10.3: Language Processing and Parsing

10.3.1 Language Processing and Parsing, from Sarah Harmon

Video Script

10.3.1.1 Language Processing and Parsing

Let's go a little more into language processing and parsing. This is what we know so far, what we have data to show, and not just hypotheses but actual brain scans. We’re able to start parsing out how a human processes language. This is also a very nascent area of linguistics; we have had theories on how language is parsed for a number of years, really for decades. However, we're only at the starting point of understanding how the brain works; the technology is only starting to allow us of view into how this is happening. A lot of what is presented here, hopefully, will either be supported in the future, or shown to be different; we will see.

Let's start with the specific processing of language. There are two views of this: top-down and bottom-up, and that has everything to do with the input versus the output. When we're talking about processing, we're talking about when you hear language and you understand it—competence, in this case. Bottom-up processing would be that you get your input from your ear (or if you're signing, from your eyes), then you process the small to the big. You start with the sounds and the phonemes, then you go to the puzzle pieces and the morphemes. From there, you start processing how they combine in that phrase, and then the semantic or pragmatic interpretations. That's one way to view processing. The other view would be top-down processing, which would mean that you start with the overall meaning—concepts, context, pragmatic aspects, and the phrasal aspects—and then break it down to the morphology, then phonology and phonemes.

Both of those theories of processing are highly abstract, and we actually have a little bit of evidence for both. First, think about if you are in a really noisy environment and you're trying to have a conversation with somebody. You cannot hear everything that they say, and yet you follow along; you get the gist of what they're saying. That's an example of top-down processing; you're getting the overall context and then you're filling in the blanks, as it were, for the pieces that you did not hear. There have been a number of studies that showcase that if you get folks trying to pronounce isolated words, and then have them use those words in a sentence, they will make more mistakes if they pronounce the word in isolation. Interestingly enough, if you create a gibberish sentence—a sentence that fails compositionality and/or grammaticality, it could be either one—that there are errors in different ways. That's an example of both bottom-up and top-down parsing; bottom-up because you're focusing on that lexicon in isolation and therefore, you're focusing on the phonemes and then going up from there. Top-down because in a sentence you're able to use the context clues to fill in the gaps.

Speaking of gaps filling in gaps, I think of this kind of like Mad Libs—the game where one person has a story or a paragraph with certain words out, and they ask their friends or colleagues or audience for a specific kind of word a noun, a color, a verb, an action, and then you fill in the gaps. Filling-in-the-gap experiment is pretty much the same thing; you leave out certain key information and the audience, the person you're talking to or writing to, fills in the gap and frequently will use the context clues to fill in the gap. Here's the interesting piece: that's mostly a top-down parsing, but there's also a bottom-up parsing, if they're given a partial clue. Let's say they were given a sentence and a key verb was missing. If they only heard one sound initial sound of that verb, they might fill in the gap with the appropriate term, thereby building up from the bottom, from the sound, and building up to the context. We see a number of cases of both bottom-up and top-down processing. There isn't one way that we do this; it's both.

The final example of this has to do with segmentation; I’ll give you a great example. You see in IPA this term here: [grede]. Unless you know the context, that could be two possible meanings. It could be grade A as an something was given the grade of A, so there's a syllable break, as it were, between the [d] and the [e]. Or, it can be a grey day, so the syllable break would be between the first [e] and the [d]. Unless you have the context, you don't know. This is where arbitrariness really comes into play; I keep bringing this term up throughout the course, and I told you it was the most important of all the hallmarks of human language. This is an example; we cannot parse something unless we understand the context. This is just English, but this is true for any language, where you have any level of homophony.

With respect to processing, we also have to factor in anticipation and priming; you'll see a little bit more about priming in the next section that Catherine Anderson includes, but she mixes up a few things. That's why I’m not going to do it here. Anticipation, especially with respect to phonology, is a signal that you are fluent in a language. If you hearken back to the very last section of language acquisition, the previous chapter, we talked about what fluency really meant and that the average person's definition of fluency is lower, and does not satisfy the definition that a linguist or a foreign language professional would count as fluency in a language. Anticipation is one aspect; if I’m fluent in a language, I can anticipate based off of what the person is saying. The sound and the context will help me parse out everything else. If I’m not fluent in that language, I have a harder time parsing that information and anticipating what will happen. The same is true with priming; that's focused more semantic and morphologic input, where certain cues will lead you to think about other things. If I’m talking about a picnic that I’m organizing, and I’m trying to figure out who should bring what, and I say, “Well, first of all, I kind of would like to have a hot dog; somebody have hot dogs at the at the picnic.” That might lead other people to say, “Well, we should bring burns. Maybe we could do hamburgers or other meats or meat substitutes. Oh, we need condiments like mustard, ketchup, mayo and pickles.” One seed, as it were, grows into all sorts of connections, and that is what priming does.

