This article is part of a series entitled Redividing Linguistics. See also:
Morphology and Syntax: What’s In A Name?
- Syntax and Semantics: The Search For Meaning
- Semantics and Pragmatics: The Reality Interface
In computer science, compilers translate the relatively human-friendly languages — Java and C++, for example — into the monotonous sequences of additions, memory loads, and logical operations which the computer’s hardware understands.
Syntax and semantics are crucial elements of the process. Before a program can be translated, it must be ensured to be well-formed. This involves checking both that the program’s recursively nested functions, blocks, and statements are properly organized — syntax — and that the program doesn’t feature contradictions in meaning — semantics. The two processes are entirely distinct, and can even be performed in separate passes.
Not so in linguistics. In natural languages, syntax and semantics are not only hard to rigorously distinguish, but they’re inextricably intertwined.
Syntax describes the well-formedness of sentences. Generative grammars are formal systems which dictate how abstract phrases can be replaced by abstract sub-phrases — through so-called production rules — recursively. These rules dictate grammatical sentences’ complex nesting structure of clauses, phrases, and words. Production rules aren’t enough, though. Given a sentence generated by production rules, we must still ensure that various dependencies across words — fittingly called syntactic dependencies — are adequately upheld. (These dependencies actually resemble the checks characterized by computer scientists as semantic.)
Semantics in linguistics, on the other hand, studies the meanings of well-formed sentences. Linguists distinguish between pragmatic meaning — which requires facts from the real world — and semantic meaning — which is purely linguistic. The latter is semantics’ concern. Linguists don’t study which sentences are true. This would require pragmatic knowledge. They study entailment relations between sentences. (One sentence is said to entail another if provided one is true, the other becomes true.) Entailment relations permit linguists to isolate purely linguistic knowledge.
In linguistics — unlike in computer languages — syntax and semantics are deeply linked.
Negative Polarity Items
Consider the following sentences: (The * symbol indicates an ungrammatical sentence.)
“a. The elves did not tell anyone that the fairies could like witchcraft.
“b. *The elves told anyone the fairies could not like witchcraft.
“c. Nobody told anyone that the fairies could like witchcraft.
“d. *The news that Shylock did not leave surprised Portia at all.” (Fromkin, et. al, 222)
Why are sentences (a) and (c) acceptable, while sentences (b) and (d) are not? The very complex relationship is described in detail in (Fromkin, 198-223). Linguistics writes, “A sentence containing an NPI is well formed only if this NPI is in the c-command domain of a negative element.” (223) The technical details need not worry us here.
What we have here is a complex syntactic dependency. For a sentence to be generated by the language’s grammar is not enough. Its constituents must also uphold certain relationships. In our particular case, NPIs (negative polarity items) must stand in a special relationship with so-called negative elements. This relationship is a syntactic one.
But how do we determine which words are negative elements? Linguistics turns explicitly to semantics! Defining so-called decreasing determiner phrases “on the basis of their entailment relations” (404) — that is, semantically — Linguistics concludes that “all decreasing DPs are negative elements, in the sense required by the NPI Licensing Rule.” (405) Negative elements have been characterized semantically.
Syntax demands, then, a relationship involving entities whose identities can only be determined semantically. Structure cannot be checked without also knowing meaning. This phenomenon will prove to be pervasive indeed.
Negative elements in particular are characterized semantically. What about the constituents of other syntactic dependencies?
Verbs typically demand that their arguments satisfy particular properties. These requirements comprise a much broader class of syntactic dependencies, called selectional restrictions. Consider the following sentences:
“a. This moat surrounds his castle.
“b. *This moat surrounds two minutes.” (Fromkin, et. al, 197)
Why is (b) incorrect? The verb “surround” enforces additional restrictions on the semantic properties of its arguments. This verb’s object, Linguistics writes, “must refer to a concrete object like castle, not a measure of time like minutes.” (197) Whether a verb’s arguments satisfy its selection requirements depends on semantic information about those arguments.
Further, how do we know a verb’s selection requirements in the first place? How do we know, for example, that “surround” requires a direct object, of type determiner phrase, which happens to be a concrete object, while “arrive” demands no object at all? Verbs’ selection requirements must themselves be semantically derived.
We did not use entailment to construct this semantic information. Is it still semantic? I’m not sure how I’d use entailment to conclude that “two minutes” is not a concrete object or that the verb “surround” demands one. I’d still consider this information suitably semantic. It appears more semantic than did our abstract characterization of negative elements.
Selection requirements are cast as syntactic dependencies. To uphold them and even to formulate them, though, we need semantic information. The stipulations of syntax are implemented using the fruits of semantics.
We’ve advanced steadily upon syntax, continually demonstrating that progressively more-integral components of its system of dependencies require semantic information to formulate and uphold. What about the core generative grammar itself?
A similar phenomenon confronts us. The English grammar might stipulate — to pick out a handful of production rules — that a Determiner Phrase produces a Determiner followed by a Noun Phrase, which, in turn, produces a Noun possibly followed by a Prepositional Phrase. These rules are purely (morphological or) syntactic. Checking whether they’re being upheld, though, would, again, require semantic information. We must know whether any given word happens to be a determiner, or a noun, or a preposition. We must know what it means.
The lexical category of a word is generally considered part of its syntactic information. The grand message, though, has arrived. It’s not clear that a word’s lexical category should be classified differently than its selection requirements. It’s not clear why these selection requirements should be classified differently than its entailment information. It’s not clear at all, indeed, why any of this information should be classified as syntactic or semantic.
Syntax and semantics are complex notions involving rules and information. The required information, though, is divided across the headings of syntax and semantics. Why can’t we unite it? Uniting this information under a single banner, we could generate grammatical sentences, check “syntactic” dependencies, and study “semantic” entailment, using a single system.
Let’s redivide theoretical linguistics. Rules for word assembly and sentence formation — traditionally, morphology and syntax — can be classed together as structure. The information required for the implementation of this structure — traditionally, divided across syntax and semantics — can be classed together as meaning. The two are closely linked.
Next, we’ll investigate meaning.
Reference: Fromkin, V, et. al. (2000). Linguistics: An introduction to linguistic theory. Malden, Mass.: Blackwell.