At the end of my previous post discussing Carissa Hessick’s paper “Corpus Linguistics and the Criminal Law,” I said that I would follow up with another post “making the affirmative case for the relevance of frequency data in determining ordinary meaning.” This is that post.
Given that subject, you might wonder why I’ve titled this post “Meaning in the framework of corpus linguistics.” The answer is that corpus linguistics has not only provided a methodology for investigating meaning, it has also generated important insights about word meaning. (That was the subject of the paper I presented at the BYU symposium in February, which will be published, along with the other papers from the symposium, in a special issue of the BYU Law Review.) I’ll draw on those insights when I talk about frequency analysis, and I thought it would be helpful to make them explicit.
THERE ARE A VARIETY OF DIFFERENT WAYS to think about word meanings. One of them is the way that I see as characteristic of how lawyers and judges tend to think: the meaning of a word is more or less equated with its dictionary definition, and then the definition is in effect read into the statute. If you’ve read a lot of cases, you’ll probably recognize the pattern:
The issue here is what “flood” means. Webster’s Dictionary defines “flood” to mean, “a great flow of water over what is usually dry land.” Therefore, the plaintiffs must show that the water in their basement resulted from a great flow of water over what is usually dry land.
Under this approach, the dictionary entry is treated as if what it defines is the concept flood rather than the word flood. The dictionary entry is being used as stating the conditions determining whether a particular instance of water on the ground qualifies as a flood. Considering the role that dictionaries have come to play in legal interpretation, it is no small irony that many lexicographers would say that the definitions they write aren’t intended to serve that purpose.
A different way of thinking about word meanings and the function of dictionaries (one that characterizes how I approach these issues) derives from corpus linguistics, and in particular from corpus lexicography, along with a helping of pragmatics. I’ll refer to this as the usage-based approach to word meaning, in contrast to the dictionary-based approach that’s described above.
The usage-based approach has a number of facets, one of which is that “meaning is something you do.” That’s a quote from the late Adam Kilgarriff, a computational linguist whose dissertation investigated the nature of dictionary word senses, and who did important work at the intersection of lexicography and computer science. The same idea has been expressed, in more detail, by two of the most important lexicographers of the past 50 years, Patrick Hanks and Sue Atkins. Hanks says that “[in] the everyday use of language, meanings are events, not entities.” And in the Oxford Guide to Practical Lexicography, Atkins and her coauthor say, “[M]eanings and dictionary senses aren’t the same thing at all. Meanings exist in infinite numbers of discrete communicative events, while the senses in a dictionary represent lexicographers’ attempts to impose some order on this babel.”
Kilgarriff describes dictionary senses, as distinguished from meanings, as representations of “corpus clusters”—groupings of corpus lines in which the word in question is used in more or less the same way. The clustering process inevitably requires judgment calls, including the decision whether to lump related senses of the word into a single broad, general cluster or instead to split them into separate narrower, more specific clusters. Whatever the decision is, the various clusters provide the basis for the word’s entry, with a separate sense corresponding to each cluster. And although Kilgarriff. described this process in the context of using corpus data, lexicographers do much the same when they work from examples of usage (called “citations”) that were collected by hand.
This process of grouping corpus lines into clusters is what underlies many of the uses of corpus analysis in legal interpretation. As you can see, it uses essentially the same method of collecting and analyzing data as dictionary-making does. But there is an important difference. In writing dictionary entries, the goal is to write a short explanation summarizing what the uses in each corpus cluster have in common, without attempting to capture every nuance of the word’s semantics. In legal interpretation, on the other hand, issues of word meaning typically turn on matters of nuance: does a few inches of water in the basement count as a “flood”; does driving somewhere with a handgun in the glove compartment count as “carrying a firearm”? With that in mind, the clustering process can zoom in on the specific aspects of meaning and use that are at issue.
ONE OF THE PREMISES of the usage-based approach to word meaning, in both corpus-based analysis and traditional lexicography, is that the patterns of actual usage constitute the subject of inquiry—the raw data to be analyzed—and that those patterns provide evidence of the various meanings that a word is used to convey. And in fact, actual usage does more than simply provide evidence of a word’s meaning(s), it determines those meanings. To state it differently, word meaning arises from usage.
