VOX - An Extensible Natural Language Processor

1. INTRODUCTION

The ability to learn is one of the most important charac­ teristics of intelligent systems. To approach such an abil­ ity, we first must build systemS"" that can accept new knowledge automatically . . By continually enhancing the extensibility capabilities of such systems, we can begin to address the problems of general learning.

Critical to extensibility is the underlying knowledge representation. The more powerful and flexible the knowledge representation, the more easily extensibility capabilities can be built and improved.

VOX (Vocabulary Extension System) is a Natural Language Processing system that emphasizes automatic extensibility. In VOX, extensibility capabilities are developed hand-in-hand with the knowledge representa­ tion. The knowledge representation, called Conceptual Grammar [5], supports a bottom-up study of language, by representing both very general and very specific knowledge. As generalizations about language are :!iscovered, they are incorporated into the representation.

Currently, VOX allows automatic addition of vocabulary and action-oriented events and scenarios. The user may build knowledge hierarchies of scenarios, events, nouns, verbs, adjectives, and other parts of speech, as weil as specifying a variety of semantic and syntactic information

1.1 EXAMPLES

We will illustrate how VOX works by adding the se­ quence of events for a simple Naval 'attack' scenario:

ship searches for ship.

ship sights ship. ship approaches ship. ship attacks ship.

ship damages ship.

We will add the words, the individual events, and the en­ tire scenario to the system. Then, we will show a text analysis example that uses this knowledge. (User inputs are in boldface, in the examples below.]

MACRO NOUN EXAMPLE:

Enter singular form of noun: ship Enter plural form of noun: ships

Enter synonym or more general concept: platform

Macro noun is an extensibility capability for adding nouns. The words 'ship', 'ships', as weil as the more abstract concepts ship(noun) and ship(np) are added to the knowledge base by macro noun. A phrase like "the 3 green ships" will be found to be equivalent to ship(np), for example. By specifying 'platform', the user places 'ship' into a conceptual hierarchy of nouns already con­ taining 'platform'. ('Platform' is a Navy word for any­ thing that a missile can be fired from. Thus, a base, a submarine, and an aircraft are all platforms.)

Macro verbis similar to macro noun. In addition to con­ cepts for words, verb, and verb phrase Ievels, macro verb creates concepts for the event and frame Ievel. For exam­ ple, in adding the verb search, the concepts search(event) and search(frame) will be created.

Assurne that all the word-level items in the simple attack scenario have been added using macro noun, macro verb, and macros for other parts of speech. Next, we add an event:

Thu work u 1upporkd by the Naval Ocean Sy&terru Center, under NOSO Grant N66001-83-0-oess.

MACRO EVENT EXAMPLE:

Enter an event:

ship search location for ship

1 2 3 4 5

Semantic information

Enter position of the following:

actor = 1

act = 2

object = 5

instrument = _

location = 3

time=_

Enter a concept that the event suggests: search

Syntaxinformation

Enter position of subject: 1

Enter voice of act (active or passive) : active

Optionality information

Enter starting points of event: 1

Enter end points of event: 5

Enter skipping points for event:

Element 1 can skip to: _

Element 2 can skip to: 4

Element 3 can skip to: _

Entry for new event = ship-search

Macro event Iets the user add standard events to the knowledge base. These events are templates, and will match much more than the literal words "ship search lo­ cation for ship". Macro event uses the abstract concepts ship(np), search(vp), rather than the word-level .concep~s. The event added is treated not just as a semant1c restrJc· tion, but as a full-fledged concept. The concept <ship· search-location-for-ship> is stored in the knowledge base under the entry 'ship-search'. We can use concepts such as this to add new scenarios, as will be shown below. This specific event-concept is added to a hierarchy of events by suggesting the generic 'search' event.

The user specifies the semantic case (actor, act, location, etc.) of each element in the event. The user specifies that the phrase starts with element 1 and ends with element 5. The syntactic component of VOX's grammar handles in­ complete forms such as "ship searched the area", so the user need not specify that element 3 is a possible end of the event phrase. The user specifies that element 3 can be skipped over; that is, "The ship searched for the sub­ marine", omitting a location element, is correct English. The user specifies this because it varies on a case-by-case basis. For example, in the sentence "Ship conducted at­ tack on submarine", "attack" could not be omitted.

Assurne that all events of the simple attack scenario have been entered using macro event. Next, we· invoke macro frame to add the entire scenario.

