"I Don’t Believe in Word Senses"

Word sense disambiguation assumes word senses. Withinthe lexicography and linguistics literature, they areknown to bevery slippery entities. The first part of the paperlooks at problemswith existing accounts of ‘word sense’ and describesthe various kinds of ways in which a word's meaning candeviate from its coremeaning. An analysis is presented in which wordsenses areabstractions from clusters of corpus citations, inaccordance withcurrent lexicographic practice. The corpus citations,not the wordsenses, are the basic objects in the ontology. Thecorpus citationswill be clustered into senses according to thepurposes of whoever or whatever does the clustering. In theabsence of suchpurposes, word senses do not exist. Word sense disambiguation also needs a set of wordsenses todisambiguate between. In most recent work, the sethas been takenfrom a general-purpose lexical resource, with theassumption that thelexical resource describes the word senses ofEnglish/French/...,between which NLP applications will need todisambiguate. Theimplication of the first part of the paper is, bycontrast, that wordsenses exist only relative to a task. Thefinal part of the paper pursues this, exploring, bymeans of asurvey, whether and how word sense ambiguity is infact a problem forcurrent NLP applications.     

Acquisition of selectional patterns in sublanguages

When implementing computational lexicons it is important to keep in mind the texts that a NLP system must deal with. Words relate to each other in many different, often odd ways this information is rarely found in dictionaries, and it is quite hard to deduce a priori. In this paper we present a technique for the acquisition of statistically significant selectional restrictions from corpora and discuss the results of an experimental application with reference to two specific sublaguages (legal and commercial). We show that there are important cooccurrence preferences among words which cannot be established a priori as they are determined for each choice of sublanguage. The method for detecting cooccurrences is based on the analysis of word associations augmented with syntactic markers and semantic tags. Word pairs are extracted by a morphosyntactic analyzer and clustered according to their semantic tags. A statistical measure is applied to the data to evaluate the sigificance of any relations detected. Selectional restrictions are acquired by a two-step process. First, statistically prevailing coarse grained conceptual patterns are used by a linguist to identify the relevant selectional restrictions in sublanguages. Second, semiautomatically acquired coarse selectional restrictions are used as the semantic bias of a system, ARIOSTO_LEX, for the automatic acquisition of a case-based semantic lexicon.     

Behaviour Synthesis of the Erect Stance for a Biped Control

In this paper we use free fall approach to develop a high level control/command strategy for a bipedal robot called BIPMAN, based on a multi-chain mechanical model with a general control architecture. The strategy is composed of three levels: the Legs and arms level, the Coordinator level and the Supervisor level. The Coordinator level is devoted to controlling leg movements and to ensure the stability of the whole biped. Actually perturbation effects threaten the equilibrium of the human robot and can only be compensated using a dynamic control strategy. This one is based on dynamic stability studies with a center of mass acceleration control and a force distribution on each leg and arm. Free fall in the gravity field is assumed to be deeply involved in the human locomotor control. According to studies of this specific motion through a direct dynamic model,the notion of equilibrium classes is introduced. They allow one to define time intervals in which the biped is able to maintain its posture. This notion is used for the definition of a reconfigurable high level control of the robot.     

Similarity-Based Models of Word Cooccurrence Probabilities

In many applications of natural language processing (NLP) it is necessary to determine the likelihood of a given word combination. For example, a speech recognizer may need to determine which of the two word combinations eat a peach and eat a beach is more likely. Statistical NLP methods determine the likelihood of a word combination from its frequency in a training corpus. However, the nature of language is such that many word combinations are infrequent and do not occur in any given corpus. In this work we propose a method for estimating the probability of such previously unseen word combinations using available information on most similar words.We describe probabilistic word association models based on distributional word similarity, and apply them to two tasks, language modeling and pseudo-word disambiguation. In the language modeling task, a similarity-based model is used to improve probability estimates for unseen bigrams in a back-off language model. The similarity-based method yields a 20% perplexity improvement in the prediction of unseen bigrams and statistically significant reductions in speech-recognition error.We also compare four similarity-based estimation methods against back-off and maximum-likelihood estimation methods on a pseudo-word sense disambiguation task in which we controlled for both unigram and bigram frequency to avoid giving too much weight to easy-to-disambiguate high-frequency configurations. The similarity-based methods perform up to 40% better on this particular task.     