When children first learn to understand and use language, they acquire the meanings of words by hearing them used by the people around them and inferring from those uses what the words mean. This process continues during and even after childhood. Take the word kerfuffle, which seems to me to have increased in the frequency of its use over the past decade or so (although that may just be the recency illusion in action). If you’re like me, you understand a kerfuffle to be a minor-league brouhaha, and you acquired that understanding as if by osmosis, from seeing and hearing the word being used in real-world contexts.
Moreover, in order for a particular sense of a word to stay a part of the language, it has to continue to be used in that sense. Consider the word ascertain. When Samuel Johnson wrote The Plan of an English Dictionary in 1747, in an effort to get the support of a wealthy patron, he said that he wanted to produce “a dictionary by which the pronunciation of our language may be fixed, and its attainment facilitated; by which its purity may be preserved, its use ascertained, and its duration lengthened.” To us today, Johnson is saying that he wanted to find out the use of the English language, which seems strange to our ears and out of sync with his desire for “the pronunciation of our language to be fixed” and “its purity…preserved.” But that’s because we understand the word based on how it’s currently used, and that’s not how Johnson used it. Rather, he used it to mean “to make certain; to fix; to establish” (to quote the definition he ultimately wrote for his dictionary). The OED lists that sense of the word as being obsolete, and the most recent example it gives for that sense is from the U.S. Constitution, in 1789: “The Senators and Representatives shall receive a Compensation for their Services, to be ascertained by Law, and paid out of the Treasury of the United States.” So despite Johnson’s desire to ascertain English usage (in the sense of making it certain and fixed), he was unable to ascertain (in that sense) that use of ascertain. Because that use died out and was ultimately forgotten, ascertain no longer means what it did in the 18th century.
In short, it is the continuing use of a word in a given sense that maintains that usage as part of the language. And just as the usage that is encountered by individual speakers of English determines what English words mean for them, overall usage over time determines the meanings of words for the language as a whole. Even when a word sense remains stable over time, it is the repeated use of that sense (along, perhaps, with widespread circulation of texts in which the sense is used) that keeps that sense alive.
WHAT THIS HAS TO DO WITH CORPUS LINGUISTICS is that it lays the foundation for understanding why corpus data is relevant to the decision of ordinary-meaning issues. To begin with, word-meaning is a matter of convention, and corpus data is
direct evidence of the ways of using the language that are conventional. Assuming that the corpus includes a representative and balanced set of texts, it should provide a picture of the conventional ways in which words are used. And the more frequent a particular sense of the word is in relation to other senses that occur in comparable contexts, the stronger is the argument that the usages in the former category are conventional. That is obviously relevant to the issue of ordinary meaning.
When one considers the conventionality or ordinariness of a given way of using a word, the focus will tend to be on the word’s meaning as some kind of abstract entity independent of the reader or listener’s understanding of the use of the word. But conventionality can be relevant to the issue of how the usage is understood. It seems reasonable to expect that when a word is used in a conventional way, it will be understood in accordance with its conventional meaning. After all, conventions of word meaning are tacit agreements about meaning, and those agreements are shared across the entire speech community. Indeed, they are a big part of what makes the speech community a speech community. So one would expect readers and listeners to understood conventional forms of expression in accordance with the conventions.
But that is not all. It seems reasonable to expect that the higher the frequency with which a particular sense is associated with a particular type of context, the more likely it is that when the word is used in similar contexts in the future will be understood to have the same meaning. In this way of looking at the issue, the corpus data is seen as a rough representation of the input—what people hear and read—that shapes people’s understanding of word meanings.
It is known that the frequencies with which specific patterns and constructions occur has an effect on how language is learned and mentally processed. It doesn’t strike me as a big leap from that to the conclusion that the strength of the association between the use of a word in a particular type of context will depend at least in part on how frequently one encounters that word being used, in that type of context, to express that meaning. Or at least that seems reasonable if the idea of “types of context” can be adequately defined.
