A Monthly Column for Word Lovers
How Do You Solve a Problem Like Meronymy?
Given a suitable context, if a farmer told you that her hand was under the knife, you would probably understand that this was a sentence about an employee and a surgical operation, despite there being no mention in the sentence of a person. There's also no mention an operating room, a doctor, a hospital, or any of the other props or venues associated with surgery. But you got the meaning in the phrase "under the knife" in the same way that you got employee from hand, a clipped version of hired hand. Your understanding of these phrases is probably not based on inferring the relationship of "hand" to employee or "knife" to surgery; chances are that you know these terms because you've heard them before. Maybe the first time you heard them you had to do that kind of interpretation; or maybe you looked up the terms in a dictionary or maybe someone glossed them for you. In any case, what hand and under the knife have in common is that they're both instances of meronymy, in which a part of something is used as an expression for the whole.
English majors may already be raising their hands or racing to the comments section to protest that this isn't meronymy; some will say that we're actually talking about metonymy, and others may chime in with synecdoche. Well, everyone's right in this game, albeit in a slightly different way. The ways in which expressions substitute parts for wholes, or features for whole entities, is a slippery business. The figures of speech that stray from the literal to instantiate a part-whole or a feature-entity relationship constitute a significant part of what makes linguistic expression so rich, flexible, complex, and interesting. For human speakers, it's a lifelong learning opportunity and a palette from which to paint word pictures and create new expressions. For computers trying to understand language, it can be an intractable problem.
Linguistic expression aside, meronymy is simply a name to characterize the relationship that is also called "part-whole" or "is-a-part-of." Metonymy is the substitution of the name of an attribute or feature of something to stand for the thing itself. So meronymy and metonymy are two different things, right? It would be convenient to draw a solid line between them, but it turns out that we don't really do this very effectively in our heads, thus making it harder to define a consistent division between the two for computational purposes. Studies have shown that when people are asked to list properties of objects, they do not distinguish between attributes and parts. One reason for this may be that part is a highly polysemous and high-frequency word in English. As a noun it consistently has more than a dozen senses in most dictionaries. By some counts it is the tenth most common noun in English, and it is the second-most common word appearing in definitions of nouns in one dictionary study. Finally, among the many definitions of feature and some of attribute, also both highly polysemous, you will often find a few instances of the word part.
How would a computer, tasked with deciding in a given sentence whether the name of a part or feature is being used to indicate the whole, go about its work? It's at least a two-part exercise, and the first part presents fewer challenges. In a lexical database like the Visual Thesaurus, for example, hundreds of relationships are coded as part-whole, such as leg-chair (component part relationship), relative-family (member-of relationship), and flesh-body (constituent material). If these were the only kinds of meronymy, the first part of the problem might be solved. But meronymy is not a monolith, and there are many different types of it, as well as different ways to divide it up. To put it a different way, "a slice is a part of a pie in quite a different way than a lieutenant is a part of an army." (That's a quote from linguist Lynne Murphy in her book Semantic Relations and the Lexicon.) One classification of meronymy has at least three additional kinds: portion-mass (exemplified in grain-rice), feature-activity (as in dating-adolescence) and area-place (as in Chicago-USA). Relationships such as these, if not definitionally encoded as participating in a part-whole relationship, would be challenging for a computer to decipher.
The second, and much more challenging part of the task is to decide, in a given sentence, whether a word is used meronymically or whether it represents a more literal meaning of its referent. The human approach to this may go something like this: you consult the possible interpretations of the word, calling on your lexicon (literal meanings, figurative ones, and idioms you are already familiar with) and use an instinctive approach that would capture possible matches. If you find an interpretation of the sentence that makes sense in the context, you stop. If nothing works there, your native language ability would probably prompt you to make a few inferences based on meronymic or metonymic relationships till you found a meaning that suited the context.
The problem for computers here is twofold: they don't really have a good handle on what makes sense (i.e., is consistent with the consensus view of reality) and what doesn't. And it's hard to imagine what kind of complex programming would be required for a computer to perform an operation analogous to the one undertaken by the human imagination when it consults possible figurative meanings of a word based on part-whole and feature-entity relationships. The most sophisticated computer today can only capture a fraction of these, and so what a computer can do is not so different from the fable of the blind men and the elephant: each of three blind men has an opportunity to touch a part of an elephant and then describe what the creature is like. Of course, they all disagree because each has experienced only a part of the beast.
There's an allusion in the title of the Lounge this month that some readers will have noted, to a Rodgers and Hammerstein lyric. The song is "Maria" from The Sound of Music, and its first two lines are:
How do you solve a problem like Maria?
How do you catch a cloud and pin it down?
The problem of meronymy is exactly that: the human mind holds a cloud of word and meaning associations evoked by a word in context and pins down the associations that matter for the sentence at hand, discarding for the time being the myriad other content of the cloud. Will computers ever be able to do that? Human languages have had millennia to develop their webs of word and meaning relationships that provide speakers with ready ways of understanding sentences like "They pumped him full of lead" or "Do you want paper or plastic?" or "The strings were drowning out the brass." Computers today are taking halting steps to try to catch up.