Blondin-Massé, Alexandre; Harnad, Stevan; Picard, Olivier; and St-Louis, Bernard (2013) Symbol Grounding and the Origin of Language: From Show to Tell. In, Lefebvre, Claire; Cohen, Henri; and Comrie, Bernard (eds.) New Perspectives on the Origins of Language. Benjamin
Arbib, M. A. (2018). In support of the role of pantomime in language evolution. Journal of Language Evolution, 3(1), 41-44.
Vincent-Lamarre, Philippe., Blondin Massé, Alexandre, Lopes, Marcus, Lord, Mèlanie, Marcotte, Odile, & Harnad, Stevan (2016). The Latent Structure of Dictionaries. TopiCS in Cognitive Science 8(3) 625–659
Organisms’ adaptive success depends on being able to do the right thing with the right kind of thing. This is categorization. Most species can learn categories by direct experience (induction). Only human beings can acquire categories by word of mouth (instruction). Artificial-life simulations show the evolutionary advantage of instruction over induction, human electrophysiology experiments show that the two ways of acquiring categories still share some common features, and graph-theoretic analyses show that dictionaries consist of a core of more concrete words that are learned earlier, from direct experience, and the meanings of the rest of the dictionary can be learned from definition alone, by combining the core words into subject/predicate propositions with truth values. Language began when purposive miming became conventionalized into arbitrary sequences of shared category names describing and defining new categories via propositions.
Subscribe to:
Post Comments (Atom)
PSYC 538 Syllabus
Categorization, Communication and Consciousness 2021 Time : FRIDAYS 11:35-2:25 Place : ZOOM Instructors : Stevan Harnad & Fernanda Pere...
-
Categorization, Communication and Consciousness 2021 Time : FRIDAYS 11:35-2:25 Place : ZOOM Instructors : Stevan Harnad & Fernanda Pere...
-
Turing, A.M. (1950) Computing Machinery and Intelligence . Mind 49 433-460 I propose to consider the question, "Can machines think?...
-
Searle, John. R. (1980) Minds, brains, and programs . Behavioral and Brain Sciences 3 (3): 417-457 This article can be viewed as an atte...
In response to the results of the Cangelosi and Harnad study described in the reading, I was led to consider what the results would be if the instructive group had not previously grounded some important and relevant categories. In this case, I would think that the benefit seen in the instructive learning creatures would disappear. Thus, if a baseline number of important categories are not learned through direct sensorimotor interaction, would it be the case that we would expect the inductive creatures to have the survival advantage? Ultimately, the advantages that hearsay provides does not mean we can get away with being told everything.
ReplyDeleteWithout the capacity to learn categories by direct sensorimotor induction, none of the creatures could learn the categories at all. (This was a toy model, but one group could learn only by induction, the other by either induction or hearsay, and that second group was the one that won in the competition.) Read the Latent Structure paper: what's needed is not "important" categories, but what?
DeleteAfter reading 8b, I found the discussion about how language is learned by both "induction" (sensorimotor trial and error learning) and "instruction" ("hearsay", or propositions) very interesting. It seems that after we acquired enough words (around 10% of the dictionary) through induction, we learned the rest of them through instruction because it is the most evolutionary economic way -- the species who have the ability to learn by "hearsay" survived. To me, this argument seems to suggest:
ReplyDelete1. The "instruction" part of language acquisition capacity depends on mutation (which seems to support 8a's argument that language is evolved by natural selection).
2. Evolution is lazy because it does not code every word but gives us the learning capacity through "hearsay".
Furthermore, the experiment that chimpanzees are not able to learn our grammatical rules seemed to suggest we are the species with evolutionary-selected language capacity. However, I was wondering since language is also symbol manipulation, is it possible that other species are just adopting different language rules to communicate that we cannot learn as well (but humans are so egocentric to assume we are the 'only' species that passed the evolutionary selection)?
1. The smallest number of words in a dictionary that remain if we keep recursively removing every word that can already be defined out of the remaining words is called the Kernel of the dictionary. The Kernel is about 10-15% of the dictionary. Every dictionary has just one Kernel. But the Kernel is not the smallest number of words out of which all the other words in the dictionary can be defined.
Delete2. The smallest number of words out of which all the other words in the dictionary can be defined is the Minset (Minimal Grounding Set). The Minset) is between 750 and 1500 words, depending on the size of the dictionary. So it is much smaller than 10% of the dictionary. But each dictionary has many different Minsets. All are in the Kernel. All are the same minimal size.
3. In principle, you would only need to ground one Minset directly through direct sensorimotor learning to be able to learn all other words and potential categories (in the dictionary – plus all encylopedias, textbooks, as well as all the things that can be learned from direct sensorimotor experience) through verbal definitions and descriptions alone (“hearsay”), starting from just that one grounded Minset. But no one is suggesting that we just ground one Minset directly and then spend the rest of our lives in a dark room in front of a screen, in a lifelong T2 chat with others who have seen and been in the world. We keep on doing direct sensorimotor grounding throughout our lives. [Why?]
4. Yes, we evolved the genes that give us the capacity to learn categorize (“to do the right thing with what kind of thing”), to name our categories, and to combine those category names to produce and understand (“mirror”) subject/predicate propositions that define and describe further categories -- in other words we evolved the capacity for language. But it was not through one gene with an “Instruction” mutation. [What do you think the steps might have been to evolve, from the (remarkable and powerful) nonverbal learning capacities of other species, to human linguistic capacity?]
5. Nonhuman species have extremely powerful learning, social and communication capacities. And, yes, we are anthropocentric, assuming other species know, and can do, much, much less than they really can, and do. But language seems to be a special and unique form of communication: symbolic and propositional. As suggested by the Strong Church/Turing Thesis, we can not only computationally model just about anything, but we can also verbally describe just about anything – in any language (as long as we have the sensorimotor capacity to ground the language). And all languages are intertranslatable. So you can say anything in any language (though not necessarily with the same number of words). Hence if nonhuman species had comparable communication codes, they too would be translatable into human language, and we could discuss it with other species.
