humanlike results. Readers who choose to skim
through the technical details are nevertheless
encouraged to reflect on how much reasoning people
actually apply to language understanding — automatically, without effort, and usually without even
noticing that there is anything to reason about.
Matching Recorded Constraints
The simplest case of language analysis is illustrated
by the example A brown squirrel is eating a nut. For this
input, like any other, the LEIA must disambiguate
each lexeme (that is, understand it as an instance of
a particular concept in its ontology) and combine
those interpretations into an overall semantic representation like the one in figure 1.
The representation in figure 1 is read as follows.
The first frame is headed by a numbered instance of
the concept INGEST, concepts being distinguished
from words of English by the use of small caps. 6
INGEST- 1 has three contextually relevant property values: its AGENT (the eater) is an instance of SQUIRREL, its
THEME (what is eaten) is an instance of NUT-FOODSTUFF,
and the TIME of the event is the time of speech, which
must be computed by the agent, if possible, using the
procedural semantic routine find-anchor-time (this
routine has not yet been launched at the stage of
analysis shown here). The next frame, headed by
SQUIRREL- 1, shows not only the inverse relation to
INGEST- 1, but also that the COLOR of this SQUIRREL is
BROWN. Since we have no additional information
about the nut, its frame — NUT-FOODSTUFF- 1 — shows
only the inverse relation with INGEST- 1. Developer
views of text meaning representations also include
many types of metadata, such as which word of input
gave rise to each frame, which lexical sense provided
the given interpretation, and so on.
Abstracting away from details of particular implementations of OS, 7 let us work through the analysis
process. First the input is syntactically parsed using a
parser developed externally from our system. 8 Then
the LEIA attempts to align the parse with the syntactic expectations recorded in the lexicon for the words
in the sentence. For example, when it looks up the
verb eat, it finds three senses: one is optionally transitive and means INGEST; the other two describe the
idiom eat away at in its physical and abstract senses
(The rust ate away at the pipe; His behavior is eating
away at my nerves!). Since the idiomatic senses require
the words away at, which are not present in our
input, they are rejected, leaving only the INGEST sense
as a viable candidate. 9 A simplified version of the
needed lexical sense is shown in figure 2.
The syntactic structure (syn-struc) zone says that
this sense of eat is optionally transitive: it requires a
subject and can be used with or without a direct
object. The semantic structure (sem-struc) zone says
that this sense of eat means INGEST. Each constituent
of input is associated with a variable in the syn-struc:
the subject is $var1 and the direct object is $var2.
Those variables are linked to their semantic interpretations in the sem-struc (^ is read as “the meaning
of”). So the word that fills the subject slot in the syn-struc ($var1) must first be semantically analyzed,
resulting in ^$var1 (the meaning of $var1); that concept can then be used to fill the AGENT role of INGEST.
For example, given our input A brown squirrel is eating
a nut, the LEIA links the word squirrel to $var1 then
semantically analyzes it as SQUIRREL before using it to
fill the AGENT role of INGEST. An analogous process
occurs for $var2/^$var2. The ontology, for its part,
constrains the valid fillers of the case-roles of INGEST
as shown in figure 3.
This excerpt from the ontological frame for INGEST
— which actually contains many more properties
and expectations about their values — says that its
typical AGENT (that is, the basic semantic constraint
indicated by the sem facet) is an ANIMAL; however, this
constraint can be relaxed to SOCIAL-OBJECTs (for example, The fire department eats a lot of pizza). Similarly,
Figure 1. The Text Meaning Representation for
A Brown Squirrel Is Eating a Nut.
AGENT SQUIRREL- 1
THEME NUT-FOODSTUFF- 1
AGENT-OF INGEST- 1
THEME-OF INGEST- 1
Figure 2. A Simplified Version of the INGEST Sense of Eat.
eat-v1: “ingest,” as in “He was eating (cheese).”
subject $var1 INGEST
root $var0 AGENT ^$var1
directobject $var2 (optional) THEME ^$var2