Field of Science

Results (Round 1): Crowdsourcing the Structure of Meaning & Thought

Language is a device for moving a thought from one person's head into another's. This means to have any real understanding of language, we also need to understand thought. This is what makes work on language exciting. It is also what makes it hard.

With the help of over 1,500 Citizen Scientists working through our VerbCorner project, we have been making rapid progress.

Grammar, Meaning, & Thought

You can say Albert hit the vase and Albert hit at the vase. You can say Albert broke the vase but you can't say Albert broke at the vase. You can say Albert sent a book to the boarder [a person staying at a guest house] or Albert sent a book to the border [the line between two countries], but while you can say Albert sent the boarder a book, you can't say Albert sent the border a book. And while you say Albert frightened Beatrice -- where Beatrice, the person experiencing the emotion, is the object of the verb -- you must say Beatrice feared Albert -- where Beatrice, the person experiencing the emotion, is now the subject.

How do you know which verb gets used which way? One possibility is that it is random, and this is just one of those things you must learn about your language, just like you have to learn that the animal in the picture on the left is called a "dog" and not a "perro", "xiaogou," or "sobaka." This might explain why it's hard to learn language -- so hard that non-human animals and machines can't do it. In fact, it results in a learning problem so difficult that many researchers believe it would be impossible, even for humans (see especially work on Baker's Paradox).

Many researchers have suspected that there are patterns in terms of which verbs can get used in which ways, explaining the structure of language and how language learning is possible, as well as shedding light on the structure of thought itself. For instance, the difference (it is argued) between Albert hit the vase and Albert hit at the vase is that the latter sentence means that Albert hit the vase ineffectively. You can't say Albert broke at the vase because you can't ineffectively break something: It is either broken or not. The reason you can't say Albert sent the border a book is that this construction means that the border owns the book, which a border can't do -- borders aren't people and can't own anything -- but a boarder can. The difference between Albert frightened Beatrice and Beatrice feared Albert is that the former describes an event that happened in a particular time and place (compare Albert frightened Beatrice yesterday in the kitchen with Beatrice feared Albert yesterday in the kitchen).

When researchers look at the aspects of meaning that matter for grammar across different languages, many of the same aspects pop up over and over again. Does the verb describe something changing (break vs. hit)? Does it describe something only people can do (own, know, believe vs. exist, break, roll)? Does it describe an event or a state (frighten vs. fear)? This is too suspicious of a pattern to be accidental. Researchers like Steven Pinker have argued that language cares about these aspects of meaning because these are basic distinctions our brain makes when we think and reason about the world (see Stuff of Thought). Thus, the structure of language gives us insight into the structure of thought.

The Question

The theory is very compelling and is exciting if true, but there are good reasons to be skeptical. The biggest one is that there simply isn't that much evidence one way or another. Although a few grammatical constructions have been studied in detail (in recent years, this work has been spearheaded by Ben Ambridge of the University of Liverpool), the vast majority have not been systematically studied, even in English. Although evidence so far suggests that which verbs go in which grammatical constructions is driven primarily or entirely by meaning, skeptics have argued that is because researchers so far have focused on exactly those parts of language that are systematic, and that if we looked at the whole picture, we would see that things are not so neat and tidy.

The problem is that no single researcher -- nor even an entire laboratory -- can possibly investigate the whole picture. Checking every verb in every grammatical construction (e.g., noun verb noun vs. noun verb at noun, etc.) for every aspect of meaning would take one person the rest of her life.

CrowdSourcing the Answer

Last May, VerbCorner was launched to solve this problem. For the first round of the project, we posted questions about 641 verbs and six different aspects of meaning. By October 18th, 1,513 volunteers had provided 117,584 judgments, which works out to 3-4 people per sentence per aspect of meaning. That was enough data to start analyzing.

As predicted, there is a great deal of systematicity in the relationship between meaning and grammar (for details on the analysis, see the next section). These results suggest that the relationship between grammar and meaning may indeed be very systematic, helping to explain how language is learnable at all. It also gives us some confidence in the broad project of using language as a window into how the brain thinks and reasons about the world. This is important, because the mind is not easy to study, and if we can leverage what we know about language, we will have learned a great deal. As we test more verbs and more aspects of meaning -- I recently added an additional aspect of meaning and several hundred new verbs -- that window will be come clearer and clearer.

