6. Beyond Similarity copy.docx

CRITICAL READING: CORNELL NOTES Beyond Similarity Name: Date: 30 August 2023 Section: Lecture 6 Period: Questions/Main Ideas/Vocabulary Notes/Answers/Definitions/Examples/Sentences Limitations of the Paradigm The experiments we have looked at so far have told us a lot about the way h...

CRITICAL READING: CORNELL NOTES Beyond Similarity Name: Date: 30 August 2023 Section: Lecture 6 Period: Questions/Main Ideas/Vocabulary Notes/Answers/Definitions/Examples/Sentences Limitations of the Paradigm The experiments we have looked at so far have told us a lot about the way humans make inferences from premises to conclusions. Importantly, they further demonstrate that category structure is graded, and that similarity relations underlie this structure. However, there are a number of limitations to the paradigm that should be pointed out. It Isn’t Always Clear What the Underlying Similarity Ratings Used in the Model Are Based on Heit & Rubenstein (1994) constructed arguments such as: Given that tuna/rabbits have blood with 2 – 3% potassium, how likely is it the whales will have blood with 2 – 3% potassium? Given that tuna/rabbits eat a large amount of food at one time, how likely is it the whales will eat a large amount of food at one time? They found that when the predicate was biological (blood potassium levels), induction was stronger for biologically similar animals (rabbits and whales). Similar results have been found in developmental studies of concept development (person has a spleen/likes to sing – mechanical monkey vs. worm). In other words, similarity is context dependent, and this dependency may not be captured in empirical data such as pairwise similarity ratings or featural overlap. There Are Situations in Which Similarity Doesn’t Appear to Play a Strong Role in Induction Lin & Murphy (2001) found that in some situations, induction was stronger for thematically related categories than for similar categories. Cats host the e. spinola bacteria strain. Therefore, lions/kitty-litter also host the e. spinola bacteria strain. Toothbrushes host the e. spinola bacteria strain. Therefore, teeth/hairbrushes also host the e. spinola bacteria strain. Cats are more similar to lions than kitty-litter, and toothbrushes are more similar to hairbrushes than teeth. But induction was stronger for the low similarity pairs (cat – kitty-litter and toothbrushes – teeth). The Predicates Aren’t Entirely Blank In other words, we may be relying on deeper knowledge to make out decision, not just the similarity structure of the category. It is possible to replace the properties with phrases such as “has property X”. But real-world properties often contain meaningfully structured information. Given that we want to understand inductive reasoning in the real-world, it places researchers in a difficult position regarding internal and external validity. Knowledge Effects Clearly, there is something going on that goes beyond simple similarity relations. The reversal of the assumption of premise monotonicity in the Heusen et. al experiments can be explained in regards to higher-order knowledge effects. Knowing that there is a difference between Brahms, Shostakovich and ACDC, changes the likelihood of extending from the premise to the conclusion. Voorspoels et. al In a similar experiment involving negative evidence, Voorspoels et. al demonstrate that the likelihood of extending from the premise to the conclusion depends upon where the participants believe they’re receiving the information from. If they believe it is coming from a ‘helpful’ or ‘knowledgeable’ source, monotonicity is violated. If they believe the information is simply being randomly sampled from the environment, monotonicity holds. Similar results have been found in studies investigating inductive generalisation. Participants are presented with evidence of something like observations of bandicoot foraging patterns and then asked to make generalisations from this evidence to other time points. The extent to which participants extend the generalisation, appears to be related to the amount and structure of the observations. Participants also varied in the degree of extension based upon whether or not they believed the information was coming from an expert or was being randomly sampled from the environment (but this varied across individuals). Examples of Knowledge Effects If the only information we have is related to size, and the new stimulus is halfway between the size of A and B, people are equally likely to categorise it as either an A or a B. But if we add more information: People tend to want to categorise it as a pizza, because we have higher-order knowledge about the relative likelihood of coins or pizzas being 5 inches in diameter. Lin & Murphy (1997) This is a tuk: Half the participants were told it’s for hunting. Noose for catching the animal. Hand guard. Handle. Rope for tightening the noose. Half of the participants were told it’s for fertilising plants: Loop for hanging up in storage. Tank for storing liquid fertiliser. Knob for tuning on/off. Outlet pipe for liquid. Participants were shown examples of objects with one or more features missing and asked ‘is this a tuk?’. Their willingness to accept an object as a tuk depended on: The condition they were in (hunting/fertilizing). The feature that was missing. Participants in the hunting condition tended to say this was a tuk but participants in the fertilizer condition didn’t. Why? It would appear that the ‘noose’ is a critical feature for hunting, but it is missing the ‘tank’, which is a critical feature for fertilising. Participants in the fertiliser condition tended to say this was a tuk, but participants in the hunting condition didn’t. Why? It is missing the ‘noose’ (critical for hunting), but it has the ‘tank’ which is critical for fertilising. The new objects were of equal similarity to the original – they were both missing a single feature. Participants drew on higher-order knowledge about which parts were functionally critical based upon their original descriptions. In other words, their categorisation decisions weren’t simply based upon similarity. Category Learning The features used in most lab-based category learning experiments are fairly arbitrary (colour, size, shape, etc). The features associated with real-world categories tend to have meaningful associations (has wings, feathers, roosts in trees, etc). It appears that our knowledge of the way these features interact, influences our ability to learn category structures. Participants were asked to learn categories that were structured like this, in which there was some sort of integration amongst the features. Or this, in which there were an equal number of features but there was no meaningful integration amongst the features. Participants learned the integrated categories in around half the time it took to learn the neutral categories. Again, higher-order knowledge helps us to integrate the features into a ‘jeep’ or ‘snowcat’ which is easier to encode than ‘arbitrary vehicle 1’ and ‘arbitrary vehicle 2’. Word Associations An alternative approach to measuring semantic structure is via word associations. Generation frequency/associative frequency is one way of doing this. De Deyne et. al (2018) ‘Small World of Words’ study has association norms for 12,000 cue words. Subsequent studies have collected more, and in various different languages. Do a good job of capturing similarity relationships, but also show deeper associations between items. Can show ‘chained’ associations that go beyond category memberships. The Word Association Game You are presented with a word (the cue), and you have to respond with the first word that comes to mind. Playing the game feels effortless, automatic and often entertaining. Generating a word associate is easy and indeed, responding with a word that isn’t the first thing that comes to mind turns out to be quite difficult.