Network Science in Cognitive Psychology

Michael S. Vitevitch, Michael S. Vitevitch

• Publisher: Routledge
• Year: 2019
• ISBN:
  9781000740943
• Edition: 1
• Topic: Psychologie
• Subtopic: Kognitive Psychologie & Kognition

This volume provides an integrative review of the emerging and increasing use of network science techniques in cognitive psychology, first developed in mathematics, computer science, sociology, and physics. The first resource on network science for cognitive psychologists in a growing international market, Vitevitch and a team of expert contributors provide a comprehensive and accessible overview of this cutting-edge topic.

This innovative guide draws on the three traditional pillars of cognitive psychological research–experimental, computational, and neuroscientific–and incorporates the latest findings from neuroimaging. The network perspective is applied to the fundamental domains of cognitive psychology including memory, language, problem-solving, and learning, as well as creativity and human intelligence, highlighting the insights to be gained through applying network science to a wide range of approaches and topics in cognitive psychology

Network Science in Cognitive Psychology will be essential reading for all upper-level cognitive psychology students, psychological researchers interested in using network science in their work, and network scientists interested in investigating questions related to cognition. It will also be useful for early career researchers and students in methodology and related courses.

1
THE INFLUENCE OF SOCIAL NETWORK PROPERTIES ON LANGUAGE PROCESSING AND USE

Shiri Lev-Ari

Language is a social phenomenon. We learn, process, and use it in social contexts. In other words, our social environment shapes our linguistic knowledge and use. To a degree, this is trivial. A child exposed to Japanese will become fluent in Japanese, whereas a child exposed to only Spanish will not understand Japanese but will master the sounds, vocabulary, and grammar of Spanish. It is by tracking the patterns in our environment that we learn language. Therefore, individual differences in our social environment could influence the input we receive and, consequently, what we know about language, how we use it, and how we learn from our input. In this chapter, I will first review briefly past research on the role of statistical learning in language acquisition. Then I will take up the case of individual differences in social network properties. I will show that social network properties influence the distributional properties of our input, and, consequently, the product of our statistical learning from it. I will show that social network properties thus influence our linguistic knowledge, how well we understand other speakers, and how malleable we are to the influence of our interlocutors.

Statistical Learning

Language is a structured system. Sounds and words do not occur randomly but are characterized by regularities. Learners are sensitive to these regularities and exploit them when learning language. For example, the sounds of our language are not spread uniformly over the entire sound space but fall into categories, with the sounds of each category forming a normal distribution around its peak. This distributional property provides the learner with valuable information about the number of categories that need to be learned and what they are. After all, languages differ in the distinctions that they make, and so a unimodal versus a bimodal distribution along a certain variable is informative regarding whether the language makes a distinction along this variable. Unsurprisingly, learners are sensitive to the shape of the distribution of their input and use it to infer the number of categories. Thus, infants exposed to a unimodal distribution along the Voice Onset Time (VOT) continuum do not distinguish sounds on the two ends of the distribution, whereas infants exposed to a bimodal distribution do (Maye, Werker & Gerken, 2002). Another way by which statistical learning helps with cracking the linguistic code is by extracting co-occurrence information. When we speak, we do not clearly separate words. This becomes evident when we listen to a language we do not understand. In those cases, we often find it difficult to tell when one word ends and another one begins. Sounds, however, do not appear in a random order but are governed by phonotactic rules and language-specific probabilities. Thus, certain sounds are more likely to follow one another within words than across words. Infants are sensitive to these transitional probabilities and use them to segment words (Saffran, Aslin & Newport, 1996). This sensitivity is important for isolating words and therefore for lexical acquisition. Of course, whole words also do not follow each other in a random order either but follow grammatical constraints. For example, adjectives precede rather than follow nouns in English. Learners are sensitive to these transitional properties as well and can extract them to learn the grammar of a language (Thompson & Newport, 2007).

Statistical learning is also an integral aspect of meaning extraction. To understand what a word means, we need to know what it refers to. Although this might seem simple in theory, in practice, there are many potential referents in every situation. Therefore, one mechanism for understanding what one is referring to, that is, to acquire word-meaning mappings, is cross-situational learning. Although there are many potential referents for every word in every situation, there are fewer potential referents that co-occur with the word across many situations. Thus, if the word ball occurs in a context of a puppy playing with a ball, we might mistake it as referring to the puppy. But if it also occurs in another context in which the puppy is absent but the ball is present together with other toys, we would be able to infer that the word refers to the ball as it is the only object that appears in both contexts. Indeed, infants seem to engage such learning strategies to learn word meanings (e.g., Akhtar & Montague, 1999; Smith & Yu, 2008). When it comes to learning verbs, the task is harder as the referents are more abstract. The acquisition of a verb meaning can be achieved, though, by learning the probabilities of subcategorization frames that appear with the word (Gleitman, 1990). If a verb like see is followed by a noun phrase (e.g., See the doggy!), then its meaning is perceptual, but when it is followed by a sentence (e.g., Let’s see who’s calling), its meaning is more abstract and closer to determine. Similarly, whether a verb appears in a transitive or an intransitive frame can provide cues to its meaning. By tracking the probabilities of certain syntactic frames then, learners can acquire word meaning, a process known as syntactic bootstrapping (Gleitman, 1990).

