Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Words and Meanings Words and meanings, are they the same things? Onomatopoeia A type of word that sounds like the thing it is describing. English examples: beep, wow, buzz, pop, bang, mumble, quack Chinese examples: 噓、哈啾、咕嚕、滴答滴答、叮噹叮噹、批瀝啪啦 English examples: fl- cluster is often (but of course not always) associated with rapid movement: flame, flap, flare, flash, flicker, flurry, flutter But… Cross-language variation in onomatopoeia DOG English bow wow French gnaf gnaf Albanian ham ham Japanese wan wan Chinese wang wang Indonesian gong gong Shakespeare said… “A rose by any other name would smell as sweet.” Words and meanings are related but separate entities… Translation argument --meanings for which there are no single words. eg., Yiddish schlep “to move a heavy and usually bulky item from place to place” --a word for which we don’t know the meaning. eg., adagio Imperfect mapping argument --one word can have more than one meaning (ambiguity) eg., bug, ring --one meaning can have more than one word (synonymy) eg., sofa/couch, pail/bucket Elasticity argument: --a word can have different meanings depending on the context in which it occurs. eg., “strong” coffee, “light” class load Questions: How do words relate to meanings? Is it a one-to-one relationship? How are words organized and accessed in our minds? How are meanings represented in our minds? How do children learn the mappings of words and meanings? How are two languages (or more) represented in bilingual minds? The Study of Words Questions: What is the smallest form in which a word is stored in our mental lexicon? What are the most important factors that influence lexical access and organization? Word primitives “Words” (or lexemes) are primitives. -- eg., book, books, bookish, bookshelf -- save processing time but take up more memory space “Morphemes” are primitives. --eg., book, -s, -ish, shelf --same memory space but take more processing time Research method Lexical decision task table vanue daughter tasp coref hunter Reaction time: How much time it takes to decide whether a given words is a real word or not. Evidence supporting “morpheme” as word primitive It takes longer to process multimorphemic words than words composed of a single morpheme. (eg., indecision, deciding) Pseudo-multimorphemic words take longer time to recognize than real prefixed words or unprefixed words. (eg., result, interest vs. recall, interchange vs. table) Speech errors from normal healthy people. (eg., she washed uped the dishes; the dog bited the cat.) Aphasic speech The evidence suggests that… People store morphologically complex words as individual morphemes, which are combined in speech production, or stripped in comprehension. But note… Inflectional morphemes and derivational morphemes may be processed in different ways. Derivational morphemes are more firmly attached to root morphemes Inflectional morphemes are more likely to be added on as we speak. Frequently occurring multimorphemic words (disagree, impossible) often have separate lexical entries from their root morphemes. What about compound words? Gingerbread, butterfly? Whether compound words are processed as one single lexical entry or multiple lexical entries depend on how transparent the meanings of composite morphemes are. Semantically transparent: teaspoon, buttonhole Semantically opaque: butterfly, mushroom How do we know? Semantic priming task pea primes beanpole bread does not prime butterfly How are words organized and accessed in our mental lexicon? Factors influencing word access and organization Frequency (lexical decision task, naming task) Imageability and concreteness and abstractness (recall task) Semantics (meanings) (word association, category naming, semantic priming, aphasia) Grammatical class (word association, speech errors, frequency effect for open-class words but not for close-class words, Broca’s aphasia) Phonology (tip of tongue, speech errors) What did we learn in the previous class? What’s the relationship between words and meanings? What questions do we ask about the mental lexicon? What is the primitive form of word that is represented in our mental lexicon? How do we know? How are words organized in our mental lexicon? How do we know? Models of lexical access Serial Search Models Words are searched one at a time. Parallel Access (or Direct Access) Models --Perceptual input about a word can activate a lexical item directly --Multiple lexical entries are activated in parallel Forster’s autonomous search model Three access files: orthographic, phonological, semantic/syntactic The orthographic and phonological access files mostly contain information about the beginning parts of words—the first few letters of their spelling or first few sounds with which they begin. When an access file directs the search to the appropriate lexical bin, entries are searched one by one until an exact match to the perceptual representation is found. When the relevant lexical entry is retrieved, it is checked against the input in a post-access check. Evidence supporting the model: It takes more time to reject non-words than to accept real words. Logogen Model (Morton) Words are accessed by being activated to a certain threshold. Each word has its own “logogen” which tabulating the number of features that a lexical entry shares with a perceptual stimulus. Each logogen has its individual threshold, which is the amount of “energy” that will be needed to access that lexical entry. Any logogen is accessed when the total activation reaches a