Hi! I had a thought on question 3. Is the fact that we continue to learn via sensorimotor experience after grounding our language's Minset (this assumption that the minset is grounded first is supported by the fact that the minset is encompassed by the kernel, which is learned at an early age) evidence that it is not always most efficient to ground through instruction? I am compelled by the principle that we only need to ground the minset via sensorimotor learning in order to learn the language and categories to some extent, but situations where understanding comes more naturally when you see something unfold in front of you, or engage with something via sensorimotor interaction seem to challenge this intuitive assumption. For example, if we were adapted to be able to ground everything from our understanding of the minset, why do we feel a deeper sense of understanding when we carry out a chemistry lab ourselves than when a teacher tells us about a chemistry lab?
DeleteThis is all just speculation, but there are advantages to both direct sensorimotor grounding and indirect verbal grounding. What are they?
DeleteIn response to 4, does this relate to the discussion about how certain individuals were motivated to engage in direct sensorimotor learning in the environment, then became motivated to share that information with their kin to aid survival/reproductive success. Thus, the motivation to increase survival/reproductive success of the self and kin helped lead to the development of human linguistic capacity?
DeleteYes, but what do you think the steps were that got our ancestors from the capacity for unsupervised and supervised category learning that we share with so many other species, to the capacity for language? No one knows the correct answer, but it is worth thinking about, because it really did happen.
DeleteIndirect verbal grounding is more advantageous for grounding abstract concepts, like "socialism" for example. In contrast, direct sensorimotor grounding is more advantageous for grounding concrete, tangible categories, such as "red".
DeleteIn response to the question: "We keep on doing direct sensorimotor grounding throughout our lives. [Why?]", the first direct reason is that, while during the evolution we gain the ability of language and indirect verbal grounding, we have never lost our ability to do sensorimotor grounding. So language is an enormous enhancement of fitness, but not a replacement of sensorimotor grounding. The second reason is that as language spread beyond kinship to a wider community, the within species competition increases and there could be conflict of interests. Then the could be inaccurate or even deliberately misleading for the instructor to gain more benefits, then we would need to verify the truth by ourselves through direct sensorimotor grounding.
DeleteFollowing the discussion on Prof. Harnad’s point 3, I think one of the reasons for continuing our sensorimotor learning and grounding is that new objects, technologies and multisensory experience(like immersive experience) are continued to develop and for an experience that can be learnt through actually sensorimotor interactions, this is the most accurate and direct way to do so. Hearsay is efficient to describe abstract things, although in-person experience can be more powerful. Especially in the covid situation, I think many people has realized the importance of direct sensorimotor interaction rather than written words, simulations or virtual interactions.
DeleteFurthermore, to prove this point further, infants are a actually only able to learn through other manner such as media/models or videos as long as there is some sort of interaction available or some sort of social contingency available to the child. Therefore, this data actually reinforced the statement made where in-person experience may actually be more powerful.
Delete“[Categorization] excludes only continuous skills, such as walking, swimming, basketball playing; we can categorize some of the discrete choice-points of their dynamic actions, but the skills themselves are continuous rather than categorical, and their dynamic “shapes” are certainly not arbitrary but congruent with the realworld objects and states on which they operate…” (p. 6 of 16)
ReplyDeleteThis quote brought to mind our discussion of discrete and continuous in distinguishing the real from the simulated. As the quote describes above, all categorization, which includes the behavioral processes surrounding language, is discrete. Simulations are discrete in that they can only estimate and selectively represent the continuous functions in our world in a limited manner. Language works likewise, representing objects, not as they actually are but in descriptive estimations. Language also generates internal images and structures that might be able to be thought of as the virtual world of language. I hadn’t considered it before, but is it possible to think of language as a sort of simulation of the real world?
Hi Genavieve!
DeleteYou are right, language is indeed a simulation of the real world:) We can consider language as a kind of algorithm with rules (syntax) and meanings that are semantically interpretable (semantics). Just as a computer simulation is an algorithm simulating the real world, language can be thought of as a comparable thing simulating the real world (just with a different type of algorithm).
Hi Genavieve and Xingti,
DeleteI had the same thought while reading this paper! A lot of the terms used in it reminded me of our discussions on computation, especially the rule-based, shape-based, and semantically interpretable aspects of computation. Since, like this paper argues, the shapes we use as symbols (like the squiggle that represents the number 2, for example) are entirely arbitrary, the only way we can understand what 2 + 2 = 4 represents is by assigning meaning to all of the symbols in that proposition (‘2’, ‘+’, ‘=’, and ‘4’). Like Xingti says, language can and does simulate the real world, and I would add that we understand language because of the meaning we have assigned to individual words through symbol grounding of sensorimotor experience. This could even explain why some people often visualize scenarios while reading: in their heads, they are creating simulations of the words and propositions they are reading. The reading talks about how “language itself began earlier than vocal language,” so we may have been simulating long before we could describe to others what simulation even is. Was that simply not necessary or evolutionarily advantageous at the time, or were we just not yet physically able to do so? I think this question ties very well into the issue of how “telling” has taken precedence over “showing” in evolution, since we learn better from word of mouth (“instruction”) than from direct experience (“induction”).
Computation and language are only analogous in some respects. (And computation is also a part of language.) They are both formal codes, but not the same kind of code. Language is grounded, computation is not.
DeleteThere are also two (related) senses of “simulation.” One is just “formally encoded,” whereas the other is “simulated for the senses,” as in VR (Virtual Reality) where computer code is used to control the output from goggles and gloves to your eye and hand. VR is not just computation; the goggles and gloves are dynamic physical devices, controlled by the computational code.
Language is just words (symbols). What distinguishes language from just computational code (squiggles) is words’ sensorimotor grounding: Directly grounded words (category names) are somehow connected to feature-filters that can pick out the sensorimotor features that distinguish members from non-members of the category to which the word refers. This (somehow) enables the brain to imagine the referent (visually, acoustically or tactually) when the category-name is spoken or thought. This is a bit like a VR simulation, somehow generated by the brain through the connection between the internal symbol (the category name) and its connection with its sensorimotor grounding. Something like this also seems to happen with categories that are grounded indirectly (i.e., verbally), like “zebra,” in which the image is generated by the direct grounding of “striped” and “horse” even if you’ve never seen a zebra.