A Full Bayesian Approach to Curve and Surface Reconstruction

When interpolating incomplete data, one can choose a parametric model, or opt for a more general approach and use a non-parametric model which allows a very large class of interpolants. A popular non-parametric model for interpolating various types of data is based on regularization, which looks for an interpolant that is both close to the data and also smooth in some sense. Formally, this interpolant is obtained by minimizing an error functional which is the weighted sum of a fidelity term and a smoothness term.The classical approach to regularization is: select optimal weights (also called hyperparameters) that should be assigned to these two terms, and minimize the resulting error functional.However, using only the optimal weights does not guarantee that the chosen function will be optimal in some sense, such as the maximum likelihood criterion, or the minimal square error criterion. For that, we have to consider all possible weights.The approach suggested here is to use the full probability distribution on the space of admissible functions, as opposed to the probability induced by using a single combination of weights. The reason is as follows: the weight actually determines the probability space in which we are working. For a given weight , the probability of a function f is proportional to exp(– f² _uu du) (for the case of a function with one variable). For each different , there is a different solution to the restoration problem; denote it by f. Now, if we had known , it would not be necessary to use all the weights; however, all we are given are some noisy measurements of f, and we do not know the correct . Therefore, the mathematically correct solution is to calculate, for every , the probability that f was sampled from a space whose probability is determined by , and average the different f's weighted by these probabilities. The same argument holds for the noise variance, which is also unknown.Three basic problems are addressed is this work: Computing the MAP estimate, that is, the function f maximizing Pr(f/D) when the data D is given. This problem is reduced to a one-dimensional optimization problem. Computing the MSE estimate. This function is defined at each point x as f(x)Pr(f/D) f. This problem is reduced to computing a one-dimensional integral.In the general setting, the MAP estimate is not equal to the MSE estimate. Computing the pointwise uncertainty associated with the MSE solution. This problem is reduced to computing three one-dimensional integrals.     

Symbols and Computation A Critique of the Computational Theory of Mind

Over the past several decades, the philosophical community has witnessed the emergence of an important new paradigm for understanding the mind.¹ The paradigm is that of machine computation, and its influence has been felt not only in philosophy, but also in all of the empirical disciplines devoted to the study of cognition. Of the several strategies for applying the resources provided by computer and cognitive science to the philosophy of mind, the one that has gained the most attention from philosophers has been the Computational Theory of Mind (CTM). CTM was first articulated by Hilary Putnam (1960, 1961), but finds perhaps its most consistent and enduring advocate in Jerry Fodor (1975, 1980, 1981, 1987, 1990, 1994). It is this theory, and not any broader interpretations of what it would be for the mind to be a computer, that I wish to address in this paper. What I shall argue here is that the notion of symbolic representation employed by CTM is fundamentally unsuited to providing an explanation of the intentionality of mental states (a major goal of CTM), and that this result undercuts a second major goal of CTM, sometimes refered to as the vindication of intentional psychology. This line of argument is related to the discussions of derived intentionality by Searle (1980, 1983, 1984) and Sayre (1986, 1987). But whereas those discussions seem to be concerned with the causal dependence of familiar sorts of symbolic representation upon meaning-bestowing acts, my claim is rather that there is not one but several notions of meaning to be had, and that the notions that are applicable to symbols are conceptually dependent upon the notion that is applicable to mental states in the fashion that Aristotle refered to as paronymy. That is, an analysis of the notions of meaning applicable to symbols reveals that they contain presuppositions about meaningful mental states, much as Aristotle's analysis of the sense of healthy that is applied to foods reveals that it means conducive to having a healthy body, and hence any attempt to explain mental semantics in terms of the semantics of symbols is doomed to circularity and regress. I shall argue, however, that this does not have the consequence that computationalism is bankrupt as a paradigm for cognitive science, as it is possible to reconstruct CTM in a fashion that avoids these difficulties and makes it a viable research framework for psychology, albeit at the cost of losing its claims to explain intentionality and to vindicate intentional psychology. I have argued elsewhere (Horst, 1996) that local special sciences such as psychology do not require vindication in the form of demonstrating their reducibility to more fundamental theories, and hence failure to make good on these philosophical promises need not compromise the broad range of work in empirical cognitive science motivated by the computer paradigm in ways that do not depend on these problematic treatments of symbols.     

