Mental Models of Meaning by P.N. Johnson-Laird (1981). Chapter in Elements of Discourse Understanding (Cambridge University Press).
I always pause a moment before diving into material that is decades old, wondering if perhaps the entire field has moved on and I just didn’t get the memo. Yet I have found, especially in AI research and computational linguistics, a lot of gems have laid buried that are worthy of revisiting and polishing off, and while I’m not sure if this is indeed one, at least I can practice my polishing techniques.
The chapter kicks off with my favorite kind of question - “what is meaning?” Mmmmmm. This is legit something I recommend asking at your next dinner party. The author references three predominate aveneus of addressing that question from with in semantic theory (psychology):
- words are represented in mental dictionary that decomposes meaning into sementic features
- words are linked in a network that represents their semantic relations
- “meaning postulates” are employed to specify the semantic relations between words
The author then goes into a bit more detail about each, knocks them around a bit, and then lays out an approach they advise called procdural semantics. Let’s follow along.
Decompositional theory of meaning
This theory has historic precedent, namely in semantic network theories, and essentially postulates that “the semantic interpretations of a sentence is obtained by replacing its words with their dictionary definitions, and combining them according to the syntactic relations of the sentence”. Some challenges to this theory are offered, particularly in regards to the nuances of decomposing words to their semantic constituents (e.g. modeling nuances of meaning depending on context), yet by in large the author extols the theory’s general merits, at least in broad strokes.
Semantic network theories
This section starts to sound more like what I’m used to in neural networks with NLP. The central assumption of semantic network theories is shared as being “that the lexicon should be treated as a networkthat interrelates representations of lexical items by a variety of inferential links”. The author is quick to point out that if you’re wondering, “wait a minute, if in decompositional theories we are breaking down words into their syntactic relations and then mapping their relations, wouldn’t those maps be networks and aren’t these really just the same thing?“, you are not alone. Others in the field have offered up such critique.
Meaning postulate theory of semantics
A reseaarcher by the name of Carnap put forth an idea in 1956 of “meaning postulates”, which argued that each word in a setence has a correlary word concept, and that these concepts are related to one another with meaning postulates.
The Smackdown (i.e. the case against all three of the above)
First of all, I love the quoted dictum of Fillmore (1974), that “issues in semantics that have no conceivable application to the process of comprehension cannot be very important for semantic theory.” Facts!
The author’s biggest beef with these theories is that the assume an “automomy” of semantics, that is to say they assume “the meaning of any sentene can be established entirely independently from what it may refer to”.
The Proposal: procedural semantics and mental models
Alright, let’s see what the author has in store for us.
“The psychological theory of meaning that I wish to advance assumes that the mental representation of a sentence can take the form of an internal model of the state of affairs characterized by the sentence”.
I feel it necessary to put on the table just how much my personal mental modeling prior to 2016 was predicated on finding somatic representations of the concepts being expressed. I called these representations “geoflora” with the idea that my feeling of a concept space was a multidemensional affect topology that would grow and develop over time as nuances of understanding were gleaned from learning. This style of representation has helped me excel in communication, culture building, the arts and management, but growing up it presented many challenges in learning, for instane, the significance of the difference between 15.72 and 15.83. I simply could not somatically distinguish the two, so for years such details were simply not graspable.
With that personal digression behind us, I return to the author’s share that, in principle, sceptics have argued “it is impossible to determine the nature of an internal representation.” I’m wondering what assumptions about modeling, technology, and human ingenuity are bounding that belief of impossibility. I have greater hope in our ability to model meaning as we advance as a species.
“The procedural theory postulates many internal proesses that cannot be couched in meaning postulates… . It represents the meanings of quantifiers without having to specify which inference schemata they permit, and builds up internal representations in which these logical characteristics are merely immanent”.