Language begins with categorization (“doing… etc.”). Arbitrary “names” for categories come later. But before vocal naming comes bodily imitation (“mirroring”); that’s why “Stevan Says” language did not begin vocally but gesturally, through pantomime. Pantomime is not language. It is just imitation, including purposive communication, as in pointing, only much more specific. But the critical transition to language is not from gestural imitation to vocal communication; nor is it just the transition from iconic imitation to arbitrary gestures (gestural “squiggles”). The critical transition is from categorizing and arbitrary gestural “names” to subject/predicate propositions. [How/why do you think that might have happened?]
To answer "The critical transition is from categorizing and arbitrary gestural “names” to subject/predicate propositions. [How/why do you think that might have happened?]", as mentioned in 5.2, there is a survival advantage of freeing hands from gestural "names" to do something else, and when the gestural "name" are not available in sight. And since the categories names are arbitrary, it doesn't matter what sensory form they take, in our case and many other species' cases it is the readily available sound through auditory system and vocal system. There are other sensory forms too, for example, olfaction and pheromone which are very common ways of communicating in many species! (sorry for being off the topic) However, this is still not language, because more importantly, combining old "names" into new "names" to form propositions is the key to form languages.
DeleteA few times throughout this paper, the idea is mentioned that vocal language developed after language itself. I was a little bit unclear on what this “language itself” would be. Does this mean that language as an internal symbol system used to categorize the world developed before this system was expressed vocally to others? This is what I was thinking it meant based on the sentence, in connection to the mushroom foragers, "since we do not believe language started vocally anyway -- think of the category “names” not as vocalizations but as observable actions”, which I took to mean language initially may have been how we internally categorized what we observed. But then, Prof. Harnad’s response above makes me think language being prior to vocal language means something like language came in the form of non-verbal demonstrations first? So rather than there being some internal language developed before it was expressed verbally, the idea is just that non-verbal communication came before verbal communication?
ReplyDeleteI was thinking about how the origins of language itself being prior to the origins of vocal language would relate to what selection pressures would have motivated the evolution of “language itself”. At first it seems like being able to verbally communicate with others is the main reason to share a language structure (when reading the Pinker and Bloom paper I was definitely picturing a verbal language when the authors were discussing parity in communications protocols, but non-verbal communication would also share the benefits of verbal communication). I wonder, though, if even just having a similar internal categorization system could be useful enough to have been selected for. Say I observe Bob eating a mushroom. If Bob’s internal categorization structure that led him to identify this mushroom as edible is similar enough to my internal categorization structure, I will be more likely to accurately code this mushroom as edible through observing Bob’s actions (generally, Bob’s interactions with the world will make more sense to me if I am categorizing the world in a similar way to how Bob is when deciding which actions to take).
1. Language is not just speech (vocal language): Gestural languages (e.g., the sign languages of the nonhearing) are all real languages, with the full (nuclear) expressive power of language, and dictionaries, completely intertranslatable with vocal languages.
Delete2. According to the gestural theory of the origin of language, language started as gesture. One reason for that hypothesis is that imitation and pointing could first make the link with what was being referred to, and gestural imitation is much richer and broader than vocal imitation. Once an iconic (analog) gesture (and its natural [mirror] link to its analog referent) become arbitrary, a social convention, simplifying the imitative gestures users have been sharing, it “inherits” the natural grounding of the arbitrary symbol as it sheds its iconicity. But neither vocal nor gestural names is yet language (or even reference), any more than pointing is language. What’s needed is propositions.
3. Language is “linguistic” communication (“verbal” has a vocal bias), meaning propositional communication. Content words (i.e., words with referents) are category names, whether they are vocal or gestural. And the names are arbitrary symbols, whether they are vocal or gestural.
4. I think (i.e., “Stevan Says”) internal language began with external language (for communication), not before. But I think Chomsky would say thought either precedes internal language or it is the same thing as inner language (I-language), so it precedes E-language. I’m not sure exactly what Chomsky thinks about this or why. But trust him (a giant) not me (a pygmy)! In any case, this has nothing to do with whether the language is vocal or gestural.
5. Yes, categorization and category learning came long before language. But if you see someone who knows eating a mushroom, you might be able to conclude that it’s safe to eat that particular mushroom. But seeing him eat it won’t give you the features that distinguish edible mushrooms from inedible ones. (In the mushroom simulation, “edible” had to be learned the hard way, for all creatures. So did “markable.” The only thing that could be learned from “hearsay” (which in that nonvocal world was just the observation that creatures were doing “return” with mushrooms that were both “edible” and “markable,” both of whose features everyone had already learned from direct experience. So the only thing they needed was to observe that the “returnable” mushrooms were the ones that had the features of both the edible mushrooms and the markable mushrooms. (This was a bit of a cheat in an extremely small toy simulation.)
6. Yes, shared, grounded categories make communication easier, but not enough to allow you to guess the features of a new category (or the meaning of a new word) without a either a definition or some sensorimotor learning.
I found the idea of the “kernel” words in this reading to be interesting. From what I’ve understood, learners can pick up new categories through observation, or unsupervised learning. These new categories are always formed by a certain set of other categories/qualities/descriptions that always refer back to other words which have been learned either by induction or instruction. As the real-life simulation discussed in this reading has found, learning by instruction is much more efficient than learning by induction. That is to say, the lesser of your categories or words are grounded in your sensorimotor experience, as induction learning would have it, the more fit you would be to survive. Collecting information about the world more efficiently means less mistakes like eating poison mushrooms (not good for survival!!). It would be interesting to see if there is a connection between this efficiency of instruction, and the limit discussed later in the text, of how many words are actually being used in the dictionary in order to define all the other words: this is the kernel. To what extent can one reduce the number of categories learned by instruction: where does efficiency meet the proposed limit, the kernel? Would this look something like an optimisation function? Are these related at all?
