Quarterstaff: Tomb of Setmoth
The internet describes Quarterstaff: The Tomb of Setmoth (1988) as the “holy grail” of Infocom games. Infocom are the legendary creators of the Zork series of “interactive fiction” games (snooty-talk for text adventures), and Quarterstaff is a rare Mac-only release. I ordered it from a catalogue at the age of 8 based purely on the title (for some reason the word “quarterstaff” was a much bigger part of my vocabulary as a child than it is today). Quarterstaffs (quarterstaves?) were my favourite weapon as a child (for some reason, having a favourite weapon was a much bigger part of my identity as a child than it is today (but, in case you’re interested, now it’s a Remington Standard 2)).
Unlike a Choose Your Own Adventure book (which were coming into their own as a genre at around the same time), Infocom’s games featured a sophisticated text parser. So instead of being presented with a limited set of actions to choose from, you could just tell the game what you want your character to do in plain English and he would do it. In Quarterstaff this text interface is combined with a graphical map and occasional pictures of the people and objects you encounter. Here’s a quote from the Help menu:
Quarterstaff tries to provide a complete simulation environment for the user. However, like most simulations, it isn’t as complete as we would like it to be. There are two aspects of our simulation technology which you might wish for but will find lacking. First, it isn’t possible to pour liquids. Second, it isn’t possible to talk to monsters. Quarterstaff has been designed in such a way where neither of these factors will affect your ability to complete the game.
To this day I can’t tell whether this is intended as a joke, or is just a sign of very limited imagination on the part of the designers. Needless to say, there are lots of things I can type in that the program doesn’t understand. Pouring liquids is the least of my worries. I can’t “hug Bruno” (although I can kiss him), I can’t “use Eolene as a human shield,” I can’t even “run away”. Despite its promise of a complete simulation, you really only end up using a limited set of verbs: “attack” “get” “drop” “eat” “greet” “push” etc. There are some clever set-pieces of syntax to figure out, such as “Put Chief Torturer in Thumb Screws”, but it always feels a bit like a game of “guess the phrase the designers were thinking of”.
But Quarterstaff tries to do something very interesting. I see Quarterstaff as a proof-of-concept for how natural language parsing could be used to structure an RPG. Modern RPGs have all but abandoned this idea and have basically gone back to the Choose Your Own Adventure style of play, where you are presented with a limited set of options to choose from (think about the point-and-click conversation system in Dragon Age or Fallout). In the 80s there were high hopes for computer natural language recognition, hopes that we now realize were naive. It turns out language is much more complex and messy than we thought. Chatbots these days are pretty good at parsing language, but terrible at knowing anything about the world.
There are some clever set-pieces of syntax to figure out, such as “Put Chief Torturer in Thumb Screws”, but it always feels a bit like a game of “guess the phrase the designers were thinking of”
But here’s the thing: in an RPG all the objects you have to refer to are tidy little virtual objects that only function within the logic of the game. All the relevant facts about those objects are already a part of the game. This means that one of the biggest hurdles with understanding language—knowing information about objects in the real world—has already been jumped.
So why don’t we have an RPG that we can interact with in English? It could use a simplified version of English, or even have a Verb menu, like the one in Quarterstaff, but I think it would give a fantastic new dimension to RPGs. Imagine having to actually talk an NPC into joining your party using your own charisma instead of your character’s; or learning the secrets of a town by chatting with locals at the tavern; or tricking a group of goblins using a combination of bravado and clever lies; or ending a war through real diplomacy. I imagine it being based on a massive relational database: Dwarves are suspicious of Elves but hate Goblins, they are Stubborn and have immense Pride, never show Fear etc. All of these qualities can have values attached to them. Here’s how I imagine a conversation going:
Me: Hey Borrin, will you come and help me raid the dungeons?
(Borrin fails check against Suspicion of Elves)
Borrin: It’ll be a cold day in hell before I help an Elf like you.
Me: There are goblins in these dungeons that need killing.
(Borrin passes check against Hatred of Goblins)
Borrin: I hate Goblins, but nobody who enters those dungeons comes out alive. Only a fool would go in there.
Me: What, are you scared of a few goblins?
(Borrin fails combined check against his Dwarven Pride mutliplied by his Hatred of Goblins)
Borrin: Bring me my battleaxe! I’ll show you how a Dwarf deals with Goblins!
Obviously it would be impossible to deal with every eventuality, but the game could help you make well-formed sentences. In Quarterstaff, if you type “Drink potion” it will ask you “Which potion?” and list all the various potions in your possession. If you type “Use wand” it will ask you to be more specific, and suggest a construction like “Shoot gun at Johnny.”
Unlike real-world natural language recognition (i.e. a good chatbot), I think good RPG natural language recognition could be attained by using brute-force. With enough relational information about each object and person in the game, plus a bunch of verbs and prepositions (Goblins can be IN dungeons; Dwarves can be scared OF Goblins), I think it would be possible to make a truly satisfying RPG. If Infocom came close in the 80s (Quarterstaff isn’t a great game, but it’s certainly playable), imagine what a modern studio could do with a thousand times the resources.
If Infocom came close in the 80s (Quarterstaff isn’t a great game, but it’s certainly playable), imagine what a modern studio could do with a thousand times the resources.
Essentially Quarterstaff represents a diverging path in the evolution of RPGs. It had point-and-click graphical elements and a sophisticated text parser. The genre followed the graphical path but in so doing turned its back on its interactive fiction roots. But imagine the sorts of amazing RPGs could we have if we had explored the other path too?
Loved your article, and I was with right up to your conclusions. I’m a CS guy getting my Phd in statistical modeling with an emphasis on natural language processing (essentially, the arcane art of teaching machines to understand human speech) and I can tell you that your sample conversation is, unfortunately, as out there as the comments about pouring liquids in the original game materials. I love computer games, especially RPGs, and have long dreamed of exactly what you described. What I have learned the (very) hard way is that even the most basic of NLP domains tend to be extremely brittle when exposed to actual human speech. Setting aside for the moment the number of variables needed to hold information about how each group feels about one another (n^2*m by my count, where n is the number of distinct ethnic/socio-economic groups and m is the number of individuals in the game, unless you intend for all dwarves to feel the exact same way about all elves), just the simple sentences you pose would be quite challenging. Assume for the moment we have all the variables in a fast, highly accessible relational database. Now imagine the player types (I won’t even bother to add in the imperfections introduced with speech recognition) ‘Hey Borrin, will you come and help me raid the dungeons?’. At this point you have what is known as a logical entailment/classification problem. In short, the computer needs to be able to parse that sentence and deduce that you, the player, are asking him, the NPC, for help with something. Not ‘he wants my money’ or ‘he wants my advice’ or even ‘he just wants to chat’, but that you are asking him to perform a specific action for you that will lead to a check against a specific variable. This is incredibly hard to do. By way of comparison, in a limited domain entailment challenge known as the RTE Pascal challenge held every year, the very best software known to exist performs at ~80% accuracy (and by limited, I mean answering yes/no questions). Some more current work such as IBM’s Watson software shows promising results by leveraging the power of a supercomputer and a most-popular vote over several techniques, but even there it is strictly limited to identifying facts, not parsing emotions, motives, and other subtleties that humans all but take for granted. In short, while I hope that such a game might exist some day, I wouldn’t count on it for a long, long time. Hugs to Bruno, I.