I will get to that, but first I need to note that I’m extrapolating from what I know of the literature, not reporting on what it says; I can’t point to any work that has been done on this specific question. Nevertheless, the assumption I’m discussing is, as far as I know, consistent with the fact that frequency effects are so widespread.
So what do I mean when I talk about there being identifiable “types of context”? To answer that question, I need to get down to the nitty gritty of what meaning means in the framework of corpus linguistics.
ONE OF THE INSIGHTS CORPUS LINGUISTICS HAS GENERATED is that although we are used to thinking that the basic units of linguistic meaning are words, it would be more faithful to reality to think of the basic units of meaning as being phrases and other multiword constructions. In isolation, many words, especially frequently-occurring words, are ambiguous. Think of all the times you’ve looked in a dictionary and seen multiple definitions for a single word. But when those same words are used in a sentence, the ambiguity disappears. The context acts to disambiguate the word; that is well known, but it leaves open the question of what specific aspects of context are responsible for that effect.
Before the 1980s, that wasn’t an issue that was studied by English-language lexicographers, by mainstream linguists, or by scholars of legal interpretation. But beginning in that decade, one of the pioneers of corpus linguistics and corpus lexicography, John Sinclair, argued that in analyzing language in use, one had to look beyond individual words to multiword expressions and other patterns. He said that “many if not most, meanings require the presence of more than one word for their normal realization” and that there is a “direct connection” between meaning and “patterns of co-selection among words.” And he showed that the sense in which a word is used will often depend on the grammatical environment in which it appears.
Let me give some examples relating to the verb carry, which was at issue in Muscarello v. United States. What we’ll see is that the word’s meaning changes depending on the grammatical construction it appears in and the words that fill the other slot(s) in the construction. We’ll start with intransitive uses, of which I can only think of two. First, there are uses in which the verb’s subject is a noun denoting a sound: sound carries better over water, her voice really carried. Constructions like these are used to express the idea that the sound or voice can be heard from far away. In contrast, in the context of parliamentary procedure, the motion carried means that the motion was approved. Similarly, contrast the meaning of the intransitive sound carries and voice carries with the meaning of transitive equivalents such as these:
As for [Johnny] Hodges, the altoman’s rich, golden sound carried a message about the possibilities latent in music that surely affected Coltrane’s own conception. (link)
Per usual, she spoke so properly her voice carried a hint of a British accent even though she’d been born and raised in Georgia. (link)
Rather than communicating something about the loudness of a sound, these examples communicate something about the content of the sound — about what impression the the sound conveys to those who hear it.
When we look at transitive verbs more broadly, we see a pattern in which the intended sense of carry is selected by the choice of the noun denoting what is being carried (i.e., the direct object in active-voice constructions, the subject in passive-voice constructions): carry a bag of groceries versus carry a message versus carry a germ versus [of a candidate for election:] carry a state. Also relevant is the choice of the noun denoting who or what is doing the carrying. If we refer to a horse or mule as carrying a burden, the burden will be understood to be physical, but if we make the same statement as to a person, the burden will often be understood as something intangible: a burden of responsibility, of work, etc. Similarly, the meaning of carries a lot of weight differs depending on whether we are talking about a truck or about somebody’s opinion.
All of this merely scratches the surface. And what it suggests is that instead of talking about the meanings of words in isolation, it makes sense to talk about the meanings of the patterns in which words occur. A pattern would typically consist of a grammatical construction, in which the word in question fills a specified slot (e.g., the verb in a transitive construction), together with either the words that can fill the other slot(s) or the semantic characteristics of such words (e.g., human, animate, physical object, organization, mental process).
ONE OF THE BENEFITS OF THIS WAY OF THINKING is that it provides a way to talk in a disciplined way about different types of contexts, and to identify corpus examples in which the word at issue in a given case appears in the same type of context that the statute uses it in. To show what I mean, I’m going to continue using carry as an example, and discuss it in the context of the issue raised by Muscarello: whether the act of driving somewhere with a handgun locked in the glove compartment falls within the ordinary meaning of the phrase “carry[ing] a firearm.” (The Supreme Court held that it does.)