ReplyDelete1. Observation (unsupervised learning) is not instruction. (See above about how the toy simulation was a bit of a cheat, because the feature of “returnables” was just the features of “edibles” plus the features of “markables.”) Instruction is verbal: telling the learner what the distinguishing features of the new category are. For that the features have to already be grounded and named. And it requires propositions.
Delete2. Verbal learning is much faster, easier, safer and more powerful than sensorimotor learning (unsupervised or supervised). But sensorimotor learning is essential too, not just to ground feature-names but because verbal definitions are approximate. “A picture (object) is worth a thousand words (of description).” The definition of a “pterodactyl” could be extended and extended, to include more and more features and facts, but seeing a pterodactyl always adds other features that were left out in the grounded verbal approximation. (Yet even the features learned through direct supervised learning are still approximate: why?)
3. Language increases fitness, but not because fewer sensorimotor categories is better. Why?
4. The relatively tiny size of a MinSet says something about the power of verbal definition and description, but it does not imply that the smaller the MinSet the better. Definitions will be shorter the more categories a language has “lexicalized” (which means: to coin a word to name the category instead of always using a long verbal description). This is also related to “chunking” (q.v.). (The size of the MinSet seems to depend partly on the total number of words in a dictionary. And a language’s lexicalized vocabulary is always growing.)
5. The MinSet is not the same as the Kernel: What’s the difference?
6. But just as having a language does not mean you only need to ground one MinSet and then for the rest of your life you can learn everything else by texting from a dark room, it is also not the goal to minimize learning from verbal instruction! (What made you think that it was?)
2- real things are continuous in features, while there is a fundamental limit of categorization (it is discrete) . Even with an unlimited number of words, you would always be approximating the features of the object because you can't recreate something continuous with something that isn't
Delete3- is it because it made it easier to share knowledge and cooperate.
5- The kernel is the list of words left when you remove all words with which you can no longer can make new words with (repeat until you can't remove any words). It's approximately 10-15% of the dictionary. The MinSet is a list of word with which you can describe all other words in the dictionary (and it is inside the kernel).
6- Is it because words are just approximations, and basing other words on previous approximation makes them even less accurate; grounding more Minset makes your language system hold more information which makes it more useful (more accurate approximation) -Elyass
2- the limit is more than just discreteness/continuity (read about Watanabe's Ugly Duckling Theorem in Week 6).
Delete3- re-read the advantages of verbal instruction over trial and error induction.
4- The Kernel (there is only one) is the smallest sub-dictionary inside the full dictionary from which you can reach all the rest by definition. The MinSet is the smallest set of words in the Kernel from which you can define all the other words (there are many MinSets, same minimal size, but none of them is a dictionary: why?).
6- Yes, verbal descriptions of a category's distinguishing features are approximate, but sensorimotor feature-detectors are approximate too. And it's easier revise the verbal description by adding more verbal descriptors than to update sensorimotor feature-detectors by trial and error. Nor is it clear that categories or their features "fade" as they become more combinatory and abstract. But definition can certainly be brought closer to the ground if it is accompanied by an image.
The MinSets contained within the kernel only define things outside of itself, and not inside — the MinSet is the smallest set of words to define all other words in the dictionary, and thus, if (like in a dictionary) one of the MinSet words could be defined by another MinSet word, the set would not be as small as it can be. For instance (and this is not a great example because these words can be broken down further and would thus not be in the MinSet…), the word “red apple” could not be in the MinSet, because “red” and “fruit” (assuming they are in the MinSet) can define “red apple”. Adding “red apple” only increases the set unnecessarily, because whatever word could be defined by “red apple” in the rest of the dictionary could be defined first by “red” and “fruit” instead. This is what makes the MinSet not a dictionary.
DeleteAccording to the reading, primitive forms of communication that lead to language occurred when we realized that we could acquire knowledge and categories from others doing the right thing with the right kind of thing without needing to interact with it directly ourselves. Language has probably evolved from pointing and miming. Pointing appears to be spontaneous to humans, it doesn’t appear that other very intelligent animals (such as great apes) use pointing. Furthermore, the paper describes that we had the necessary “cognitive components” before language, but our motivation and sociability enabled us to use these capacities for language. What does it imply for the evolution of language, from then to now? Were these necessary ‘cognitive components’ related to universal grammar? Nonetheless, other animals are very smart, social and communicate in some way, why do they not have language?
ReplyDeleteLanguage has a nuclear power. This is because it is able to produce an infinite amount of categories and propositions with truth values. As categorization (doing the right thing with the right kind of thing) is very important, and learning is faster and simpler though instruction rather than induction, language is therefore the most powerful way to learn categories. Nonetheless, we still need to learn certain categories through induction (sensorimotor learning), but how much induction is actually necessary?
Lola, your explanation and analysis is quite interesting. The note you made about pointing is also something I wish to expand on. Following your analysis, I wonder if pointing is a human trait due to our grounding of words and thus, the ability to directly link the word with the physical object — ex. Pointing to an object when the question is “what is x” is a correct and appropriate response.
DeleteFurthermore, I believe that your point on ‘why do other smart and social animals not have language’ — is in development. If we look at the case of “Bunny the talking dog” it is interesting to contemplate if she has some sort of understanding of the use of human language. She is able to ask questions, express her emotions, and use propositions with truth values (such as “dad upstairs” or “sound outside” then going upstairs or outside to show/confirm her statement). It appears Bunny is using mostly the instructive method of learning, such as in the case of her owner pushing the ‘ball’ button and Bunny getting different toys until she gets the ball, is rewarded and associates the ball with the ‘ball’ button. However, there must be some sort of indirect grounding or inductive learning taking place, because Bunny would not be able to generate her own strings of words and propositions if she were to be using instructive methods since the beginning of her learning.
Please read my other Replies. UG is a special case, for several reasons. (What are they? and what is UG?) But there's much more to language than UG. It is unlikely (but not impossible) that UG has anything to do with the question of whether the reason only humans can learn and use language is cognitive or motivational.
