Sequential sampling models of human text classification

Text classification involves deciding whether or not a document is about a given topic. It is an important problem in machine learning, because automated text classifiers have enormous potential for application in information retrieval systems. It is also an interesting problem for cognitive science, because it involves real world human decision making with complicated stimuli. This paper develops two models of human text document classification based on random walk and accumulator sequential sampling processes. The models are evaluated using data from an experiment where participants classify text documents presented one word at a time under task instructions that emphasize either speed or accuracy, and rate their confidence in their decisions. Fitting the random walk and accumulator models to these data shows that the accumulator provides a better account of the decisions made, and a “balance of evidence” measure provides the best account of confidence. Both models are also evaluated in the applied information retrieval context, by comparing their performance to established machine learning techniques on the standard Reuters‐21578 corpus. It is found that they are almost as accurate as the benchmarks, and make decisions much more quickly because they only need to examine a small proportion of the words in the document. In addition, the ability of the accumulator model to produce useful confidence measures is shown to have application in prioritizing the results of classification decisions.     

An analysis of sequential variables in Pavlovian conditioning employing extended and limited acquisition training

Sequential theory’s memory model of learning has been successfully applied in response contingent instrumental conditioning experiments (Capaldi, 1966, Capaldi, 1967, Capaldi, 1994 and Capaldi and Miller, 2003). However, it has not been systematically tested in nonresponse contingent Pavlovian conditioning experiments. The present experiments attempted to determine if several sequential variables affect responding in Pavlovian situations as they do in instrumental ones. Of primary concern here were the effects on extinction of number of NR transitions (the number of times a nonreinforced trial is followed by a reinforced trial), N-length (the number of successive nonreinforced trials that precede a reinforced trial), and percentage of reinforcement (50 versus 100%) following either extended acquisition training (Experiment 1, 720 trials) or limited acquisition training (Experiment 3, 24 trials). In agreement with a sequential analysis, N-length increased resistance to extinction more than number of NR transitions following extensive training with the opposite occurring following limited training. In Experiment 1, greater resistance to extinction was associated with 50% than with 100% reinforcement, a partial reinforcement extinction effect (PREE). Experiment 2 examined an anomalous finding obtained in Experiment 1. A major theoretical difference between instrumental and Pavlovian conditioning has been held to be the greater ease of producing a PREE in instrumental than in Pavlovian conditioning (Kimble, 1961 and Mackintosh, 1974). However, the findings obtained here suggest that the probability of obtaining a PREE and other Pavlovian extinction effects, as in instrumental conditioning, increases along with the effectiveness of the sequential variables employed.     

The dynamics of choice among multiple alternatives

We consider neurally based models for decision-making in the presence of noisy incoming data. The two-alternative forced-choice task has been extensively studied, and in that case it is known that mutually inhibited leaky integrators in which leakage and inhibition balance can closely approximate a drift-diffusion process that is the continuum limit of the optimal sequential probability ratio test (SPRT). Here we study the performance of neural integrators in n?2 alternative choice tasks and relate them to a multihypothesis sequential probability ratio test (MSPRT) that is asymptotically optimal in the limit of vanishing error rates. While a simple race model can implement this ‘max-vs-next’ MSPRT, it requires an additional computational layer, while absolute threshold crossing tests do not require such a layer. Race models with absolute thresholds perform relatively poorly, but we show that a balanced leaky accumulator model with an absolute crossing criterion can approximate a ‘max-vs-ave’ test that is intermediate in performance between the absolute and max-vs-next tests. We consider free and fixed time response protocols, and show that the resulting mean reaction times under the former and decision times for fixed accuracy under the latter obey versions of Hick's law in the low error rate range, and we interpret this in terms of information gained. Specifically, we derive relationships of the forms log(n-1), log(n), or log(n+1) depending on error rates, signal-to-noise ratio, and the test itself. We focus on linearized models, but also consider nonlinear effects of neural activities (firing rates) that are bounded below and show how they modify Hick's law.     

Heuristics as beliefs and as behaviors: the adaptiveness of the "hot hand"

and analyzed reasoning by asking: what are the reasoner's goals? This emphasizes the adaptiveness of behavior rather than whether a belief is normative. Belief in the "hot hand" in basketball suggests that players experiencing streaks should be given more shots, but this has been seen as a fallacy due to failure to find dependencies between players' shots. Based on their findings, I demonstrate by Markov modeling and simulation that streaks are valid allocation cues for deciding who to give shots to, because this behavior achieves the team goal of scoring more. Empirically I show that this adaptive heuristic is supported by the fallacious belief in dependency, more so as skill level increases. I extend the theoretical analysis to identify general conditions under which following streaks should be beneficial. Overall, this approach illustrates the advantages of analyzing reasoning in terms of adaptiveness.     

A Bayesian Theory of Sequential Causal Learning and Abstract Transfer

Two key research issues in the field of causal learning are how people acquire causal knowledge when observing data that are presented sequentially, and the level of abstraction at which learning takes place. Does sequential causal learning solely involve the acquisition of specific cause‐effect links, or do learners also acquire knowledge about abstract causal constraints? Recent empirical studies have revealed that experience with one set of causal cues can dramatically alter subsequent learning and performance with entirely different cues, suggesting that learning involves abstract transfer, and such transfer effects involve sequential presentation of distinct sets of causal cues. It has been demonstrated that pre‐training (or even post‐training) can modulate classic causal learning phenomena such as forward and backward blocking. To account for these effects, we propose a Bayesian theory of sequential causal learning. The theory assumes that humans are able to consider and use several alternative causal generative models, each instantiating a different causal integration rule. Model selection is used to decide which integration rule to use in a given learning environment in order to infer causal knowledge from sequential data. Detailed computer simulations demonstrate that humans rely on the abstract characteristics of outcome variables (e.g., binary vs. continuous) to select a causal integration rule, which in turn alters causal learning in a variety of blocking and overshadowing paradigms. When the nature of the outcome variable is ambiguous, humans select the model that yields the best fit with the recent environment, and then apply it to subsequent learning tasks. Based on sequential patterns of cue‐outcome co‐occurrence, the theory can account for a range of phenomena in sequential causal learning, including various blocking effects, primacy effects in some experimental conditions, and apparently abstract transfer of causal knowledge.     