pre-designated threshold. Many types of information are used to access the target word. If several entries are activated to threshold, the one with the highest count wins and is “recognized”. It then slowly returns to its resting level. This model can account for frequency and priming effects. There is counter-evidence for the model. Connectionist Models (McClelland & Rumelhart) The models use the analogy of the brain and neurons to develop models of cognition. In the models, lexical access is instituted in “neural nets” composed of nodes and connections between these nodes. Nodes are of three types: input, output, and hidden nodes. Each functional level of the hidden nodes represents different aspects of words—eg., their visual, orthographic, phonological and semantic natures, etc. Connections between layers, and between nodes in the same layer, can be either excitatory or inhibitory. The models also emphasize direct access to lexical entries, simultaneous activation of multiple candidates, and the use of many types of information to access a target word. The models can account for frequency and priming effects. Cohort Model (Marslen-Wilson) The model was designed to account only for auditory word recognition. When we hear a word, all of its phonological neighbors get activated. This set of words is known as the “word initial cohort”. The list of word candidates is narrowed as the auditory input proceeds serially. Initially, the cohort model depended heavily on an exact match between a spoken word and its phonological representation in the lexicon. But now the model is revised: the system chooses the best match to fit an incoming word. The model can account for frequency and priming effects. Which model (serial or parallel processing) accounts for word access and organization better? Parallel Processing Model The Study of Meanings When we know the meaning of a word, what do we know? Bachelor Unmarried male Is Pope a bachelor? Eligible, unmarried male A divorced man? Despite having a lifelong relationship with Simone de Beauvoir, Jean-Paul Sartre died a bachelor? What does “meaning” mean? Philosophical theories of meaning Reference Theory --The meaning of a word is the object to which that word refers to in the real world (that is, its referent) --Problems: it doesn’t account for words that do not name things (such as function words), nor does it account for abstract terms such as “freedom” or unreal objects such as “unicorns” Ideational Theory --Words denote ideas rather than things. --Problems?? Alternative Theories --Words have no meaning independently but are based on their connection to other words and sentences within the language. Conceptual primitives What are the smallest units of meaning? Are concepts defined in a rule-like fashion, with clear conceptual boundaries? Is it sufficient to represent a concept as a list of features? Theories of conceptual structure Feature Theories Knowledge-based Approaches Feature Theories The smallest unit of meaning is feature. Features can be perceptual (gray, large, tall), functional (used to transport people), microstructural (composed of hydrogen and oxygen molecules), or societal/conventional (supreme ruler). The Classical View The Family Resemblance View Define “triangle” Is a closed figure Has three sides Has angles that add up to 180 degrees. The Classical View Any concept has necessary and jointly sufficient features that all instances of that concept share. Eg., “triangles” All triangles are considered to be equally good triangles. Concepts are defined in a rule-like fashion, with clear boundaries. Define “bird” Can fly Lays eggs Has feathers Builds nests Is small Has bones The Family Resemblance View For any concept, there are no necessary and sufficient features, but there are characteristic features — attributes common to many exemplars of a category. Instances of a concept may overlap in some traits but not in others, and they do not share a single set of features. Some instances of a category or concept are more representative than others. Categories are said to have a graded structure. The best example of a concept or category is known as the prototype, which has the most characteristic features for the category. People use the prototype of a concept as a reference point to make category judgements. Fuzzy boundaries— some instances of one category can overlap significantly with other categories An interesting experiment showing that many categories have a graded structure… Armstrong, Gleitman, and Gleitman (1983) Task: subjects had to rate category exemplars on a 1 (best) to 7 (worst) scale Results: -- apple (1.3), olive (6.5) -- 13 was a better odd number than was 57! More task: semantic verification task Results: -- subjects were quicker to respond to prototypical items of a category than to peripheral items. -- subjects were quicker to respond to “13 is an odd number” than to “57 is an odd number”! -- subjects were quicker to respond to “A mother is a female” than to “A waitress is a female.” ! Conclusions from the study… Categories have a graded structure. People may represent even concepts with necessary and jointly sufficient features in a graded way. Differences between the classical view and family resemblance view The classical view: --Conceptual boundaries are well defined. --It is sufficient to represent a concept as a list of features. The family resemblance view: --Conceptual boundaries are not well defined (fuzzy boundaries) --Concepts are graded as to typicality within a category. --A list of the most characteristic features