DeleteWhile reading "The Latent Structure of Dictionaries", I made a connection to what I learned about language acquisition in a class on developmental psychology. Namely how the three layers of a dictionary (Core, Kernel, Dictionary) are analogous to a baby's psycholinguistics as they learn.
ReplyDeleteParents talk to their infants using infant directed speech, which uses a mix of words the infant understands and words they don't understand. Infant directed speech is useful for teaching language because it uses the child's zone of proximal development, which is a lot like the structure of a dictionary. There is a core of known words, an outside layer of mostly unknown words and words that bridge between the two like the kernel.
Using infant directed speech is a lot like "instruction" for language development because you are exposing the child to language they do not know yet couched in language they understand, which gives them a sense for what the new words mean.
Is there a database of the words in infant-directed speech? Are there more of them in the Kernel than the rest of the dictionary? In the Satellites? In the Core? In the MinSets? How do they fit in the Dictionary (vegan) Doughnut?
DeleteI thought this was a really interesting way of interpreting the concepts of the dictionary to how language is learned in infancy. I think there is a database of the word in infant-directed speech, though obviously the database varies per infant, these words and sentences tend to be simpler. I would hypothesize that most of them are in the kernel than in the rest of the dictionary and many of them in the MinSets. When speaking to babies and toddlers, adults are more prone to speak in concrete terms like “good” and “no”. Though it also really also depends on how parents speak to their babies, as well as what babies may overhear that is not directed toward them. This opens up more exposure to the rest of the dictionary. To Laurier’s point, language acquisition builds overtime, and new words and meanings are bridged from previously learned words and meanings.
DeleteSounds like a potential research project (the locus of the words of child-directed speech in the hidden structure of the dictionary).
DeleteHi Laurier, I haven't taken the class you're referring to, but I did study childhood development a bit in the past, and it's really interesting how you linked this week's topics to children's zone of proximal development. I wonder if there is any link between how much overlap there is between MinSets and infant-directed speech and an individual's language development. For instance, if a parent uses more grounded language/more words overlapping with a MinSet, does a child have a greater ability to learn words via instruction?
DeleteTwo thoughts I had while reading this paper:
ReplyDelete1) "So maybe that’s what evolution did with our species. Because we were more social, more cooperative and collaborative, more kin-dependent—and not necessarily because we were that much smarter—some of us discovered the power of acquiring categories by instruction instead of just induction, first passively, by chance, without the help of any genetic predisposition." This passage reminded me of a book I read for another class, "Mothers and others" by the anthropologist Sarah Hrdy. In this book, Hrdy makes the hypothesis that the reason why humans evolved to have this *desire* to communicate and cooperate with others (unlike apes who can, but don't show interest towards it expect when it benefits them) is because of the fact that homo sapiens raise their children communally. Unlike most great apes who are very protective of their child, human mother in hunter-gatherer society need the help of other members of the group ("allomothers") to care for the child, and that means that kids that are better at communicating and requesting the care of others will have an evolutive edge.
2) The difference between learning through induction vs instruction, and the statement that you cannot learn everything through instruction, made me think of the notion of qualia. Could things that are supposedly impossible to describe, like colors or the feeling of pain, be an example of categories that can only be learned through induction? Of course, you could learn by instruction that touching this plant will hurt, so you would be able to categorize that plant. But knowing what pain feels like must be useful for categorization in general.
1. Sarah Hrdy's point about humans applies just as much to elephants. See all the "allomothers" in the Jabulani Elephant Herd. And, while you're at it, does Bunny the Dog have language? Why? or Why not?
Delete2. Of course the only way we can learn what it feels like to feel (feel anything: touching, tasting, hearing, seeing, pain) is by being able to feel. And that immediately leads to the "hard problem" of how and why we feel, rather than just "do" (including detecting sensory input and producing motor output, without feeling a thing, like a toy robot, or even a T3, if Eric could be a T3 zombot). We're not going to get an answer to that in this course, just a better understanding of the question.
But just as we can ask how many words have to be grounded directly thought sensorimotor learning so that all other words can be grounded out of them through combinatory definition, we can ask how many sensory modalities need to be grounded directly so that the rest can be understood from words. (Think of Helen Keller, who was congenitally blind and deaf, yet could manage with her remaining senses. -- And don't we all have a good enough approximate sense of what it feels like to "see" with the bat's sonar? (But this does not mean that a color-blind person's sense of "purple" isn't a lot vaguer than ours...)
I would think that our ability to develop language goes beyond a desire to communicate because this same desire should also appear among the “allomothers” in the elephant herd and in Bunny. In both these examples, there is one party, either the baby elephants or Bunny, that rely on another group to be taken care of. It would be desirable then to be able to communicate exact needs, like when you are hurt or hungry.
DeleteThat being said, I have already assumed that Bunny does not have language. I believe this is the case because I do not believe that she would be able to use her current vocabulary to make meaning of new words, or in other words, learn language instructively. I think Bunny only knows the words she knows because of her direct experience. For example, she knows to press the “ouch” button when she is hurt, but if she were introduced to the word “pain” I don’t believe she would know what this meant and how it is associated with “ouch” until she was hurt again.
In addition to language being grounded and computation not being grounded (as mentioned by Prof Harnad above), I would like to add other aspects of natural language (which was discussed during class and section 3.4 of the above reading) that shows language cannot be purely computational.
ReplyDeleteComputation is semantically independent (semantically interpretable yes but still independent). The concept of “autonomy of syntax” in language is parallel to such property of computation. The sentence “Colorless green ideas sleep furiously” is semantically nonsensical but still grammatically correct which demonstrate that syntax and semantics are distinct. Syntax is independent of semantics.
HOWEVER, although the sentence is semantically nonsensical, it is NOT semantically EMPTY. We can still somewhat make sense of the sentence metaphorically which was well illustrated through the poem by Hollander recited during class. I believe we can make sense of such “nonsensical” sentences because those words (symbols) are already semantically grounded in our minds (the interpreters).