What is really cool is to see how categorical language is. The video below talks about the brain dictionary, and it shows the hypothesis that we have been working with for some time about a mental lexicon. It seems that we do actually see in different parts of the brain categories of lexicon, and they seem to pop up in different areas of the brain.

10.3.1.2 Failure to Parse

If we talk about parsing and processing, we also have to talk about what happens when we don't parse correctly, when there's something going on. If we talk about arbitrariness, we also have to talk about ambiguity. We talked a little bit about ambiguity with respect and syntax and semantics. There's more to it; there are two specific principles—The Principle of Minimal Attachment and The Principle of Late closure—that factor heavily into ambiguity. The Principle of Minimal Attachment is that you build the simplest structure consistent with the grammar of the language. If something doesn't make sense, it could be in violation of the Principle of Minimal Attachment, because you got too complicated; you need to keep it simple and straightforward. Don't try and attach extra pieces. The Principle of Late Closure says that we attach incoming material to a phrase that is currently being processed, we anticipate and prime ourselves; if what we're priming and anticipating doesn't match what happens, there's ambiguity. Both of these are factors when we say we don't understand something, and you can also include the Principle of Compositionality, which we talked about in the Meaning chapter. They all work together; if there's ambiguity, something is breaking down. Either the structure doesn't follow what we expect—that's Minimal Attachment—or the semantics and the content doesn't match what we expect—that's Late Closure.

What we can do sometimes is backtrack and use a garden path, as it were, to backtrack and figure out what the person was intending to say. This, especially happens with respect to presupposition and implicature. If we didn't understand the context that the speaker was intending, then we'll backtrack or use the garden path, as it were, to go back to where we were to try and piece it together. You can see, this all the time in a conversation. If you have a conversation with somebody, and you aren't quite picking up what they're saying, or you're not quite understanding it, it means that you're not able to parse it. It's ambiguous. You usually have some kind of way of stopping the conversation and go back to the part that you last understood. It could be as simple as saying, “Wait just a second, let me go back,” or some version of that phrasing. Or you can say, “Hold up; say that again,” and that's another way to cue your speaker that they're not making sense. This is something called dischordia. We signal dischordia to explain that scenario when the person you’re talking/signing to says something that you are able to parse. This definitely can lead to re-analysis. By the way, if you think about historical linguistics, when we talked about language change and specifically re-analysis, that an apron really started off in life as a napron, so there was a question as to priming and parsing, probably a Late Closure error. You can also see it in acquisition of lexicon from other languages, when we borrow something and then add or subtract from it. We see this in a variety of ways, and this all has to do with parsing.

We can even talk about bilingual parsing, when we are talking about aspects of language that are not native to us, and the use of context clues. We talked about how if I said [si] to a Spanish-English bilingual, depending on the context depends on whether it is a body of water or an affirmative comment, and that has everything to do with context clues. It also ties into knowing your audience, whether you know the person that you're talking to speaks the language that you're intending to use. Implicature and presupposition is also part of this. We build as we learn a language, whether it is our native language or one we have acquired, whether it is our first language or our 10th language we constantly build all of these pieces in such that parsing becomes a high level of processing language.

Of course, there are errors in parsing. When we have slips of the tongue—I meant to say bridge and I said bork—or word substitutions—I can't think of sofa, so I’m going to say couch, but for whatever reason that semi-synonym may not work for the context. There are blending errors, which is when instead of combining lexicon that work together, like smog from smoke and fog, I say skog instead and there's an error in the blending. Spoonerisms are when you swap out full syllables and swap them, and then create a metastasis that may create a whole different word. Metaphor, when it's not shared meaning, and the other person doesn't understand the context, all of these are errors in parsing, in comprehension.

What is also important to remember is that every single human being innately and subconsciously shows that they have an error in parsing and comprehension, and it can be followed by a conscious thought and even statement. When we talk about animal communication, that is something that we don't see much of; we don't see errors in parsing. You don't see a cat telling another cat, “Hey, I didn't understand you. Would you say it again?” We don't (so far) know that they do that.