One aspect of the input that is argued to be important for statistical learning is its variability. In general, variability is argued to boost learning. Infants are more successful at learning the VOT distinction between /b/ and /p/ if exposed to 18 speakers instead of one, even when the amount of input they receive is the same (Rost & McMurray, 2009, 2010). Japanese speakers succeed better at learning the /l/—/r/ distinction in English when exposed to productions from five speakers rather than one (Lively, Logan & Pisoni, 1993). Similarly, English speakers are better at understanding a novel Chinese-accented speaker if they were previously exposed to four different Chinese-accented speakers than if they were previously exposed to only one speaker, again, even when the amount of input across conditions is matched (Bradlow & Bent, 2008). These effects are argued to be due to the greater variability that there is in input that comes from many compared with few speakers. To directly test the role of input variability, Rost and McMurray (2010) manipulated variability along different dimensions and showed that in phonological acquisition in infants, it is the variability along the irrelevant dimensions (e.g., prosody and pitch) that enables learners to understand which aspects of the stimuli are stable across contexts and important for categorization and which ones are not. In contrast, Sumner (2011) found that at least when it comes to adults, variability along the relevant dimension (VOT) facilitates adaptation to a new speaker. Variability has also been argued to boost learning by preventing the learner from relying on item-specific learning, and forcing the need to generalize across items, as there are too many specific items to memorize (Gómez, 2002). Variability, then, facilitates learning, and it might do so in several different ways.

Up until now, I reviewed literature that examines the role of statistical learning in language acquisition. It might therefore seem like the importance of statistical learning is restricted to infancy or the acquisition of a new language. This is not the case. Language learning is a process that continues throughout our lives. That is, it is not the case that our representations are set after a certain age. Throughout our lives we sample the input in our environment and update our representations in accordance with it. Adult native speakers exposed to /t/s with atypical VOTs shift their boundary between /d/ and /t/ in accordance with the new input (e.g., Kraljic, Brennan & Samuel, 2008; Norris, McQueen & Cutler, 2003). Our linguistic representations in our native language can even be influenced by exposure to a new language (e.g., Cook, 2003; Flege, 1987, 1995; Major, 1992; Sancier & Fowler, 1997). Thus, a speaker of Portuguese immersed in an English environment might start producing in Portuguese VOTs with a longer lag than is characteristic in Portuguese, reflecting an influence of English, which is characterized by longer VOTs (Sancier & Fowler, 1997). Even without atypical exposure, languages change, and so speakers exposed to changing distributions gradually change their own representations. For example, an examination of Queen Elisabeth II’s Christmas addresses indicates that her pronunciation, just as the pronunciation of many other British speakers, has changed with the years, leading even the queen to diverge from the Queen’s English (Harrington, Palethorpe & Watson, 2000).

To conclude, there is vast literature that shows that throughout our lives, we rely on statistical learning to extract patterns from the linguistic input in our environment and that the environment thus shapes our linguistic knowledge and abilities. In the remainder of this chapter I will show how individual differences in our social environment, by influencing the distributional properties of our input, lead to individual differences in our linguistic performance and in the robustness and malleability of our representations.

Social Network Size and Linguistic Abilities

People differ in the sizes of their social networks. Some people tend to interact with only a few people, whereas others might interact with a wide range of people. This is reflected in people’s holiday greeting habits: some people might send cards to only a few people, whereas other would send greeting cards to more than 350 people (Hill & Dunbar, 2003). Our careers, hobbies, and habits might also lead us to interact with quite different numbers of people. Imagine a doctor and a nurse, who spend their days interacting with a changing roster of patients of all walks of lives, and in contrast, a programmer who works from home. Presumably, the doctor and the programmer receive radically different linguistic input. Would they consequently differ in their linguistic abilities?