(I’m not sure I have grasped the concept 100%, please correct me.)
Hence, we CAN say that syntax (of natural language) is NOT independent of meaning but we CANNOT hastily say that indicates [syntax IS dependent on semantics]. Similarly, we also CANNOT say that [semantics IS dependent on syntax] because we CAN still somewhat make sense of sentences that does not obey UG like “Furiously sleep ideas green colorless” or “John is easy to please Mary”. All we can say is [syntax is NOT independent of meaning].
Furthermore, the article states: “What is certainly true is that, when we manipulate natural language symbols (in other words: words) in order to say all the things we mean to say, our symbol manipulations are governed by a further-constraint, over and above the syntax that determines whether or not they are well-formed: that further constraint is the “shape” of the words’ meanings” (page 6). Thus, language is not purely computational in that respect because our manipulation of words is determined by 1) well-formed syntax AND 2) the “shape” of the words’ meanings.
Intuitively thinking, this makes sense because language, which is a form of communication, serves to convey some type of meaningful message. What we say, therefore is not only governed by the formal syntactic rules (grammar) but also meaning. While we can still make sense of nonsensical sentences or UG-violating sentences because of our ability to ground and have grounded meaning, we don’t TYPICALLY PRODUCE nonsensical sentences or UG-violating sentences because of both #1 and #2 respectively.
(Does this make sense? Have I understood this correctly?)
X is independent of Y means: X is not dependent on Y and Y is not dependent on X.
Delete“Furiously sleep ideas green colorless” violates OG.
The shape of words’ meanings is the shape of their grounding (the sensorimotor feature-detectors, plus the syntax and logic of propositions).
We do produce nonsensical (or at least self-contradictory) sentences (just read what anti-vaxxers say). We can also speak in nonsense syllables (e.g., as in Lewis Carrol’s Jabberwocky), though that still cheats with some OG inflexions like -ed, -ing, and function words like “the” and “a,” so it’s both UG- and OG-compliant.
Only hypothesis-testing Chomskian linguists produce UG-violating sentences. Just trying to violate UG without knowing what UG is is unlikely to produce a violation of UG. (It’s a bit like trying to make a sound you can’t pronounce.)
I'm still confused. Does that mean the syntax of natural langue IS or IS NOT independent of semantics?
DeleteFrom the class, I got the message that syntax is NOT independent of semantics but I don't think I fully understand WHY. It is because the words are grounded?
In natural language syntax is not independent of semantics. But the dependency (constraint) is in both directions, not just in one or the other (as you seemed to be saying).
DeleteIn this skywriting, I wanted to draw parallels between how languages are characterized in the text and the definition of computation that was presented earlier in the course. In this text, natural languages are described as being symbol systems with arbitrary shapes that follow syntactic rules and have systematic meaning. In addition, natural languages are implementation-independent, as they can be taught to (or implemented by) different individuals and technologies. This follows almost word-per-word what we saw during the first week of class. Keeping in mind how computations come to be and how useful they can be, I wonder if we could adopt a more computational approach to language in order to guide our understanding of how it came to be. Maybe language emerged from the brain’s abilities to perform other computations? It would be interesting to think about what kinds of inferences could be drawn on the origins and evolution of language from a computational point of view.
ReplyDeleteBut the symbols of formal computation, though, like words, arbitrary, have no referent or meaning, whereas the symbols of natural language, are grounded: T3 (Eric) can point to, categorize and describe the referents of his (content) words. The word-shape is arbitrary, but the shape of its referent is not. Neither is the (noncomputational) machinery inside T3 that connects the word to its referent in the world.
DeleteHere I will be attempting to explain the Baldwinian effect. In this computational model, organisms learn the first few categories via trial-and-error and the following categories via trial-and-error or hearsay. Hearsay is another term for learning by instruction. Over time, organisms only learned through hearsay rather than trial-and-error to the point that the instruction learners (the ones who learned through language) had out-survived and out-produced the induction learners. Indeed, the power of language, if categorization is important, and if learning to do it in the fastest, safest and simplest way is important, then language is the most powerful way to learn (most) categories. In fact, this is exactly the reason why the hearsay organisms survived, because of the Baldwinian effect: the organisms who are more inclined or motivated to learn categorization (doing the right thing with the right kind of thing) through the most efficient way had an advantage in the form of survival and reproductive success over the rest of the group.
ReplyDeleteThat is correct. In other words, individuals that have a better predisposition to learn will be better at adapting to their environment and thus will be the ones to pass their genes more successfully (natural selection). That is how species end up having a predisposition to learn adaptive behaviours that once had to be learned from scratch.
DeleteIn this case, the behaviour is learning categories through language (verbal instruction): the survival and reproductive success of individuals that are better at this eventually result in a biological predisposition for language.
I have a slightly long-winded question:
ReplyDeleteA symbol system like math is independent of both hardware and meaning. People can interpret these otherwise meaningless systems so that they say something. An example of this would be a math equation. However, despite our interpretations, a computational algorithm or logical proof doesn’t need meaning to be executed. In other words, a human being doing a math problem or a computer executing a program doesn’t use meaning to do the steps. Indeed, formal symbols manipulation is being executed according to a series of rules. This brings me to my question; When it comes to language, symbols are grounded and as such are meaningful. So, c language can be “done” formally like ungrounded symbols?
This reminds me of the Chinese room argument. When we treat language as purely computational (as in the CRA), one can learn the rules of the symbol system and how to do the right thing with the right kind of thing but without any grounded meaning. I think the answer to your question is that computational language can be done formally like ungrounded symbols but they too will be ungrounded and therefore (I think) kind of useless for the purposes of human communication. A T2 robot can learn language computationally but according to Searle it will never "understand".
DeleteI agree with Lucie. I think generally we use computational languages to perform some sort of action or we have some goal in mind. Although these languages might not directly link to our spoken language, I think they would still require meaning in order to act according to our intentions (as in we would place meaning onto the symbols in a certain way). Just by creating the language with the intention for it to do something I think would prevent computational languages from being done formally.
