Natural Language and Inference in a Computer Game

Natural Language and Inference in a Computer Game Malte Gabsdil and Alexander Koller and Kristina Striegnitz Dept. of Computational Linguistics Saarland University, Saarbr¨ucken, Germany {gabsdil|koller|kris}@coli.uni-sb.de Abstract We present an engine for text adventures – computer games with which the player interacts using natu- ral language. The system employs current meth- ods from computational linguistics and an efficient inference system for description logic to make the interaction more natural. The inference system is especially useful in the linguistic modules dealing with reference resolution and generation and we show how we use it to rank different readings in the case of referential and syntactic ambiguities. It turns out that the player’s utterances are naturally restricted in the game scenario, which simplifies the language processing task. 1 Introduction Text adventures are computer games with which the player interacts via a natural language dialogue. Texts describe the game world and how it evolves, and the player can manipulate objects in this game world by typing in commands; Fig. 1 shows a sam- ple interaction. Text adventures were very popu- lar and commercially successful in the eighties, but have gone out of fashion since then – mostly be- cause the parsers were rather limited and forced the user into very restricted forms of interaction. We describe an engine for text adventures that attempts to overcome these limitations by using current methods from computational linguistics for processing the natural language input and output, and a state-of-the-art inference system based on de- scription logic (DL) to represent the dynamic state of the game world and what the player knows about it. The DL prover is used in all language-processing modules except for parsing and surface realization, and supports the inferences we need very well. This shows in particular in the modules for the resolution and generation of referring expressions. By keeping track of the true state of the world and the player’s knowledge in separate knowledge bases, we can evaluate definite descriptions with re- spect to what the player knows. In generation, such inferences allow us to produce smaller while still sufficiently informative references. Another interesting aspect which we discuss in this paper is the treatment of syntactic and referen- tial ambiguities that come up in understanding input sentences. Here, too, the player knowledge restricts the way in which the input should be interpreted and guides the resolution process. We use inferences about the player knowledge to rule out inconsistent analyses and pragmatic heuristics to possibly select the preferred one. Players of a text adventure are effectively situ- ated in a game world and have to accomplish a specific task, which severely restricts the utterances they will naturally produce. For example, they will typically only refer to objects they could “see” in the simulated world. This simplifies the language processing tasks tremendously. The scenario of the game can be freely specified and tailored to different applications and levels of complexity. Along with the modularity of our implementation, this makes our system an interesting testbed for evaluating more …

Content truncated to 500 words