The first corpus analysis of this issue was done by Stephen Mouritsen in his 2011 article, “The Dictionary Is Not a Fortress: Definitional Fallacies and a Corpus-Based Approach to Plain Meaning” (pdf). He found that among the potentially relevant corpus lines he examined, those that were consistent with the Court’s reading were only about one-sixth as common as those that were not. However, Mouritsen didn’t perform the kind of pattern analysis I’ve been discussing (at least not explicitly). Therefore, I undertook such an analysis as part of the paper that I wrote for this year’s law-and-corpus-linguisitics symposium at BYU. You can read the complete writeup of my analysis in that paper; for this post I will focus on what is most important. (The summary below is adapted from the paper.) [Update: Be sure to read the comments I have added about the paper at the end of the post. Also, note that the corpus data that the analysis is based on is available here.]
In order to filter out corpus lines in which carry was used in a sense that was not potentially relevant, I used the idea of semantically categorizing the collocates of the verb. Deciding on the appropriate categories was easy. The statute prohibits actions by humans, and firearms are tangible objects, so the relevant corpus lines would be those that denoted the carrying of tangible objects by humans. Thus we have the initial categories human and object, and the basic conceptual schema human carry object.
Within the set of corpus lines instantiating that schema, I was interested in seeing what was indicated by each line with respect to the manner in which the object was carried: in the person’s hands or arms, strapped to their back, in the trunk of their car, and so on. Therefore, I attempted to categorize the manner of carrying that was reflected in each concordance line that denoted the carrying of a tangible object by a human. I set up two categories, each one corresponding to a separate sub-schema: human carry object in vehicle, which encompassed events analogous to those in Muscarello and human carry object, which covered everything else.
Consistent with what Mouritsen had found, the corpus lines categorized as human carry object in vehicle—i.e., those corresponding to the defendants’ conduct in Muscarello—were greatly outnumbered by those categorized as simply human carry object. The data also pointed to an additional conclusion, one that was possible only because of the kind of pattern analysis I have referred to.
With respect to the both of the categories, I looked to see how often the manner of carrying was explicitly encoded in the corpus line (for example, by a prepositional phrase such as in his arms or in the trunk of his car.) And each category’s results were virtually the mirror-image of the other’s. Within the human carry object category, the great majority of lines did not encode the manner in which the object was carried, leaving that fact to be inferred. The opposite was true of the category human carry object in vehicle; most of the uses in that category did explicitly encode the fact that the object was being carried in a vehicle, as in the following examples:
He dug out the flashlight he carried in the glove box and clicked it on.
… no reason to keep carrying stuff in our trunk
Two were charged with stripping parts from a parked car and three with carrying a sawed-off shotgun and drugs in a car, police said.
In the front seat, I’m carrying a gadget slightly larger than an electric shaver.
In all but one of the remaining lines, the manner-of-carrying information was partially encoded, by which I mean that there were explicit references to vehicles (underlined in the examples below), which prompted inferences that the object was carried in that vehicle:
Shagren said what sparked the proclamation was concern over truck drivers carrying dairy products not being able to drive more than 12 hours a day…
He was thrown into the back of a deuce-and-a-half truck and carried south in a pile of wounded and dying men…
We returned with the truck to carry loot for the soldiers, mainly furniture.
Only one corpus line instantiating the human carry object in vehicle schema did not explicitly refer to a vehicle, but even that line can be understood as referring to vehicles by implication:
and the other delivery men at Yasgur’s were planning to carry milk to the hordes…
The conclusion I drew from all of this was not only that the pattern [human] carry [object] is seldom used to refer to events in which the object is carried in a vehicle, but also that the pattern is not used to express that meaning without the text or utterance containing some kind of overt indication to that effect. In other words, the data suggested that the pattern [human] carry [object] has a default meaning. In the absence of something in the text indicating something different, that pattern is used to refer to acts of carrying in which the person is physically bearing the object—in their hands, in their arms, on their back, and so on.