DeleteThe article Symbol Grounding and the problem of language discusses the finding that when induction-only learners were placed in competition with instruction learners, within a few generations the instruction learners had both out-survived and reproduce the induction learners. For me this raised questions about why instruction learners were able to survive for much longer – what exactly about instruction learning is significantly more sustainable than induction learning? Shouldn’t first-hand sensorimotor experiences with the categories they’re learning confer some advantages such as being able to learn the categories more efficiently since their conceptions/understandings can actually be grounded through experience?
ReplyDeleteLearning categories through sensorimotor experience is important because it allows us to ground symbols to their referents in the world. But it is time-consuming, and somehow dangerous. Take the example of mushrooms: To learn which ones are edible and which ones are poisonous, you'd have to risk getting sick if you only go through trial and error. Having someone else verbally describe the features that distinguish these two types will save you a lot of stomachache and reduce your risk of dying after having mistakenly eating a poisonous mushroom!
DeleteAnother example: Inuit cultures have different categories for "ice you can walk on" or "ice you can make drinkable water of". The advantage of sharing the features of each one of this through language saves the trouble of having to walk on the ice to see if it cracks under your feet or to drink water that may be unhealthy to be able to learn these categories through trial and error.
Remember that we also mentioned that we keep learning through sensorimotor interaction through life, because acquiring all the categories through language from a minimal set can end up being very time consuming as well. So both sensorimotor (trial and error) learning and verbal instruction learning have its role in adult cognition.
The advantage of language is that it saves us the time and trouble to interact with every possible category out there in order to learn it.
I am going to try to summarize my understanding of this article with an example situation. If a child were to be brought up in a small room, and was not able to fully interact with the environment and was not taught language, they would still have Universal Grammar. However, because they were not able to expand upon this by using sensorimotor capacities to ground words. As well, they were never taught language or about objects through instruction. Therefore, they would not have the same capacity and understanding for language as children who were taught language through instruction, and used sensorimotor capabilities to ground symbols/words. However, if you removed the child from the room and attempted to teach them a language, it is likely that they would be able to grasp basic concepts of language and how to structure it, because of the principle of Universal Grammar, which is inborn and therefore doesn’t need to be taught.
ReplyDeleteThe example of mushroom category learning through instruction and induction in this essay was quite illuminating. From experience alone, learning by instruction does tend to be more beneficial than learning by induction, since in the case of instruction, you can better avoid making the same mistakes as your instructor. However, based on my own experience learning things, when you learn something through induction, it tends to stick in your head more. For example, if I’m learning how to distinguish between different kinds of mushroom, I would likely remember which one is toxic better if I got sick from it myself than if someone told me it makes people sick. This kind of direct experience just tends to get more deeply ingrained due to the unpleasant nature of poisoning. However, I also recognize that learning through direct experience is certainly riskier, and I’d rather learn that a mushroom is lethal by instruction than by induction—especially since I may die before even getting the chance to put my newfound knowledge into practice. So from an evolutionary standpoint, it’s easy to see why everyone having to learn all categories through induction would not be beneficial.
ReplyDeleteHi Milo. I agree with your observation that when you learn something through induction, it tends to stick in your head more. I think the last section of this paper sheds some light to this observation. Although we gain a huge benefit from learning categories through instruction, the categories we learned through instructions do need to be refined through sensorimotor induction. This point can be seen more generally in our everyday learning process. It is important that we learn from the instructions of teachers. But it is equally important to our learning that we practice them after class and apply them in real life. Hence, I think the most efficient way of learning categories should be a combination of instruction and induction.
DeleteI want to clarify my thinking after this week’s lecture: Miming itself is not a way to create or distinguish between categories. Rather, miming is a way to nonverbally communicate basics such as what is happening or what somebody wants. Miming may also be an important part of how nonlinguistic communication transitioned into linguistic communication through propositions. This miming is very different from sign language in that it sign language is just as complex as spoken language and miming does not involve categorization, propositions, syntax, etc. Is that correct?
ReplyDelete"For those readers who have doubts, and think there is something one can say in another
ReplyDeletelanguage that one cannot say in English: please tell us —in English—what that something is, and
why one cannot say it. You will find that you have generated your own counterexample."
I believe that this argument is not valid, as it is just as possible that there is a term in one language that cannot be translated properly to English, or another language - at least not with the same accuracy, and thus, the translation, even using more words, cannot be stretched enough to be exactly what is meant by the foreign term. Thus, perhaps a natural language cannot express any and every proposition, as languages can be limited by the flexibility, vocabulary, and precision of the terms used.
That being said, I do believe the general idea and message of the argument - that languages can be translated to some degree and carry the same meaning in another language - but there can be some limitations based on the limitations of the languages themselves, if that makes any sense.
That’s a good point, Alex. Being from a small island whose endemic language (Mauritian Creole) is only spoken by the local people, I can definitely relate. There have been countless times when I was thinking of an expression in that language, but could not find a proper translation for it in English. Even if I did manage to translate it, the translation did not quite grasp the feeling of what the expression was trying to convey.
DeleteThis makes me think of Searle’s Chinese Room Argument. It feels like something for me to understand Mauritian Creole, because I grew up speaking it. Whatever symbol grounding I did growing up was in that language. Thus, I think that if someone who has never spoken the language just uses Google translate to reply to whatever I say in my native language (it’s not even on the list of languages that Google can translate, but hopefully one day), they won’t fully understand it, because they have not grounded at least the Kernel through sensorimotor interaction or indirect verbal grounding.
Hi Shandra, thank you for the reply and the example. Being monolingual, this idea was only something I have heard from other bilingual or multilingual speakers, rather than something I could experience for myself.
DeleteI believe this idea of limitations between languages is related to the main question of the last Blondin Massé et al. reading. "The size of the smallest set of words that
can define all the rest – a graph’s “minimum feedback vertex set” or MinSet – is about 1% of
the dictionary"(Blondin Massé et al., 2016). This idea of a subset of words that can be used to describe all other words would be different in other languages - the percentage, the quantity - and I believe, symbol grounding would be different in each language. Some languages do not have words for things that other languages do, and this backs up my argument in my above skywriting.
