![[Figure 1]](fig1.gif)
Figure 1 - Flow of Deductive Logic
The INFANT System was originally developed as a
conversational natural language processing system. It was later
adapted for practical use as an interface for standard operating system
tasks. Now it begins its third phase: as a general-purpose natural
language system that relies on massive propositional interaction within
a hybrid symbolic/connectionist framework to achieve an elementary
form of understanding.
1. Introduction
The INFANT System is a natural language processing (NLP)
system that attempts to achieve understanding through the analysis of
propositional sentence parts. Conversational input is broken down to
a hierarchy of standardized propositional forms, which are mapped
compositionally through an abstract data structure called the Semantic
Specification Language (SSL) to arrive at an overall meaning
representation. A connectionist-like network of inferences is
maintained to assist with meaning analysis.
Since its beginnings in 1985, INFANT has gone through
various phases of development. It was originally designed as a
conversational system for communication at the level of a small child.
1
This paper will review the work done in Phases I and II, and
outline plans for the work currently underway. Section 2 will discuss
the main features of the original conversational system. Section 3 will
describe the system as it performed within an operating system
domain, and summarize the results of the lab experiments. Section
4 will present a simple model for the anticipated treatment of
propositional data in the enhanced understanding system.
2. The INFANT Conversational System
Sentence analysis in the original system consisted of a number of overlapping syntactic and semantic modules. 1 Syntactic analysis was largely performed by the Predictive Wait-and-See Parser, which used linked lists to connect the words of a sentence in grammatical patterns. The parsing strategy was a form of parallel chart-parsing with a predictive phrase structure grammar, through which word expectations were progressively interlinked with subsequent matching words until only syntactically acceptable derivations remained.
Semantic analysis began with the conversion of a sentence to Hierarchical Logical Form (HLF), which expresses thoughts as a hierarchy of linked Subject-Predicate-Object propositions. For example, the sentence
Janie will have a party if she cleans her room
is converted to the following HLF:
JANIE HAVE (future) PARTY
|
JANIE CLEAN (if) ROOM
|
ROOM HER
The Subject-Predicate-Object triplet form is similar to the template
in Wilks' Preference Semantics.
3
Propositional linkage was proposed by van Dijk and Kintsch,
4
who used the term episode to
describe a connected series of propositions dominated by a single
macroproposition.
The INFANT System used the propositions from HLF to
generate new inferences with a systematic forward chaining process.
For example, the inferences
The massively interconnected knowledge base of propositions
was managed in part by a belief maintenance process.
6
Janie clean room => Janie clean place
=> Person clean place
=> Person fix place
=> Janie have party
=> Janie have fun
=> Janie happy _
would be part of the updated propositional knowledge base. These
propositions would in time be used to generate later propositions.
The highly combinatorial nature of inference generation led to an
attempt to manage it as a parallel processing operation in the
PINFANT System.
5
A number of problems plagued the early system: an
insufficient lexicon (although INFANT could learn new words);
difficulty in focusing on the main point of conversation; lexical and
structural ambiguities; failed intensional and common sense analyses;
the explosion of new inferences. It was soon apparent that a less
complex domain would have to be defined to demonstrate that the
concept of propositional interaction was of practical use.
3. The INFANT Operating System Interface
Phase II de-emphasized the connectedness of the original system and focused primarily on single sentence analysis for OS commands. It used an abstract data structure called the semantic specification language (SSL) to map propositional sentence fragments into an overall meaning representation. After user commands were translated into HLF, a compositional semantics procedure was used to match sentence propositions against the SSL and perform a bottom-up meaning analysis. The simplicity and preciseness of the propositional form facilitated its representation within the abstract framework of the SSL, and as a result it could be applied to a wide range of conversational English.
The SSL provided the link between a logical semantic form and a precise representation of meaning. It demonstrated the declarative nature of the INFANT language understanding mechanism through its direct mapping from sentence constituents into the commands possibly intended by the user.
The primary data structure in the SSL was a schema that defined a generalized propositional form derived from syntactic analysis. The most common schema form was
NVN: Predicatex(Subjectx,Objectx)
=>
{Meaning1 [Subjectx|Predicatex|Objectx] |
Meaning2 [Subjectx|Predicatex|Objectx] |
:
Meaningn [Subjectx|Predicatex|Objectx] }
(N1xVN2x: Subjectx=N1x, Predicatex=Vx, Objectx=N2x)
It should be noted that the SSL entries were encoded in phrase structure form {Subject-Predicate-Object (NVN) or Predicate- Preposition-Object (VPN)} rather than in predicate calculus notation. Thus a proposition such as ON (FILE, DISK) was represented as (FILE ON DISK). The SSL abstracted away from specific words and phrases to form a manageable set of propositional concepts, with a degree of specificity necessary only to differentiate between the subtleties of language present at the surface level.
The SSL's abstraction mechanism can be illustrated in an
example that also reveals the deceptive simplicity of language. The
sentence
Can I get FILEX on A:
is translated into two different HLFs:
(I GET FILEX / GET ON A:) {file to go on A:}
(I GET FILEX / FILEX ON A:) {file to be found on A:}
The SSL resolves the ambiguity by checking for the presence of
FILEX on A:. If FILEX is not found, the first HLF is chosen and
matched to the generalized form
PERSON GET N1 / GET IN N2 ,
which is translated by the SSL to the command
COPY N1 N2 {COPY FILEX A:} .
If FILEX is found, the second HLF is chosen, (FILEX ON A:) is
resolved to A:FILEX, and the top-level proposition (I GET FILEX)
is matched to the generalized form
PERSON GET N1 ,
which is translated by the SSL to the command
COPY N1 -- {COPY A:FILEX --} .
