Phantasmal Site > Setup > Configuring ParseD

Configuring the ParseD

The command line is the most direct interface between your players and your MUD. Phantasmal offers customization of the parser so that you can give your players the best possible experience in interacting with your game. The following is Phantasmal's idealized parsing architecture. Not all of it will be immediately available.

How the Parser Works

Tokenizing and Parts of Speech

First, the parser looks over various words and decides what parts of speech they're allowed to be. For instance, the word "fast" can be an adjective or adverb, but can also be a noun or a verb. Most words aren't quite as versatile -- "green" is a noun or an adjective, but almost never a verb, and definitely not an adverb.

Some parts of speech are gathered from your existing data. Since your objects already have a wide variety of nouns and adjectives, Phantasmal registers all of them with the ParseD. If you see a message like "I don't know the word 'clock'", it probably means that you have no object in your game with the noun 'clock'. You can add it to the game's vocabulary simply by creating a clock object anywhere in your MUD. Adjectives (and eventually plural and possessive nouns) work the same way.

Verbs are registered specifically with the parser. Note that verbs are not the same as actions in your MUD. For instance, it's quite reasonable to make the phrase "get down" mean to dance, even though 'get' usually means to take an object. Most MUDs don't allow this, or would require you to code it into the 'get' action. Verbs are categorized as either intransitive (i.e. not taking a direct object), transitive (i.e. taking a direct object, and perhaps an indirect object as well), or both. An example of a verb which is both is 'sing' -- you may say 'sing' by itself, or give it a direct object, as in "sing something by Vivaldi".

Adverbs currently don't do anything, but they'll eventually be able to be registered. Those will be important primarily for social verbs, though phrases like "walk carefully" or "sneak silently" should eventually be possible, and may even modify the action if the game so chooses.

Special words like conjunctions (and, but, yet, so, for, and, nor) and prepositions (into, around, under, behind, of, on, on top of, underneath, etc) are currently hardcoded into the parser, and are likely to be for quite some time.

This division by what parts of speech a word may be is called 'tokenizing'. The now-categorized words are called 'tokens'.

Grammar and Parsing

The parser will then examine all of these words and come up with possible sentence structures. For instance, the phrase "move the box under the table" might be parsed with the prepositional phrase modifying the box, or modifying the table. "Fast and quickly run" can mean either of two things, depending on whether 'fast' is meant as a verb or an adverb. Each sentences can each have attributes for things like grammar and perceived appropriateness.

At this point, the sentences have been parsed only grammatically. The nouns haven't been checked against available people and objects, and the verbs (and corresponding actions) haven't been checked to see if they are possible or appropriate. "Eat the red shoe behind the mother of checkered bobcat" will be considered just as valid as "go north", except perhaps for the grammatical nitpick in the former.

Verb Resolution and Binding

The next stage accomplishes verb resolution and binding. Verb resolution checks the appropriateness of certain grammatical constructions, and maps the sentence into an action. Binding matches noun phrases to objects in the environment.

The various parsed sentences from the previous stage are passed to an LPC object representing the verb. That object decides how binding should be accomplished, and calls an internal Binder API to accomplish the actual binding. Verbs that take hypothetical subjects may do no binding at all -- 'watch for hairy green forks' may be a perfectly valid thing to say even if no hairy green forks exist anywhere in the game. The 'watch' verb would simply choose not to bind that noun phrase, and instead pass it through verbatim to the 'watch' action.

For most actions, the binding will locate an object in the immediate environment. For instance, in the phrase "get the third sticky grape", binding would look for objects that might be called sticky grapes, and choose the third of them. There may eventually even be support for comparative and superlative adjectives -- "get the hottest potato in the bin". But currently no such thing exists.

More complex sentence structures are possible, and binding will occur on them in this stage as well. In the phrase "remove a pen from the big box", the big box will also be resolved by the binder.

Normally, if an object can't be found by the binder or for some other reason the sentence seems impossible, it still won't be thrown out. If the player types "get the green emerald" and there's no emerald available, it's important to receive a message like "You can't find one.". So the sentence should be marked as unlikely (in case there's some other way to parse the sentence that makes more sense), but it shouldn't be rejected outright since they may indeed have meant what they typed, even if it's not currently possible.

Action Resolution

Once the verbs have done appropriate preprocessing, the action object receives the mostly-parsed sentence structure. Action structures are fully specific to your game, and so they can decode information like whether a particular noun is edible. That's useful so you can realize that "take cake from oven and eat it" probably means to eat the cake, not the oven. The 'eat' action will rate both as possible meanings, but the cake should score higher for appropriateness, and so it should be chosen in favor of trying to eat the oven.

The actions should encode only things reasonably known to the player character. For instance, if an object is poisonous but the character doesn't know that, the parser should treat it like any other edible object. If a particular action will result in disaster but the player couldn't reasonably know, don't mark it as unlikely. Players will find ways to cheat by referring to the action ambiguously. On the other hand, if one possible interpretation is to stick a dagger into the character's own eye, the player probably didn't mean that. At least, not if there's another choice.

Finally, Choosing an Action

There are now one or more possible interpretations of the player's input that Phantasmal has scrutinized, and it's merely a matter of deciding on which. Based on appropriateness, grammar, likeliness to damage the character and other factors, ParseD will decide which interpretations are most likely. If there is only one candidate, or only one credible candidate, that action will be taken. If the obvious action is also dangerous, the player may be prompted to see if he really meant to type that. If there are multiple good possibilities, the player may be prompted which he meant.