Computer understanding of natural language

The state of research in automatic natural language processing at Yale University is described. Five programs are discussed: SAM, FRUMP, PAM, TALESPIN, and POLITICS.     

Applying automated deduction to natural language understanding

Very few natural language understanding applications employ methods from automated deduction. This is mainly because (i) a high level of interdisciplinary knowledge is required, (ii) there is a huge gap between formal semantic theory and practical implementation, and (iii) statistical rather than symbolic approaches dominate the current trends in natural language processing. Moreover, abduction rather than deduction is generally viewed as a promising way to apply reasoning in natural language understanding. We describe three applications where we show how first-order theorem proving and finite model construction can efficiently be employed in language understanding.The first is a text understanding system building semantic representations of texts, developed in the late 1990s. Theorem provers are here used to signal inconsistent interpretations and to check whether new contributions to the discourse are informative or not. This application shows that it is feasible to use general-purpose theorem provers for first-order logic, and that it pays off to use a battery of different inference engines as in practice they complement each other in terms of performance.The second application is a spoken-dialogue interface to a mobile robot and an automated home. We use the first-order theorem prover spass for checking inconsistencies and newness of information, but the inference tasks are complemented with the finite model builder mace used in parallel to the prover. The model builder is used to check for satisfiability of the input; in addition, the produced finite and minimal models are used to determine the actions that the robot or automated house has to execute. When the semantic representation of the dialogue as well as the number of objects in the context are kept fairly small, response times are acceptable to human users.The third demonstration of successful use of first-order inference engines comes from the task of recognising entailment between two (short) texts. We run a robust parser producing semantic representations for both texts, and use the theorem prover vampire to check whether one text entails the other. For many examples it is hard to compute the appropriate background knowledge in order to produce a proof, and the model builders mace and paradox are used to estimate the likelihood of an entailment.     

Inhibition and language pragmatic view in redundant data problem solving

The present study concerns redundant data problems, defined as problems in which irrelevant data is provided. This type of problem provides a misleading context [Pascual-Leone, J. (1987). Organismic process for neo-Piagetian theories: A dialectical causal account of cognitive development. International Journal of Psychology, 22, 531–570] similar to the Piagetian class inclusion task. The two experiments performed showed that adding the a-pragmatic cue ‘in total’ to the final question of redundant data problems increased both 7- and 9-year-old children's failure rate, without modifying the structure of the problem. This indicates that a non-structural linguistic cue can bias children towards failure by enhancing the lack of inhibition of the ordinary adding schema.     

Acquiring an understanding of design: evidence from children's insight problem solving

The human ability to make tools and use them to solve problems may not be zoologically unique, but it is certainly extraordinary. Yet little is known about the conceptual machinery that makes humans so competent at making and using tools. Do adults and children have concepts specialized for understanding human-made artifacts? If so, are these concepts deployed in attempts to solve novel problems? Here we present new data, derived from problem-solving experiments, which support the following. (i) The structure of the child's concept of artifact function changes profoundly between ages 5 and 7. At age 5, the child's conceptual machinery defines the function of an artifact as any goal a user might have; by age 7, its function is defined by the artifact's typical or intended use. (ii) This conceptual shift has a striking effect on problem-solving performance, i.e. the child's concept of artifact function appears to be deployed in problem solving. (iii) This effect on problem solving is not caused by differences in the amount of knowledge that children have about the typical use of a particular tool; it is mediated by the structure of the child's artifact concept (which organizes and deploys the child's knowledge). In two studies, children between 5 and 7 years of age were matched for their knowledge of what a particular artifact "is for", and then given a problem that can only be solved if that tool is used for an atypical purpose. All children performed well in a baseline condition. But when they were primed by a demonstration of the artifact's typical function, 5-year-old children solved the problem much faster than 6-7-year-old children. Because all children knew what the tools were for, differences in knowledge alone cannot explain the results. We argue that the older children were slower to solve the problem when the typical function was primed because (i) their artifact concept plays a role in problem solving, and (ii) intended purpose is central to their concept of artifact function, but not to that of the younger children.     

Children's understanding of natural language hierarchies

Three experiments examined young children's ability to evaluate the relationship between concepts as one of inclusion. In Experiment 1, the same group of 4-to 7-year-old children were given three tasks: one in which they made judgments about whether “all” or “some” members of a category were included in another category, and two tasks in which they made inferences based on knowledge of inclusion relations. The majority of children succeeded on at least one of the tasks, thereby implying that they could evaluate inclusion relations. Two further experiments provided more evidence for this hypothesis: Experiment 2 confirmed that nursery children could answer quantitative questions about conceptual interrelationts; Experiment 3 demonstrated that nursery schoolers could solve certain inference problems involving the construction and evaluation of hierarchies.     