MACRO FRAME EXAMPLE:

Enter the events of the frame:

1 ship-search

2 ship-sight

3 ship-approach

4 ship-attack

• 5 ship-damage

Briefly describe the frame: ship attacks ship

Semantic Information

Enter the main event: 4

Enter events needed in a complete text: 2 4 5

Enter concept suggested by frame: attack

Enter number of actor roles: 2

Enter a word for actor 1: ship

Actor 1 is subject in which events? 1 2 3 4 5

Actor 1 is object in which events? _

Enter a word for actor 2: ship

Actor 2 is subject in which events? _

Actor 2 is object in which events? 1 2 3 4 5

Optionality Information

Enter starting points for frame: 1 2 3 4 5

Enter end points for frame: 4 5

Enter skipping points for frame:

Element 1 can skip to: 3 4 5

Element 2 can skip to: 4 5

Element 3 can skip to: 5

Entry for frame = ship-attack

Macro frame is similar in many ways to macro event. In particular, this specific scenario is given full concept status and is placed under the entry 'ship-attack' (Note: the e~try 'ship-attack' holds both events and scenarios.) We entered it into a hierarchy of seenarios by having it suggest the generic attack(frame) concept.

Optionality information: The user specifies that the description of the scenario could start with any of the events in it and skip over any events. For example, a complete text might read "damaged sub". On the other hand, we require that an attack scenario end with an at­ tack or darnage event (event 4 or 5) .

Having entered this scenario into the system, we can make use of it to understand texts that deal with a 'ship attacking ship' scenario. Here is an example of the kind of text VOX analyzes using the scenario just entered.

VOX TEXT ANALYSIS EXAMPLE

( Constellation is message sender.)

Type message:

at 1235T bad searched area. damaged sub.

INTERPRETATION 1 OF 1. MESSAGE FEATURES

ERROR: missing event = attack

ERROR: missing event = sight

ERROR: missing object =sub

ERROR: missing actor = constellation

ERROR: missing actor = constellation

REWORDED MESSAGE

constellation had searched area for sub at 1235 t. constellation damaged sub.

Frame = ship attacks ship

When analyzing text, VOX builds a frame-based representation of the underlying meaning, which is used for checking and correcting syntactic and semantic prob­ lems. VOX reports errors found and produces a reworded version of the input text . VOX's use of extensibility ses· sion information in the above example is fairly Straight­ forward.

2. CONCEPTUAL GRAMMAR

We describe the Conceptual Grammar knowledge representation, which forms the foundation of the extensi· bility and analysis capabilities of VOX.

2.1 INTRODUCTION

Conceptual Grammar (CG) is a framework for the representation of conceptual information. The unit of knowledge in CG is the concept. A concept is an atomic representation of anything that can be verbalized. In our notation, a concept is depicted by a description of the concept enclosed in angle brackets. For example,

<aircraft carrier (noun)>

is an atomic representation of the concept "aircraft car· rier". We often omit the angle brackets and hyphenate the description, for simplicity:

aircraft-carrier or <aircraft-carrier>

Concepts can be combined to form phrases. Most phrases have associated concepts to represent their meaning. Concepts and phrases suggest, or reduce to, other concepts by means of grammar rules. Some typical rules in CG are shown in Figure 1.

FIGURE 1

In rule (A), a phrase suggests its atomic representation. The phrase "aircraft carrier" has no corresponding single English word, yet it is a well-defined object , so we represent it with an atomic concept.

Rule (B) is an example of a hierarchical rule. The hierar· chy is an important organizing principle of CG. CG has semantic hierarchies of nouns, adjectives, and some other parts of speech, as weil as events and Scenarios. For ex­ ample, a scenario where a ship attacks a submarine is a. more specific instance of one where a platform attacks a platform, which is a more specific version of the generic attack scenario, and so on. ('platform' is a Navy word for anything that missiles can be fired from.)

Rule ( C) illustrates a second organizing principle of CG ·· conceptuallevels. When we speak of 'attack', we may be talking about the word itself, the verb, the action, the event , or an entire scenario. CG treats all of these facets as explicit concepts. The lower-level semantic concepts correspond to the syntactic concepts word, verb, verb phrase, and so on. The higher semantic Ievels of event and frame (or scenario) have no precise syntactic equivalent. A frame could correspond to a sentence, a paragraph, or even a novel.