Creating Assemblies in Public Environments: Social Interaction,Interactive Exhibits and CSCW

This paper examines the use of a series of three low tech interactive assemblies that have been exhibited by the authors in a range of fairs, expositions and galleries. The paper does not present novel technical developments, but rather uses the low tech assemblies to help scope out the design space for CSCW in museums and galleries and to investigate the ways in which people collaboratively encounter and explore technological exhibits in museums and galleries. The bulk of the paper focuses on the analysis of the use of one interactive installation that was exhibited at the Sculpture, Objects and Functional Art (SOFA) Exposition in Chicago, USA. The study uses audio–visual recordings of interaction with and around the work to consider how people, in and through their interaction with others, make sense of an assembly of traditional objects and video technologies. The analysis focuses on the organised practices of assembly and how assembling the relationship between different parts of the work is interactionally accomplished. The analysis is then used to develop a series of design sensitivities to inform the development of technological assemblies to engender informal interaction and sociability in museums and galleries.     

Extremal Relations between Additive Loss Functions and the Kolmogorov Complexity

Conditions are presented under which the maximum of the Kolmogorov complexity (algorithmic entropy) K(₁... _N ) is attained, given the cost f( _i ) of a message ₁... _N . Various extremal relations between the message cost and the Kolmogorov complexity are also considered; in particular, the minimization problem for the function f( _i ) – K(₁... _N ) is studied. Here, is a parameter, called the temperature by analogy with thermodynamics. We also study domains of small variation of this function.     

States and the free orthogonality monoid

Let (X, #) be an orthogonality space such that the lattice C(X, #) of closed subsets of (X, #) is orthomodular and let (, ) denote the free orthogonality monoid over (X, #). Let C₀(, ) be the subset of C(, ), consisting of all closures of bounded orthogonal sets. We show that C₀(, ) is a suborthomodular lattice of C(, ) and we provide a necessary and sufficient condition for C₀(, ) to carry a full set of dispersion free states.The work of the second author on this paper was supported by National Science Foundation Grant GP-9005.     

Improving A Solution's Quality Through Parallel Processing

The primary purpose of parallel computation is the fast execution of computational tasks that are too slow to perform sequentially. However, it was shown recently that a second equally important motivation for using parallel computers exists: Within the paradigm of real-time computation, some classes of problems have the property that a solution to a problem in the class computed in parallel is better than the one obtained on a sequential computer. What represents a better solution depends on the problem under consideration. Thus, for optimization problems, better means closer to optimal. Similarly, for numerical problems, a solution is better than another one if it is more accurate. The present paper continues this line of inquiry by exploring another class enjoying the aforementioned property, namely, cryptographic problems in a real-time setting. In this class, better means more secure. A real-time cryptographic problem is presented for which the parallel solution is provably, considerably, and consistently better than a sequential one.It is important to note that the purpose of this paper is not to demonstrate merely that a parallel computer can obtain a better solution to a computational problem than one derived sequentially. The latter is an interesting (and often surprising) observation in its own right, but we wish to go further. It is shown here that the improvement in quality can be arbitrarily high (and certainly superlinear in the number of processors used by the parallel computer). This result is akin to superlinear speedup—a phenomenon itself originally thought to be impossible.