Goal specificity and knowledge acquisition in statistics problem solving: Evidence for attentional focus

Solving training problems with nonspecific goals (NG; i.e., solving for all possible unknown values) often results in better transfer than solving training problems with standard goals (SG; i.e., solving for one particular unknown value). In this study, we evaluated an attentional focus explanation of the goal specificity effect. According to the attentional focus view, solving NG problems causes attention to be directed to local relations among successive problem states, whereas solving SG problems causes attention to be directed to relations between the various problem states and the goal state. Attention to the former is thought to enhance structural knowledge about the problem domain and thus promote transfer. Results supported this view because structurally different transfer problems were solved faster following NG training than following SG training. Moreover, structural knowledge representations revealed more links depicting local relations following NG training and more links to the training goal following SG training. As predicted, these effects were obtained only by domain novices.     

A model of problem solving with incomplete constraint knowledge

The problem-solving behavior of subjects presented with a series of Tower of Hanoi problems is examined. A production system model which incorporates elements of domain-specific knowledge into a general problem-solving framework is presented. Other models developed for the task are based on understanding of complete solution strategies and are not satisfactory models of nonexpert human performance. The current model discriminates between problem-solving behavior based on constraint knowledge and behavior based on nonspecific general search strategies. A variety of move choice and latency measures are used to compare the performance of the model to human subject performance.     

Representing object colour in language comprehension

Embodied theories of cognition hold that mentally representing something red engages the neural subsystems that respond to environmental perception of that colour. This paper examines whether implicit perceptual information on object colour is represented during sentence comprehension even though doing so does not necessarily facilitate task performance. After reading a sentence that implied a particular colour for a given object, participants were presented with a picture of the object that either matched or mismatched the implied colour. When asked if the pictured object was mentioned in the preceding sentence, people's responses were faster when the colours mismatched than when they matched, suggesting that object colour is represented differently to other object properties such as shape and orientation. A distinction between stable and unstable embodied representations is proposed to allow embodied theories to account for these findings.     

Planning paths to multiple targets: memory involvement and planning heuristics in spatial problem solving

For large numbers of targets, path planning is a complex and computationally expensive task. Humans, however, usually solve such tasks quickly and efficiently. We present experiments studying human path planning performance and the cognitive processes and heuristics involved. Twenty-five places were arranged on a regular grid in a large room. Participants were repeatedly asked to solve traveling salesman problems (TSP), i.e., to find the shortest closed loop connecting a start location with multiple target locations. In Experiment 1, we tested whether humans employed the nearest neighbor (NN) strategy when solving the TSP. Results showed that subjects outperform the NN-strategy, suggesting that it is not sufficient to explain human route planning behavior. As a second possible strategy we tested a hierarchical planning heuristic in Experiment 2, demonstrating that participants first plan a coarse route on the region level that is refined during navigation. To test for the relevance of spatial working memory (SWM) and spatial long-term memory (LTM) for planning performance and the planning heuristics applied, we varied the memory demands between conditions in Experiment 2. In one condition the target locations were directly marked, such that no memory was required; a second condition required participants to memorize the target locations during path planning (SWM); in a third condition, additionally, the locations of targets had to retrieved from LTM (SWM and LTM). Results showed that navigation performance decreased with increasing memory demands while the dependence on the hierarchical planning heuristic increased.     

Expertise as mental set: The effects of domain knowledge in creative problem solving

Experts generally solve problems in their fields more effectively than novices because their wellstructured, easily activated knowledge allows for efficient search of a solution space. But what happens when a problem requires a broad search for a solution? One concern is that subjects with a large amount of domain knowledge may actually be at a disadvantage, because their knowledge may confine them to an area of the search space in which the solution does not reside. In other words, domain knowledge may act as a mental set, promoting fixation in creative problem-solving attempts. A series of three experiments in which an adapted version of Mednick’s (1962) remote associates task was used demonstrates conditions under which domain knowledge may inhibit creative problem solving.     

Worrying about chronic pain: vigilance to threat and misdirected problem solving

Worry is reviewed to inform a cognitive-behavioural understanding of chronic pain. How worry comes to dominate is discussed as 'vigilance' and why worry comes to dominate is discussed as 'problem solving'. These two aspects are applied to worrying about chronic pain. Chronic pain is re-presented as chronic vigilance to threat that may lead to a perservation of attempts at solving the problem of achieving escape from pain. Attempts will be frustrated by the insolubility of the problem of chronic pain. The clinical and theoretical implications of this application are discussed.     