The different Ievels of semantic concepts allow semantic phrases to be represented unambiguously in CG. Note how the concept of 'ship' is used in

1. <ship(np)> <attack(vp)> <submarine(np)>

(2) <ship(noun)> <ahoy(word)> <'!'>

1. <ship(word) > <hyphen> <shape(word)>

Phrase ( 1) would match a text like "The 3 US destroyers will attack the enemy sub". Phrase (2) matches only "ship ahoy!", "destroyers ahoy!", etc. Phrase (3) matches only "ship-shape". Conceptual Ievels allow se­ mantic phrases to be represented with a high degree of precision.

An important dass of rules in CG is concerned with the transitions between semantic and syntactic phrases. In Figure 1 above, rule (D) shows a semantic concept sug­ gesting a syntactic concept, while rule (E) shows the re­ verse. Rule (E) is an example of a restriction rule. It suggests a specific noun-phrase concept corresponding to the noun on the left-hand-side of the rule. We will dis· cuss this kind of rule in more detail in the next section.

2.2. EXAMPLE

We show, step-by-step, how CG analyzes

"Green ship will fire 2 missiles at 1230pm at submarine".

Most important here is step X, which uses the rule:

Since <ship (noun)> gave rise to the noun in the left­ hand-side of the above rule, this rule suggests <ship (np)>. In essence, "green ship" has been condensed to the semantic concept <ship (np)>.

Here again, we are using a semantic restriction rule:

This rule finds the most specific possible semantic in­ stance of the verb, and suggests its corresponding vp­ level.

"2 missiles" is analyzed similarly to "green ship".

The detailed analysis of the prepositional phrases "at 1230pm" and "at submarine" is omitted for simplicity. Both will suggest the concept <adv>, which corresponds to prepositional and adverbial phrases. Also, we represent a Iist of adverbial phrases by <x>, for short .

Rule Y Iooks like

The task of rule Y is to find specific events in the data­ base; in this case, it found the event "ship attack subma­ rine". Now, rule Y is highly sophisticated, and we will describe some of the actions that it took. First , note that "ship fired missile at submarine" is syntactically am­ biguous. "At submarine" could be analyzed as a preposi­ tional phrase, or " fire ... at" could be recognized as a prepositional verb with its associated particle. Rule Y knows about both of these possibilities. It checks the kind of verb, and uses one of the rules A or B according­ ly:

If the verb can be prepositional, rule Y Iooks for rules such as

In our case, it will find and use this rule. Having found that "will fire 2 missiles at" corresponds to an 'attack' concept, rule Y Iooks for the most specific possible event of the type

<ship> <attack> <submarine>

or

<platform> <attack> <platform>

and so on. Once the most specific possible event is found, rule Y suggests it. Note that rule Y has to search through the adverbial-list to find possible prepositional particles for the prepositional verb, and that it rejected the time adverbial because the knowledge base has no information about attacks on 'time'.

Rules like Y form a critical part of CG. They not only provide a mapping from surface text to underlying con­ cepts, but also handle syntactic ambiguity in a unified and non-combinatorially explosive fashion. (Using rules A and B instead of Y would always result in two in­ terpretations, whereas Y chooses the best one.)

Another example of such a rule is

x np x <vp (passive)> x --> <specific event>

w hich handles forms like

The ship was attacked by the submarine Missiles were fired at ship by submarine Mi8siles were fired by submarine at ship

again, in non-combinatorially explosive fashion. Further­ more, this rule is able to search for active-voice events, thus allowing most event knowledge to be stored in active voice. If desired, this rule can transform passive voice sentences to active voice. Another critical rule is

<event-list> -----> <specific frame>

The task of this rule is to find a single frame (or scenario) which will unify a sequence of events, and to try to deter­ mine the causal relationship of all the events. This rule embodies some of the analyzer's frame-selection mechan­ isms.

3. RELATED WORK

VOX is a revised version of NOMAD !IJ.

CG developed from our attempt to reconcile the PHRAN approach !6] with traditional syntax theories, a.nd to sys­ temize the representation of phrasal knowledge.

We know of several efforts to build automatically extensi­ ble NLP systems: The Teacher component of UC !7], LIFER !4], KLAUS l3J, TEAM [2], and others. The suc­ cess of these efforts depends, ultimately, on the underly­ ing knowledge representation, including both conceptual and linguistic knowledge. We find that CG provides a framework for both productive and nonproductive linguis­ tic knowledge, while the other systems mentioned tend to concentrate on one type or the other. Also, CG alone provides for automatic addition of scenarios.

ACKNOWLEDGMENT

Thanks to Richard Granger for his support and for criti­ quing this paper, Karin Klein for proofreading it, and Laura Yoklavich for formatting it.