of a category is sufficient to represent the meaning of the concept. Criticism of the Feature Theories Features themselves are not well defined. The choice and weighting of features is context and task dependent. We know much more about the intension of a concept than a list of features suggests. What does the following things have in common? children, jewelry, portable TV set, photograph album, manuscripts, oil paintings Things to take out of the home during a fire. Knowledge-based Approaches Concepts must be represented and organized according to people’s theories about the world. Psychological Essentialism (Medin) People believe that things have underlying essences that make them what they are. Psychological Contextualism Items can be categorized by the context in which they are found. Psychological Contextualism What do women, fire, dangerous have in common? In the Dyirbal language spoken in Australia, the three words are preceded by the same marker balan Knowledge-based Approaches What are the smallest units of meaning? Concepts have relevant features, but these features do not define the category. Are concepts defined in a rule-like fashion, with clear conceptual boundaries? Concepts are not defined in a rule-like way, with clear conceptual boundaries. Rather, concepts are constructed as needed for different contexts or goals. Is it sufficient to represent a concept as a list of features? No. Deeper knowledge about conceptual coherence is necessary to make accurate and adult-level categorizations. How are concepts organized in our mental lexicon? How do we know? Semantic verification task “A robin is a bird.” “A robin is an animal.” “An ostrich has feathers.” “An ostrich has skin.” Semantic priming task Models of Semantic Representation Hierarchical Network Model Feature Comparison Model Spreading Activation Network Model Hierarchical Network Model By Collins and Quillian (1969) Individual concepts are represented by nodes that are organized in our minds like pyramids. The features would only be stored at the higher-level concept. Problems with the model: -- may only work for taxonomic categories, but not for more abstract concepts such as virtue, good --can’t explain frequency effects of features. --can’t explain typicality effects Feature Comparison Model By Smith, Shoben, Rips (1974) Concepts are represented as lists defining features and characteristic features. All features are assumed to be stored under all relevant concepts. Semantic verification tasks are performed by comparing the number of overlapping features of two or more concepts. Problems with the model: --features are not well defined. --some features correlate more highly than others. Spreading Activation Network Model By Collins, Collins, Quillian (1969, 1970) Concepts are represented as nodes connected to related nodes. When a single concept is activated, the “electricity” spreads to connected concepts. Closely connected concepts are activated more quickly and more strongly than distantly related concepts. What about ambiguous words? How do people process ambiguous words in context? “The men started to drill before they were ordered to do so.” “The carpenter ordered the men to drill before they were ready.” Selective Access View Exhaustive Access View Multiple meanings of an ambiguous word are activated even when the word is embedded in a biasing context. Swinney (1979): cross-model priming study Task: subjects listened to sentences containing ambiguous words and simultaneously participated in a visual lexical decision task. Example: “The man was surprised when he found several spiders, roaches and other bugs in the corner of his room.” On the computer screen, subjects saw “ant” “spy” and “sew” Dominance is a major factor in early lexical access; dominant meanings are immediately activated even when the context is biased against them. Duffy, Morris, Rayner (1988) Explored the difference between balanced (right) and polarized (yarn) homographs. Method: eye tracking Stimuli: --biasing context: “The dragon was no longer in pain once the scale was removed.” --nonbiasing context: “Once the scale was removed, the dragon was no longer in pain.” Results: --eye gaze was longer to the balanced ambiguous words than to control words. --When context was available, eye gaze to the target word was longer only in the polarized condition. Time issue in processing ambiguous words Simpson and Burgess (1985) Stimuli: prime-target pairs (bank-money, bank-river) Results: --16 ms. ISI: only dominant meanings were facilitated. --300 ms. ISI: both meanings were equally activated. --longer ISIs: the dominant meaning appears to remain active while the less-frequent meaning subsides. Conclusions about processing ambiguous words Multiple meanings of a word may be activated in parallel, with the dominant meaning occurring first. Context may speed access to several meanings, but it doesn’t restrict access to all interpretations. Lexical ambiguity resolution is a dynamic process with various interpretations racing against each other for access based on frequency of meaning and contextual biases. How do children acquire the mappings of words and meanings? Reference Principle: The assumption that words map onto a thing or an event. Whole Object Principle: Words refer to whole objects, rather than to their parts. Extendibility Principle: The assumption that a word refers to a class of objects rather than a specific object. Mutual Exclusivity Principle: Each object can have only one name. Principle of Contrast: No two words have exactly the same meaning.