As a first stab into this question, I asked native Dutch speakers to keep a log of all their conversations for a period of a typical week. This log provided information about the number of different interlocutors these people interact with on a weekly basis and the number of hours of conversation they conduct per week. As a second step, all participants came to the lab to perform a few phonological and cognitive tasks, including a test of perception of speech in noise. In that task, participants heard Dutch pseudowords, that is, words that do not exist but follow the phonotactic rules of Dutch. These words were embedded in white noise. The participants’ task was to write down what they heard, and their responses were scored for accuracy in vowel identification. Results indicated that as predicted, participants who regularly interact with more people were better at understanding speech in noise (Lev-Ari, 2018a). This in itself is encouraging and in line with the hypothesis that having a larger social network boosts linguistic performance even in adults’ native language.

One may wonder though whether it is simply the case that people with larger social networks have superior cognitive abilities, and it is these abilities that lead to better performance. To rule out this account, all participants also performed a host of cognitive tasks that measure working memory (O-Span; Unsworth, Heitz, Schrock & Engle, 2005), auditory short-term memory, selective attention (flanker task; Eriksen, 1995), and task switching (trail making task; Reitan, 1958). Results indicated that none of these tasks correlated with network size, indicating that people with larger social networks do not have superior cognitive abilities in general. Moreover, even after controlling for all these tasks, social network size predicted accuracy in the perception of speech in noise task. The benefit of a larger social network, then, is not due to individual differences in cognitive abilities among people of different social network sizes.

Similarly, one may wonder whether it is indeed the size of the social network that matters or whether people with larger social networks receive more input, and it is the increase in amount of input that underlies the benefit of larger social networks. Participants’ interaction logs, however, also provided information about amount of input (number of hours of interaction). Analyses revealed that amount of input did not predict performance on the speech in noise task and that the effect of social network size remained even after controlling for amount of input. These results thus indicate that people who have larger social networks have better phonological abilities, at least when it comes to understanding speech in noise. These results thus show that like infants and second language learners benefit from exposure to multiple speakers, so do adult native speakers.

So is this benefit specific to the phonological level, or can a larger social network boost performance at other linguistic levels as well? A common argument for why variability boosts performance at the phonological level is the lack of invariance at that level. That is, there is no one-to-one mapping at the phono-logical level. Karen might produce a different vowel from Kim when saying hot, and, in fact, her vowel might be the one that Kim uses when saying words, such as hut. Input variability is presumed to assist with dealing with such variability by exposing listeners to the full distribution of productions, indicating the relations among different tokens, and revealing which aspects of the input need to be attended to. Many-to-many mapping, however, is not unique to the phonological level. Suppose you ask your friend about the film that she saw, and she says that it was great. How enthusiastic is she about the film? Should you go and watch it? For some people, great is a highly positive adjective that is used only when something is valued positively. Other people use the word to mean something not much different from OK. Take the following two reviews of the film Spotlight from the film review website rottentomates.com:

• Very eye opening and breathtaking view of the coverups and corruption with the Catholic Church and the community, just sad. Excellent movie well put together seemed as if it was a book and I just kept flipping through pages. Yes must watch.
• Spotlight is a very interesting and important film to see. Not like a typical Oscar-Bait movie whatsoever.

Both reviews are positive. You might however feel that the top review is much more positive as it includes descriptions such as excellent and must watch, which at first blush, might seem stronger than very interesting and important—the descriptions the second reviewer uses. The number of stars those reviewers chose to give the film, however, indicates otherwise, as the bottom reviewer assigned the film 5 stars, whereas the top reviewer assigned only 3.5 stars. As these reviews indicate, at the semantic level, just like the phonological level, there could be many-to-many mapping. Karen might use the term excellent to express a level of enthusiasm that Kim would describe with a term such as nice or pleasant and would use nice for much less positively valued films. If having a larger social network size improves linguistic performance at the phonological level because it helps listeners deal with variability, would having a larger social network similarly improve people’s comprehension of evaluative language? To test that, I recruited 49 participants via Mechanical Turk, an online experiment platform. First, I asked them to indicate how many people they talk to in a typical week. Then, I asked them to read 60 restaurant reviews and estimate the number of stars that the reviewer assigned each restaurant. Results indicated that participants’ social network size predicted how accurate they were in their estimates. Those with larger social networks were more accurate at understanding how enthusiastic (or unenthusiastic) reviewers were about the restaurant they were reviewing (Lev-Ari, 2016).

Of course, these results are correlational. Therefore, the option always remains that participants with larger social networks were better at understanding the reviews not because interaction with more people boosts semantic skills but because people with better semantic skills tend to have larger networks. In other words, the direction of the effect could go in the opposite direction. Therefore, the next step was to experimentally manipulate people’s social network sizes and see whether having a larger social network causes better semantic abilities. One obvious problem is that people already have existing social networks. Therefore, to eliminate...