Author: Border, Peter
We report on a novel technique to manage pre-written lines of character dialog by treating a conversation as a game played by the speakers. Thinking of a conversation as a game instead of as an information exchange means structuring it as a series of moves, made according to rules, with some sort of score kept by the speakers. We have developed our framework by studying the dialog in theatrical plays, which has been very helpful in finding a preliminary set of conversational rules. We believe this framework will eventually allow much more realistic NPC dialog than the usual decision trees. Speakers in our system converse by participating in "conversational fragments," or short fixed sequences of lines. They make "moves" in the conversational game by negotiating transitions between fragments. Speakers have some internal state variables which describe their current standing in the conversation, as well as their current emotional and physical state. Speakers try to manage the conversation so as to to maximize some "payoff function" of their internal variables. The internal variables govern the negotiations around transitions, so speakers with a poor set of internal variables may not be able to transition to more desirable (higher payoff function) fragments. Transitions may be requested during a fragment, as well as when it has finished. We have developed our framework by studying several short theatrical plays, and analyzing the dialog in them. We have identified many types of fragments and studied the transitions between them, as well as compiling a reasonable set of characters' internal state variables. The "payoff functions" clearly vary between individuals (and probably change with time), and so far we have only an ad hoc understanding of them. If all goes well, we should have some usable text-based demonstrations working for the DIGRA conference. Previously, we tried a system with a large "dialog mesh", made of interlocking threads of dialog. The dialog manager chose NPC lines from decision trees which were "pruned" based on an NPC's emotional state variables. The "pruned mesh" system was adequate for simple encounters (such as a frantic player character asking a cantankerous NPC gas-station attendant for directions), but lacked the depth and complexity needed for realistic conversation. We found that our characters simply needed more room to play with their words, and some way to keep score, than the "pruned mesh" system could supply. Our new system avoids these problems by allowing characters to participate in fragments purely for the joy of making a "bon mot", or scoring a point on the other speaker, regardless of whether any information is exchanged, or whether anything of any practical significance is accomplished at all. We believe that this technique will allow us to create much more lifelike NPC dialog than is currently present in games, and allow NPCs to play a much more prominent social role in game worlds. We also believe that games in general desperately need to have more socially coherent characters, and that improving dialog is a critical problem in game development today. There are many possible applications of this system outside of gaming- certainly "virtual tours" would benefit immensely from livelier NPC dialog, and applications in education, health maintenance, sales, consulting, help desks, and so on are obvious.
Contact: Peter Border, Physics Dept, University of Minnesota, email@example.com
Copyright is held by the author(s).
Member of collection