Fifty-four students from various introductory computer classes
at Harold Washington College were asked to perform a variety of file
transfers, deletions, and displays in two ways: (1) with the INFANT
TALK program, and (2) with the Windows File Manager. Although
some of the students had prior experience with Windows, none were
highly proficient with the File Manager, and none of them had ever
used TALK. The results of the experiment were as follows:
Average Time Required for
Category of Student Windows File Manager TALK
With Windows Experience (10) 1.5 min. 2.5 min. Without Windows Experience (44) 9.5 min. 12.0 min. All Students (54) 7.5 min. 10.5 min.
No. of students unable to finish with Category of Student Windows File Manager TALK
With Windows Experience (10) 0 0 Without Windows Experience (44) 6 1
It was clear that experienced Windows Users required less time than Non-Users, and that the students in general required less time for the File Manager than for TALK. The latter result may have been due in part to the naturalness of using a GUI rather than textual commands. It also reflects the frequency of interpretive errors made by the TALK System. Students were often required to rephrase their requests before the system responded as desired.
A possibly significant observation from the test results was that
very inexperienced students seemed to require more help with the
File Manager than with the TALK Program. Mistakes were made
with both approaches, but the inexperienced TALK Users seemed
better able to accomplish their given task through a trial and error
process. Several of the inexperienced Windows Users were unable to
recover from File Manager errors without asking for assistance.
4. Proposed Enhancements to the INFANT System
In current work, an attempt is being made to enhance the earlier conversational system with the improvements made in the lexicon, SSL, and compositional semantics algorithm during the OS implementation. Additionally, connectionist modules are being designed to replace processes with significant associative activity. The goals of common sense understanding and natural interactivity remain as before, although the immediate objective of this new phase is to test the learning ability of the model through the gradual addition of knowledge, specifically as it might be presented in children's stories.
One of the further goals of this new development is to manage the great quantity of inferences generated by the forward chaining process. In the enhanced model a type of spreading activation is performed with propositions, which can be likened to the neural "units of thought" of a connectionist network. A standardized propositional form assists in the organization of deductively inferred propositions, as does the gradual conversion of specific to generalized forms through inductive inference. Faster, larger parallel processing machines make mass storage more feasible. And the addition of connectionist modules in a large sense encourages the generation of inferences for the required neural network training process.
As an example of the enhanced reasoning process proposed for
Phase III, consider the following sequence of statements:
(1) Janie wants to have a party for her birthday
(2) Mom says that Janie has to clean her room first
(3) I guess we won't be seeing any balloons anytime soon
Figure 1 - Flow of Deductive Logic
Figure 1 traces the reasoning process through these statements.
Qualifying information such as modal operators (e.g., Mom says Janie
(must) clean room) have been left out of the figure. Also, all
propositions and inferences are stored in gradual degrees of
generalization. For example, the proposition Janie clean room is
stored as a single instance of each the following generalized forms:
Janie clean place
Janie fix room
Janie fix place
Person clean room
Person clean place
Person fix room
Person fix place
The flow of logic in Figure 1 reflects the stepwise changes in
INFANT's beliefs:
1. Janie wants to have a party
2. Mom says that Janie must clean her room first
(Inference: If Janie cleans room, then she gets a party)
3. I guess that no one will see any balloons
(Inference: There won't be any parties)
(Inference: Janie will not clean her room)
It must be emphasized that belief values and modality play a sizable
role in such an analysis. Statement 3, for example, does not clearly
state that everyone will fail to see any party objects. This is one of
the major difficulties faced in the proposed work. However, the
strength of the INFANT approach is its reliance on a connectionist-
like analysis of generalized propositions qualified by belief values.
The enhanced model is designed to verify the practicality of this
approach on a larger scale.
5. Summary
The cognitive model of the INFANT System has as its foundation the concept of propositional sentence content -- that is, sentences can in general be broken down into discrete propositional segments that through their very form and organization contribute to the understanding process. The INFANT propositional form is highly standardized and intuitively simple, and thereby able to help resolve the demanding storage and search requirements of a symbol-based architecture. The organization of the propositional knowledge base is connectionist-like in that all old, new, and inferred units of information are interlinked in a type of "spreading activation" of propositions.
Previous work shows some of the merits of the propositional
approach. Early work laid the groundwork for the massive interaction
of propositions, while the operating system implementation showed
the practicalness of a largely computational application. It is the
purpose of continuing research to show that a working hybrid can be
formed of the symbol-based and connectionist components that will
better achieve understanding from the elementary fragments of
conversational input.
References
1 Buchheit, Paul (1991). INFANT: A Connectionist-Like Knowledge Base and Natural Language Processing System. Ph.D. Thesis, University of Illinois at Chicago. Back to text
2 Buchheit, Paul (1995). 'A Conversational Interface for File Management in an Operating System.' Proceedings of the 1995 Midwest Artificial Intelligence and Cognitive Science Society Conference, Carbondale, IL., pp. 43-47. Back to text
3 Wilks, Yorick (1976). 'Parsing English II.' In Computational Semantics (Eds. E. Charniak and Y. Wilks), pp. 155-184, North- Holland, Amsterdam. Back to text
4 van Dijk, Teun A. and Kintsch, Walter (1983). Strategies of Discourse Processing (Academic Press, New York). Back to text
5 Ng, Anna Man-hua, Evens, Martha, and Buchheit, Paul (1993). 'The Parallel Information Network for a Natural Thinking System.' Proceedings of the 7th Annual Midwest Computer Conference, University of Wisconsin-Whitewater, March 26, 1993, pp. 135-9. Back to text
6 Buchheit, Paul (1992). 'Belief Maintenance in a Natural Language System.' Fourth International Conference on Tools with Artificial Intelligence, Arlington VA, November 10-13, 1992, pp. 293-300. Back to text