Traditional approaches to language processing have been based on explicit, discrete representations which are difficult to learn from a reasonable linguistic environment—hence, it has come to be accepted that much of our linguistic representations and knowledge is innate. With its focus on learning based upon graded, malleable, distributed representations, connectionist modeling has reopened the question of what could be learned from the environment in the absence of detailed innate knowledge. This paper provides an overview of connectionist models of language processing, at both the lexical and sentence levels.
The Construction-Integration (CI) Model (Kintsch 1988, 1992, 1998a; Kintsch & Welsch, 1991) is a cognitive architecture of comprehension. It is a hybrid model that comprises the construction phase employing weak symbolic rules and the integration phase with a connectionist constraint-satisfaction algorithm. In the CI Model, comprehension emerges from an interaction between the text to be comprehended and the comprehender's knowledge and episodic memory. This article first provides an overview of the model. It then shows how the CI model can simulate discourse comprehension processes to qualitatively account for empirical data from studies on inference generation. In so doing, the author points out a simplifying assumption about the construction process of a working-memory representation in the simulation and argues that the construction phase itself can be characterized as the repetitive application of construction and integration. The paper then discusses a modification to the model by introducing the incremental construction-integration procedure and presents a case study that illustrates what insights this refinement gives into simulating inference processing.
We are capable of drawing a variety of inferences effortlessly, spontaneously, and with remarkable efficiency—as though these inferences are a reflex response of our cognitive apparatus. This remarkable human ability poses a challenge for cognitive science and computational neuroscience: How can a network of slow neuron-like elements represent a large body of systematic knowledge and perform a wide range of inferences with such speed? We briefly review work in connectionist modeling that attempts to address this challenge by demonstrating how a neurally plausible network can encode a large body of semantic and episodic facts, systematic rules, and knowledge about entities and types, and yet perform a wide range of explanatory and predictive inferences within a few hundred milliseconds.
This paper reviews our work simulating human thinking with the LISA model. Human mental representations are both flexible and structure-sensitive—properties that jointly present challenging design requirements for a model of the cognitive architecture. LISA satisfies these requirements by representing relational roles and their fillers as patterns of activation distributed over a collection of semantic units (achieving flexibility) and binding these representations dynamically into propositional structures using synchrony of firing (achieving structure-sensitivity). The resulting representations serve as a natural basis for memory retrieval, analogical mapping, analogical inference and schema induction. In addition, the LISA architecture provides an integrated account of effortless “reflexive” forms of inference and more effortful “reflective” inference, serves as a natural basis for integrating generalized procedures for relational reasoning with modules for more specialized forms of reasoning (e.g., reasoning about objects in spatial arrays), provides an a priori account of the limitations of human working memory, and may serve as a platform for understanding the neural basis of symbolic thought.
The ability to generalize—to abstract regularities from our experiences that can be applied to new experiences—is fundamental to human cognition and our abilities to flexibly adapt to changing situations. However, the generalization abilities of children and adults are far from perfect, with many clear demonstrations of failures to generalize in situations that would otherwise appear to lend themselves to generalization. It seems that people require extensive experience with a domain to demonstrate good generalization, and that their generalization abilities are best when dealing with relatively concrete, familiar situations. In this paper, we argue that people's successes and failures in generalization are well characterized by neural network models. Networks of neurons connected by synaptic weights are naturally predisposed to encode information in a highly specific fashion, which does not support generalization (as has been seized upon by critics of such models). However, with sufficient experience and appropriate architectural properties, such models can develop abstract representations that support good generalization. Implications for the neural bases and development of generalization abilities are discussed.
We constructed a connectionist model of the phonological loop, which develops an internal representation of phonology through imitation-style learning. After learning a large corpus of English monosyllabic words, the model could repeat and rehearse not only well-learned words, but also novel nonwords. The model was also able to acquire a representation of the identity of words through a self-organizing learning rule, and to recall the serial order of the words. We showed that the capacity of the phonological loop and the speed of the vocabulary acquisition depended on the rate of updating weights in the model. Furthermore, the model could replicate the effect of long-term memory on the immediate serial recall of words and nonwords.
In the infancy of human being, it is known that the number of words in speech increase drastically. We think a word acquisition boosting of this period occurs according to the fast mapping in the learning system which is controlled by a meta-information about the language situation. To explain the boosting mechanism, we propose a neural network model of the meta-information that consists of a prediction part, which is a simple recurrent neural network, and a learning evaluation part that controls the fast learning. The learning evaluation part learns a confidence of learning progress as the meta-information from a representation of recurrent network. By a computer simulation study, we show that the meta-information is learnable in spite of its luck of saliency and that the use of meta-information results accelerative learning.