Unless, of course, it turns out that not all of language is so systematic. While our data so far represent a significant proportion of all research to date, it's only a tiny fraction of English. That is what makes research on language so hard: there is so much of it, and it is incredibly complex. But with the support of our volunteer Citizen Scientists, I am confident that we will be able to finish the project and launch a new phase of the study of language.

That brings up one additional aspect of the results: It shows that this project is possible. Citizen Science is rare in the study of the mind, and many of my colleagues doubted that amateurs could provide reliable results. In fact, by the standard measures of reliability, the information our volunteers contributed is very reliable.

Of course, checking for a systematic relationship between grammar and meaning is only the first step. We'd also liked to understanding which verbs and grammatical constructions have which aspects of meaning and why, and leverage this knowledge into understanding more about the nature of thought. Right now, we still don't have enough data to have exciting new conclusions (for exciting old conclusions, see Pinker's Stuff of Thought). I expect I'll have more to say about that after we complete the next phase of data collection.

Details of the Analysis

Here is how we did the analyses. If meaning determines which grammatical constructions a given verb can appear in, then you would expect that all the verbs that appear in the same set of frames should be the same in terms of the core aspects of meaning discussed above. So if one of those verbs describes, for instance, physical contact, then all of them should.

Helpfully, the VerbNet project -- which was built on earlier work by Beth Levin -- has already classified over 6,000 English verbs according to which grammatical constructions they can appear in. The 641 verbs posted in the first round of the VerbCorner project consisted of all the verbs from 11 of these classes.

So is it the case that in a given class, all the verbs describe physical contact or all of them do not? One additional complication is that, as I described above, the grammatical construction itself can change the meaning. So what I did was count what percentage of verbs from the same class have the same value for a given aspect of meaning for each grammatical construction, and then I averaged over those constructions.

The "Explode on Contact" task in VerbCorner asked people to determine whether a given sentence (e.g., Albert hugged Beatrice) described contact between different people or things. Were the results for a given verb class and a given grammatical construction? Several volunteers checked each sentence. If there was disagreement among the volunteers, I used whatever answer the majority had chosen.

This graph shows the degree of consistency by verb class (the classes are numbered according to their VerbNet number), with 100% being maximum consistency. You can see that all eleven classes are very close to 100%. Obviously, exactly 100% would be more impressive, but that's extremely rare to see when working with human judgments, simply because people make mistakes. We addressed this in part by having several people check each sentence, but there are so many sentences (around 5,000), that simply by bad luck sometimes several people will all make a mistake on the same sentence. So this graph looks as close to 100% as one could reasonably expect. As we get more data, it should get clearer.

Results were similar for other tasks. Another one looked at whether the sentence described someone applying force (pushing, shoving, etc.) to something or someone else:
Maybe everything just looks very consistent? We actually had a check for that. One of the tasks measures whether the sentence describes something that is good, bad, or neither. These is no evidence that this aspect of meaning matters for grammar (again, the hypothesis is not that every aspect of meaning matters -- only certain ones that are particularly important for structuring thought are expected to matter). And, indeed, we see much less consistency:
Notice that there is still some consistency, however. This seems to be mostly because most sentences describe something that is neither good nor bad, so there is a fair amount of essentially accidental consistency within each verb class. Nonetheless, this is far less consistency that what we saw for the other five aspects of meaning studied.


Anonymous said...

Great article. This project sounds really interesting. I'm curious about what the boundaries might be for using the wealth of literature already digitised and available.
Is it that the language is too complex even in publications with orthodox language use (eg. newspapers or theses)? Is it too difficult to determine and control the context?
Or perhaps the research is being done.
Regards, Mike

GamesWithWords said...

Quite a bit can and is done by analyzing published texts. The difficulty is it is just language floating in the ether -- there's nothing to connect it to any actual meaning. Imagine I gave you a book in Swahili and told you to "learn Swahili." You have no idea what the book is about, and you know no Swahili. I suspect that no matter how much you analyzed that book, you wouldn't get very far in your knowledge of the language!

Martha Palmer said...

Long term we would very much like to hook this up with the exciting work going on in distributional semantics (trying to induce semantically related clusters of words from vast amounts of text, etc.). Automatic clustering of instances of verbs into classes seems like a good place to start, and that is something we are working on. Check out an upcoming EACL paper by Daisuke Kawahara on this topic.