Act first, think later: The presence and absence of inferential planning in problem solving

Planning is fundamental to successful problem solving, yet individuals sometimes fail to plan even one step ahead when it lies within their competence to do so. In this article, we report two experiments in which we explored variants of a ball-weighing puzzle, a problem that has only two steps, yet nonetheless yields performance consistent with a failure to plan. The results fit a computational model in which a solver’s attempts are determined by two heuristics: maximization of the apparent progress made toward the problem goal and minimization of the problem space in which attempts are sought. The effectiveness of these heuristics was determined by lookahead, defined operationally as the number of steps evaluated in a planned move. Where move outcomes cannot be visualized but must be inferred, planning is constrained to the point where some individuals apply zero lookahead, which with n-ball problems yields seemingly irrational unequal weighs. Applying general-purpose heuristics with or without lookahead accounts for a range of rational and irrational phenomena found with insight and noninsight problems.     

Representing time in language and memory: The role of similarity structure

Every day we read about or watch events in the world and can easily understand or remember how long they last. What aspects of an event are retained in memory? And how do we extract temporal information from our memory representations? These issues are central to human cognition, as they underlie a fundamental aspect of our mental life, namely our representation of time. This paper reviews previous language studies and reports a visual learning study indicating that properties of the events encoded in memory shape the representation of their duration. The evidence indicates that for a given event, the extent to which its associated properties or sub-components differ from one another modulates our representation of its duration. These properties include the similarity between sub-events and the similarity between the situational contexts in which an event occurs. We suggest that the diversity of representations that we associate with events in memory plays an important role in remembering and estimating the duration of experienced or described events.     

Ways of thinking about and teaching ethical problem solving: Microethics and macroethics in engineering

Engineering ethics entails three frames of reference: individual, professional, and social. "Microethics" considers individuals and internal relations of the engineering profession; "macroethics" applies to the collective social responsibility of the profession and to societal decisions about technology. Most research and teaching in engineering ethics, including online resources, has had a "micro" focus. Mechanisms for incorporating macroethical perspectives include: integrating engineering ethics and science, technology and society (STS); closer integration of engineering ethics and computer ethics; and consideration of the influence of professional engineering societies and corporate social responsibility programs on ethical engineering practice. Integrating macroethical issues and concerns in engineering ethics involves broadening the context of ethical problem solving. This in turn implies: developing courses emphasizing both micro and macro perspectives, providing faculty development that includes training in both STS and practical ethics; and revision of curriculum materials, including online resources. Multidisciplinary collaboration is recommended 1) to create online case studies emphasizing ethical decision making in individual, professional, and societal contexts; 2) to leverage existing online computer ethics resources with relevance to engineering education and practice; and 3) to create transparent linkages between public policy positions advocated by professional societies and codes of ethics.     

The influence of goal-state access cost on planning during problem solving

Two problem-solving experiments investigated the relationship between planning and the cost of accessing goal-state information using the theoretical framework of the soft constraints hypothesis (Gray & Fu, 2004; Gray, Simms, Fu, & Schoelles, 2006). In Experiment 1, 36 participants were allocated to low, medium, and high access cost conditions and completed a problem-solving version of the Blocks World Task. Both the nature of planning (memory based or display based) and its timing (before or during action) changed with high goal-state access cost (a mouse movement and a 2.5-s delay). In this condition more planning before action was observed, with less planning during action, evidenced by longer first-move latencies, more moves per goal-state inspection, and more short (≤0.8 s) and long (>8 s) “preplanned” intermove latencies. Experiment 2 used an eight-puzzle-like transformation task and replicated the effect of goal-state access cost when more complex planning was required, also confirmed by sampled protocol data. Planning before an episode of move making increased with higher goal-state access cost, and planning whilst making moves increased with lower access cost. These novel results are discussed in the context of the soft constraints hypothesis.     

The influence of goal-state access cost on planning during problem solving

Two problem-solving experiments investigated the relationship between planning and the cost of accessing goal-state information using the theoretical framework of the soft constraints hypothesis (Gray & Fu, 2004; Gray, Simms, Fu, & Schoelles, 2006). In Experiment 1, 36 participants were allocated to low, medium, and high access cost conditions and completed a problem-solving version of the Blocks World Task. Both the nature of planning (memory based or display based) and its timing (before or during action) changed with high goal-state access cost (a mouse movement and a 2.5-s delay). In this condition more planning before action was observed, with less planning during action, evidenced by longer first-move latencies, more moves per goal-state inspection, and more short (≤ 0.8 s) and long (>8 s) "preplanned" intermove latencies. Experiment 2 used an eight-puzzle-like transformation task and replicated the effect of goal-state access cost when more complex planning was required, also confirmed by sampled protocol data. Planning before an episode of move making increased with higher goal-state access cost, and planning whilst making moves increased with lower access cost. These novel results are discussed in the context of the soft constraints hypothesis.