The purpose of this study is to develop a model of the decision-making process with the feedback mental transaction in which the result of decision making influences the evaluation of selective factors. Typical examples of this feedback mental transaction, including ‘resolution of cognitive dissonance’ and ‘rationalization’, are well-known in psychology. Namely, the evaluation of decision factors is influenced not by the evaluation of the decision but by the decision itself. At first, we assumed that the decision-making process consists of two steps. The first is the process by which certain candidates are selected from all alternatives and the latter is the process by which the final one is chosen from among the screened candidates. Then, we constructed a model which represents the interactive influences between preferable alternatives and decision factors via neural networks with feedback links. We used questionnaire results about customers' car purchases to estimate the weight parameters in the neural networks. Results of the simulation using this model showed positive agreement with the questionnaire results.
The effects of the speaker's adjustment of the teaching strategy and the use of paralanguage information of speech to acquire its meaning were clarified by means of experiment: In the experiments, two subjects played a game of Pong: one of the subjects (operator) could not understand linguistically what the other one (teacher) was saying. The results of these experiments revealed the following. First, the teacher's high-pitched voice drew attention of the operator's current action. Second, the process of meaning acquisition can be regarded as reinforcement learning based on a multi-reward system (i.e., a positive reward for correct actions and a negative reward for incorrect actions, given in the form of the teacher's high-pitched voice). Finally, mutual adaptation between the subjects was observed, that is, the subjects learned to respond appropriately to each other's behavior. It is concluded that the above three phenomena are important to the process of meaning acquisition and can be viewed as the basic requirements to enable the acquisition of meaning of unknown speech, and to construct an adaptive sound interface, which can provide a natural interaction enviromnet for its user.
The constructional disabilities are cognitive impairments that show difficulties in visually guided construction of spatial patterns (copying line drawings, block patterns and finger constructions). The cause of them is less well understood. To make clear it we investigated the behavioral characteristics of a brain injured subject who had difficulties in copying finger constructions. Video tape recording of the subject's finger copying revealed that he needed visual confirmation of his own hand. It implied that the difficulty came from the disturbance of the internal feedback system i.e., the forward model. To confirm it his mental manipulation ability of finger constructions and block patterns was investigated. The result showed that the subject had disturbance in the mental manipulation of the hand. Several studies suggested that the mental manipulation was executed by the mental models of external environment including own body. Forward model is the major component of these mental models. The result of our study suggested that the critical problem of the constructional disabilities was come from the forward model impairment. And this function was also suspected to be concerned with the left lateral parietal cortex.
Yokosawa, Subramaniam, and Biederman (1996) asked their subjects to verify whether successively-presented two objects were the same or different in the basic-level naming class. When the two objects were different but belonged to the same superordinate category (e.g., a banana and an apple), the verifying time became slower than when they had no relation, independent of their perceptual similarity. This interference is termed the superordinate similarity effect, which indicates that superordinate-level information is processed automatically and will interfere even in the basic-level verification tasks. On the other hand, it has been controversial whether surface characteristics (e.g., color, texture, and brightness) affect basic-level recognition and superordinate-level processing of visual objects (Biederman & Ju, 1988; Price & Humphreys, 1989 etc.). In the present study, three experiments were conducted to examine the effect of surface characteristics upon the verification speed at the basic level and the interference at the superordinate level. As a result, surface characteristics facilitated the basic-level verification but did not affect the superordinate-level interference. This suggests that there exists direct link from surface characteristics to the basic-level representations, but not to the superordinate-level ones.
When we estimate subjective probability about “It will be fine tomorrow”, there are other possibilities about tomorrow's weather (“cloudy”, “rain”, “snow”). In other words, when we estimate subjective probability about an event (focal event), there are other events (alternative events) that are probable but not the targets of probability judgments. Previous studies suggested that alternative events affect the subjective probability of the focal event. In this study, we proposed a descriptive model of subjective probability that considers the effects of the randomness of alternative events to the subjective probability of focal events. Participants were shown various outcomes of multinomial events, and instructed to estimate the subjective probability of the specific events of the multinomial events. Participants answered subjective probabilities by verbal impression (verbal estimate) or objective percentage (percent estimate). The randomness of the alternative events was defined by the expression used in Rappoport & Budescu (1997). The results of multiple regression analysis indicated that the randomness of alternative events significantly affected the subjective probability of a focal event both in verbal estimate or percent estimate, so our model was supported. Some theoretical suggestions were discussed.