I agree with you, Alex. I think the minimal set of words to ground a language can be very different from language to language.
DeleteThere are many words(categories) in a language, which simply cannot be found in another language. And we cannot simply describe these words in another language, either. Some of them are very fundamental categories, which are directly grounded, so you cannot really describe them by words. Some of them can be described by words within their own language. But many of words (concepts/categories) used in the description do not have correspondences in another language, so in this case, it still cannot be translated.
‘If UG is central to language and is inborn rather than learned, then that would turn a big part of the question of the origin of language into the question of the origin of UG.’
ReplyDeleteMy kid sib asked me almost the same question after I introduced what UG is and proposed to consider if UG is the thing that is innate for humans, which somehow gives humans the nuclear power to generate language ability. In the following section, Prof.Harnad argues that language is a symbol system in which one can express any and every proposition, and since every proposition is also a category inclusion statement, the power of language could be considered as the power of acquiring a new composite category form other people who already knew the old categories out of which the new one was composed.
After reading this article, my question is where and when UG is working when we are acquiring a new language. Take Bunny as an example. After I reviewed more videos of Bunny pressing the ‘keyboard’ to ‘say’ something, such as ‘you we I family here,’ I don’t think Bunny violates UG; instead, Bunny really sounds like a baby learning a new language, which is violating the OG. I wonder if it is because whether those strings of words made by Bunny or the simple gestures learned by chimps convey proposition or still just pantomime is still a mystery, we cannot make the conclusion that UG is unique for humans and it might lead to a dead-end if we stick to discuss whether UG is a necessary condition for the human to evolve with language ability.
When the paper proposed the challenge to think of something one can say in another language that does not translate to English, I thought of words related to collectivist cultures. I’ll attempt to define them to the best of my ability but there will be pieces that are lacking that I cannot account for due to the missing link between English and Korean. In Korean there are two words that come to mind; “han” and “jeong”. “Han” can be described under the umbrella term of “oppression”. However, it is uniquely related to Korea due to the fact that it is a collective, societal feeling that is associated with the Japanese invasion. The second word “jeong” refers to psychosocial bonds that tie Korean society together. This stems from Korea’s history of oppression. Is the fact that I can describe the word in such a way and indicate what is lacking in an English translation enough to label it as a counterexample? Or does the fact that the word stem from a unique history and its definition is tied to the history distinguish it as an example of a word that does not have an English translation?
ReplyDeleteIt’s a weird comparison, but for me, how the kernel is to dictionaries is like how UG is to OG (although UG is innate, and the kernel is learned). More specifically, kernel words are the core of dictionaries, that cannot be deducted more to reach all the definitions of other words, and UG is the basis of language, and all the other rules (OG) are the derivative of UG that will never violate it. Hence, applying this to T3 robots, if we provide them with (or in other words, build...in them) all the fundamentals such as UG, innate CP, etc., they should be capable to learn the rest as humans.
ReplyDeleteWhat is language? It could be defined broadly as any attempt to communicate using the tools at our disposal (brain, socializing, miming etc.). It could be more simply that language is Universal grammar (UG). UG being rules that are already present at birth, not learned through experience by children. Related to communication, we can say that language is related to our ability to categorize, to do the right thing with the right kind of thing. We can use arbitrary shapes named symbols and manipulate them with rules. Then, we can associate an object in the world to those symbols by fundamentally experiencing the object with our senses and naming it(induction). Then, we could instruct others on what we have learned(instruction). Lastly, an interesting point brought in the paper is that experiencing life through our sense is still vital to Homo Sapiens and makes our ability to categorize sharper. Thus, experiencing with our senses will always have an upper hand on instruction through hearsay.
ReplyDeleteIn this paper by Blondin-Masse et al., the authors hypothesize that “cognitive components … were already available before language began” and natural selection selected for the ones that helped individuals communicatie and learn categories the most efficiently in order to increase survival. This broadly mirrors the selectionist approach in our last reading by Pinker and Bloom.
ReplyDeleteIn section 5.1, the authors also pose the question whether disposition to propose is intelligence or motivation. They are unsure as to whether or not chimps get propositions, not due to lack of intelligence, but instead due to motivation or compulsion to name and describe. I’m not entirely sure whether I agree or disagree with this statement. If we take an evolutionary standpoint, what diverted humans to have such an increased sense of motivation beyond chimps that allowed us to pick up propositions? As I read on, Blondin-Masse et. al did touch upon the fact that there wasn’t that big of a motivation and perhaps humans just acquired propositions first and proceeded to alter the environment which ultimately inhibited chimps from acquisition of propositions. I’m curious as to whether or not there are counter-arguments or advancements made from this specific point.
After reading this interesting paper, I found myself not fully convinced by the definition of language in it. This paper gives two definitions of language. Katz’s version is the “glossability” that a “natural language is a symbol system in which one can express any and every proposition, where a proposition is any statement that has a truth value. The version given by authors of this paper is “translatability” thesis: “Anything you can say in any natural language can also be said in any other natural language”.
ReplyDeleteFor “glossability”, my doubt is that whether it is too narrow a scope to limit language to just propositions? It seems to me that language is much more than mere propositions. Not only do we still keep more “primitive” use of language to point to things, and to pantomime, but we also use language in all sorts of pragmatical purposes, such as making commands, assenting, refusing and warning. We also use language for creative purposes such as writing poetry. None of them necessarily have a truth value.
For the “translatability” thesis, my observation is that there are certain words and expressions in a language which cannot be translated to another language. This observation has been discussed in some sky-writing responses above. I just want to add that one comment: In think grounding is important not only for the minimal grounding set. Words which can be fully defined by other words are still connected to the sensory-motor capacities. Maybe you can teach me what an apple is through instruction, but our understanding of the category apple is still quite different, depending our “experience” of the apple as we sensory-motorly interact with the apple Hence, some untranslatability between languages may be caused by the very different grounding of words resulted from the very different forms of life in various cultures.