This strikes me as a finding that is pretty significant for purposes of deciding the issue in Muscarello. After all, when the statute prohibited the carrying of firearms, it said nothing about the manner or mode of the carrying (other than restricting the prohibition to cases in which the firearm was carried during and in relation to a drug offense or crime of violence). Therefore, the analysis I’ve summarized provides strong support for the argument that driving with a gun in the glove compartment does not fall within the ordinary meaning of carry a firearm.
THAT CONCLUSION WAS MADE POSSIBLE only by conducting a frequency analysis. But note that the object of the frequency analysis was not simply to find the way in which carry was used most frequently. Rather, by expanding the focus from the word carry itself to the patterns in which carry was used, it became possible to limit the analysis to corpus lines in which carry was used in a context similar to the context in which it was used in the statute.
In short, the analysis combined corpus methodology with an approach to lexical meaning that corpus methodology made possible. Thus: meaning in the framework of corpus linguistics.
In closing, I have two additional points. The first is that there are almost certainly other cases in which it would be appropriate to invoke the notion of a default meaning that applies in the absence of contextual support for a different meaning. One that occurs to me off the bat is Smith v. United States, which concerned the phrase use a firearm as used in the same statute as Muscarello. So keep this issue in mind the next time you’re dealing with an issue of lexical meaning (which is probably a better label for what I’ve been referring to as word meaning).
The second point is that pattern analysis of the same general kind that I’ve discussed won’t necessarily be appropriate in all cases (just as corpus analysis won’t necessarily be appropriate in all cases). Hopefully, as we get more experience with using corpus linguistics in legal interpretation, it will be possible to identify the kinds of cases that are the best candidates for using corpus analysis, and within that category, to identify those that are good candidates for pattern analysis.
There are interesting times ahead.
Update: Asher Steinberg has posted an epic comment to this post, in which he argues that my discussion of the corpus of the corpus data regarding carry is fundamentally misguided. And I assume that his argument would apply as to my entire discussion of why frequency data is relevant. I will post a response, hopefully within a few days, but in the meantime I want to make a few quick points. [Update to the update: It took more than a few days, but my response is posted here.]
My impression from Steinberg’s comment is that he hasn’t read the paper that (as I said in the post) my discussion was based on—or at least that he hasn’t read the paper’s discussion of the corpus data. The paper makes some points that are relevant to some of what Steinberg says, but he doesn’t mention them. Hence, I assume that he isn’t aware of them. I would recommend that Steinberg, and anyone who is interested in his argument, read at least the part of the paper that devoted to the corpus analysis (pages 41-52). And if you have any questions about how the data was coded, the data can be downloaded here.
As I say in the post, my discussion there of the corpus analysis focuses on what I think is most important. In case that statement didn’t make it clear, let me say now that the post doesn’t attempt to deal with everything that is covered in the complete corpus analysis. The purpose of the post was to explain why frequency data can be relevant in deciding issues of ordinary meaning, not to comprehensively discuss the merits of the decision in Muscarello.
One of the things that the post doesn’t discuss is my analysis of the corpus lines in which carry was collocated with words denoting firearms. I say in the paper that those lines “arguably lend some support to the Muscarello majority’s interpretation,” but ultimately conclude that “arguing that this data justifies the holding in Muscarello is a stretch.” I continued: “It would be more accurate to say that the data provides hints toward a possible argument, but that the evidence is impressionistic at best and ultimately doesn’t have much substance.”
Finally, let me point out something that I didn’t mention in the post, but that is arguably relevant to the broader question of how the statute in Muscarello should be interpreted. Within the community of (presumably law-abiding) gun-owners, the phrase car carry is used to denote having a gun in one’s car or truck. (Compare open carry and concealed carry.) That’s something I wrote about in this blog’s early days, in 2011 (“Interpretive communities and car carry”). For purposes of interpreting the statute, the significance of this expression will depend on how much weight should be given to a usage that seems to be prevalent only among a subset of the general public.
Note: The title of this post is cribbed from the title of a paper by John Sinclair. And as I’ve noted, portions of the post have been adapted from my paper with the similarly cribbed title “A Lawyer’s Introduction to Meaning in the Framework of Corpus Linguistics” (forthcoming, B.Y.U. L. Rev.).