatically exploiting this knowledge in order to obtain dynamic knowledge about language analysis and production. The efficient and skillful use as well as integration into a parser and a generator is outlined.

Anwendbarkeit von Unifikationsgrammatiken für effizientes Generieren

Die prinzipielle Anwendbarkeit von Unifikationsgrammatiken fur die Generierung ist mehrfach gezeigt worden, wobei jedoch das Problem der Effizienz meist nicht untersucht wurde. In diesem Papier soll anhand einiger einfacher und fur das Deutsche typischer Beispiele die Eignung bzw. Ungeeignetheit von Unifikationsgrammatiken zur effizienten Bearbeitung der gewahlten Phanomene demonstriert werden. Abschliesend wird auch versucht, die typischen Eigenschaften der untersuchten Phanomene zu beschreiben, um aus den gewonnenen Erfahrungen (gewisse) Richtlinien fur den effizienten Einsatz von Unifikationsgrammatiken zu erhalten.

Semantic Interpretation in Natural Language Interface Applications

In this paper we will present how the practical relevance of the semantics in Natural Language utterances can be made available for database interaction without a complete understanding of the general meaning of such utterances. We will demonstrate our ideas on the treatment of adjuncts, especially of time adjuncts, on the treatment of the verbal complex and of words representing compound actions like “verify” and “complete”. Representation techniques and interpretation heuristics are sketched out that take advantage of the restricted domain and produce useful relations between critical parts of an utterance and entities of a database. This work has been done within the frame of the Hamburg subpart of the ESPRIT project LOKI whose aim is the production of a Natural Language Interface (German and English) to a database, concentrating on pragmatics and performance.

A Heuristic Model for Treatment of Vague Time Descriptions

In this paper we present a heuristic model which is capable of dealing with vague time references and modifiers. Basically, vagueness appears in two different ways: a reference can be incomplete or the degree of concreteness may be low. In order to overcome this lack of information we propose the application of defaults and probability distributions for the computation of possibility estimates about the temporal constellation of events. Conceptual descriptions of time references are assumed to be derivable from natural language expressions. They are mapped onto an appropriate numeric form providing the unification of descriptions with different types of units and concreteness. Some operations are defined realizing reasoning processes to determine adequate expectations about events characterized by several temporal descriptions. As we will show, this method provides a basis for treating questions about potential events, which cannot be answered on the basis of a true/false mechanism. Finally, some ideas about the verbalization of the numeric representation are discussed.