Effects of stroking horses on both humans' and horses' heart rate responses1

Abstract: The present study examined both human and horse heart rates (HRs) when humans stroked horses for 90 seconds; the subjective arousal levels of the humans were measured by the Tohoku Activation Deactivation Adjective Check List before and after stroking horses. Six male subjects with a positive attitude toward companion animals and 6 male subjects with a negative attitude were selected by their scores on the Pet Attitude Scale, and these two groups, together with a third group, of 6 subjects who were male members of the Doshisha University horse-riding club, participated in this experiment. The HRs of the human subjects during the first 10 seconds immediately after the stroking began were significantly higher than those obtained after that period, but these higher levels gradually returned to baseline levels. This tendency appears more clearly in the negative attitude group. The HRs of the horses increased during the first 20 seconds immediately after the human subjects of the NA group started stroking them, but gradually reduced as the stroking continued. The results of subjective arousal levels suggest a decrease in tension by stroking horses. These results suggest that a certain affectional interaction may exist between humans and companion animals.     

Simulating the Acquisition of Verb Inflection in Typically Developing Children and Children With Developmental Language Disorder in English and Spanish

Children with developmental language disorder (DLD) have significant deficits in language ability that cannot be attributed to neurological damage, hearing impairment, or intellectual disability. The symptoms displayed by children with DLD differ across languages. In English, DLD is often marked by severe difficulties acquiring verb inflection. Such difficulties are less apparent in languages with rich verb morphology like Spanish and Italian. Here we show how these differential profiles can be understood in terms of an interaction between properties of the input language, and the child's ability to learn predictive relations between linguistic elements that are separated within a sentence. We apply a simple associative learning model to sequential English and Spanish stimuli and show how the model's ability to associate cues occurring earlier in time with later outcomes affects the acquisition of verb inflection in English more than in Spanish. We relate this to the high frequency of the English bare form (which acts as a default) and the English process of question formation, which means that (unlike in Spanish) bare forms frequently occur in third-person singular contexts. Finally, we hypothesize that the pro-drop nature of Spanish makes it easier to associate person and number cues with the verb inflection than in English. Since the factors that conspire to make English verb inflection particularly challenging for learners with weak sequential learning abilities are much reduced or absent in Spanish, this provides an explanation for why learning Spanish verb inflection is relatively unaffected in children with DLD.     

Criteria instability and the isolated option effect

In two-stage choices, decision makers often compare a new (isolated) option with the winner from the first stage. Previous research has identified a choice advantage for an isolated option, ostensibly due to loss aversion. We propose an alternative mechanism suggesting that instability of the criteria used in each choice stage is the main driver of the isolated option effect. Results from a series of experiments support the criteria instability account and not loss aversion as the explanation for the isolated option effect.     

Discontinuity in the enumeration of sequentially presented auditory and visual stimuli

The seeking of discontinuity in enumeration was recently renewed because Cowan [Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24, 87–185; Cowan, N. (2005). Working memory capacity. Hove: Psychology Press] suggested that it allows evaluating the limit of the focus of attention, currently estimated at four items. A strong argument in favour of a general constraint of the cognitive system is that similar discontinuities should be observed in modalities different from the classic simultaneous presentation of visual objects. Recently, data were provided on tactile stimuli, but the authors diverged in their conclusion about the existence of such discontinuity [Gallace, A., Tan, H. Z., & Spence, C. (2006). Numerosity judgments for tactile stimuli distributed over the body surface. Perception, 35(2), 247–266; Riggs, K. J., Ferrand, L., Lancelin, D., Fryziel, L., Dumur, G., & Simpson, A. (2006). Subitizing in tactile perception. Psychological Science, 17(4), 271–272]. Following a similar rationale, our study aimed at evaluating discontinuity in the enumeration of auditory and visual stimuli presented sequentially. The clear and similar discontinuity observed in error rates, response times and given responses for both modalities favours the general capacity limit view, but also questions the size of this capacity, because the discontinuity occurred here at size 2. However, the masking of stimuli in sensory memory could not be entirely discarded.     

Sequential responding to arabic numerals with wild cards by the chimpanzee (Pan troglodytes)

One adult female chimpanzee (Pan troglodytes) was trained to respond serially to three arabic numerals between 1 and 9, presented on a cathode-ray-tube (CRT) screen. To examine the factors affecting her sequential responding behavior, wild-card items were added to the three-item sequences. When this wild-card item remained until the subject responded to the last numeral (i.e., the terminator condition), her response to the terminator at each point of the sequence was controlled by the ordinal distance between numerals. Thus, the number of responses to the terminator increased as the ordinal distance between numerals increased. When the wild-card item was eliminated by the subject’s response (wild-card conditions), the probability of responses to the wild card before the first numeral increased as a function of the serial position of the first numeral. These results were consistent with previous studies of response time and suggest both serial position and symbolic distance effects. It is suggested that the subject might form the integrated 9-item linear representations by training of possible subsets of three-item sequences. Knowledge concerning the ordinal position of each numeral was established through this training. Received: 27 October 1999 / Accepted: 22 November 1999