I~RAN AND PHRED: ANALYSIS AND I~ODUCTION USING A CO~0N 
KNOWLEDG~BASE 
Robert Wilensky 
Computer Science Dlwision, Department of EECS, University 
of Californ£a, Berkeley, Caltforn£a 94720, USA 
1 o0 Introduction 
We propose a model of language use that is derived from 
wlewl~ language processing systems as knowledge-based sys- 
tems. The knowledge that needs to be represented and organized 
here is the large amount of knowledge about what the utteran- 
ces of a language mean. In this paper, I describe some of the 
theoretleel underpinni~e of the model, and then desorlbe two 
programs, PHRAN and I~RED, that are based on these ideas° We 
have conducted a number of experiments with these systems 
that have some bearing on the utility of the model "s presumpt- 
ions, including testing these systems on other languages (Spa- 
hish and Chinese), and implementing one of them in a relation- 
al data base system. 
2.0. The assumptions of the model 
2°1° The Importance of Non-generative Lan~age 
Language user knows a great number of fact about what 
utterances of their language mean. That is, in addition to 
knowing the meanings of a large number of words, they know the 
steaL%license of a set of meaningful lingu/stto units that are 
not necessarily understood in term8 of thei~ components. Our 
conjecture is that such units constitute a very considerable 
fraction of the language knowledge needed by an intelligent 
language processor. 
- 296 - 
2.2. 3harable Knowledge Base 
In our model, it is assumed that the knowledge used 
for analysis and for production is by and laxge the same. 
That is, there is only one data base of knowledge about the 
meanings of a language's forms. By having the knowledge of 
the two components be a shared data base, only one form of 
representation is needed. Moreover, the addition of new know- 
ledge to thAs data base extends the capabilities of both sys- 
tems simultaneously. 
As thAs requirement forces knowledge to be represented 
deolaratively, the other benefits of such representations eme 
enjoyed as well. For exsunple, in this format, knowledge about 
the language is kept separate from the processing strategies 
that apply this knowledge to the understanding and production 
tasks. Thuue adding new knowledge requires only adding new 
asssztionB to the data base, not writing and debug~ new 
code. In addition, other knowledge besides the mesu~ng of a 
phrase can be easily associated with such declarative repres- 
entations. 
3.0. PHRAN and P~h~ED 
We have been developing this model of language use in 
two related programs, PHRAN (PHRasal ANs~yzer) and PHRED 
(PHRasal ~ngllsh Diction). PHRAN is a language understanding 
program written by ¥igal Arens. It reads English sentences 
and produces representations from them that encode their 
meaning. PHRED is a natural language production meohanAs~, 
developed by Steven Upstill. PHRED takes meaning representat- 
ions as input and expresses them i~ ~nglish sentences. 
Both PHRAN and PHRED shs~e a common data base of lang- 
uage knowledge. This data base contains declarative represent- 
ations about what the phrase of the English language mean. 
This knowledge is stored in the form of p attex~-conoeDt pears. 
A pattern is a phrasal construct of varying de~rees of speci- 
ficity. The concept part of a pattern-concept pair is a ogn- 
- 297 - 
ceptual template that represents the meaning of the associat- 
ed phrase. Together, these pairs associate different forms of 
utterances with their meanings, 
PHRAN understands by reading the input text and trying 
to find the phrasal patterns that apply to it. Eventually, the 
conceptual template associated with the desired pattern is 
used to generate the structure denoting the meaning of the 
utterance. PHRED produces sentences that encode an idea by 
examining the same knowledge base. 
4.0 Spanish and Chinese PHRAN 
We have build both a Spanish and a Chinese version of 
PHRAN simply by chansing the pattern-concept data base. These 
programs lend support to some of the claims we make for our 
model. We found that it was possible to rewrite most of the 
patterns into phrases of another language without havinsgthe 
knowledge encoder learn anything about the inner workings of 
the program. This suggests that a system like PHR~ could be 
designed to allow fairly easy construction of a language pro- 
cessor for a new language, or to allow for the addition of 
special purpose phrases or Jargon by some user who was not an 
expert AI programmer. 
5.0 AI and Relation Data Bases 
We implemented a version of PHRAR in a conventional da- 
ta base system. PItR.kN was re-written in EQEEL, a query langua- 
ge for the INGRES relational data base system developed at 
Berkeley. Tests were run to compare the relative perforemnce 
of the systems on various size data bases. 
The results can be summarized as follows: The LISP 
version is considerably faster when the data base of pattern- 
-concept pairs is small. However, when the data base is large 
(2000 words and 500 patterns), the EQUEL version is about 3 
times faster than the LISP version. Thus performance problems 
in natural lan~Aag@ may be solved by importing developments in 
data base technology as the size of our knowledge bases grow. 
- 298 - 
