И.С. Гудилина, Л.Б. Саратовская, Л.Ф. Спиридонова - English Reader in Computer Science (1114139), страница 9

Просмтор этого файла доступен только зарегистрированным пользователям. Но у нас супер быстрая регистрация: достаточно только электронной почты!

Текст из файла (страница 9)

The author's opinion is that machine translation is too complex for the current statistical processing methods to handle, even though these methods do not aspire to building representational models of human language capacity and rely only on the input -output behavior of such models (in machine translation, a text and its translation). He also thinks that the rule-based/corpus-based dichotomy is not as important as we think. He proposes that the real problem of machine translation as technology is that artificial intelligence researches do not generally understand how difficult the problem actually is.

I think that Machine Translators can be helpful to a person doing translations from one language into another in spite of that to obtain high-quality, fully automatic translation remains an elusive goal for the present.

While using M.T. there arise some difficulties because of that it commits quite a lot of errors. As it is known many words are polycemantic and M.T. can choose only one meaning of the word which often doesn't correspond to that given in the text. The Russian language is rather complicated grammatically; computer can not use, for example, the six Russian cases in a proper way. When translating compound or complex sentences with participal constructions it doesn't manage its job well either.

But it translates simple sentences and monosemantic words rather well. I often use M.T. as a glossary, so I don't need spend much time for searching a word in a dictionary. What do you think of that?

Unit 3

Modeling Reality

Why don't we have a complete plan for reforms? In order to play chess, one must know the rules... how to move the various pieces on the board. But it is not possible to know the situation on the chessboard after the 15^th or 25^th move.

— Vaclav Klaus, Finance Minister,Czechoslovakia

Scientists, managers, executives and government leaders are expressing increasing concern about the safety and reliability of complex computer systems. As such systems take charge of everything from phone calls to flights, we are exposed to a growing danger of man-made disasters.

Important among complex computer systems are computer models used for simulations and predictions of phenomena in areas ranging from physics to hardware engineering to socio-economic systems. Computer models have become an area of concern unto themselves. Their misuse could lead governments to adopt disastrous policies in dealing with such subjects as global warming and global economic stability. The proliferation of computer models supporting divergent points of view—for example, computer simulations supporting conflicting theories of global warming or nuclear winter—can easily mislead the lay public. Models whose results depend on assumptions about human behaviour are the most likely to produce controversial results.

In early November 1990 the Association for Computing Machinery (ACM) brought together leading scientists, business executives and government officials to discuss public-policy questions surrounding computer models. I will summarize the main points about modeling made by the principal speakers at that meeting.

Modeling Expertise

Knowledge-based systems (KBSs) are important examples of computer models. A KBS is because to many observers the majority of them have fallen short of the promise of competent performance.

John Kunz attributes part of the problem to a design-and-testing process taken from software engineering, a process that begins with a formal specification of the system's behavior and ends with an acceptance test. This process cannot take into account that the standards for expert performance can shift as a field changes. Kunz argues that, to obtain reliable KBSs, continual testing and improvement must be the standard approach. The tests must do more than compare KBS decisions with real situations; they must validate that at all times the recommended actions fulfill the purpose of the system, that the reasoning procedures are valid for the domain, and that the recommended actions are consistently endorsed and assessed as competent by human experts. Kunz recommends that the tests include simple realistic cases as well as cases that apply various stresses to the KBS. He recommends that some of the tests be retrospective (comparing KBS decisions with those of experts in the past) and that some be prospective (testing the KBS against experts in real time).

KBSs are founded on the assumption that an expert works from a complete theory of the domain. Once a theory is articulated as a set of rules and stored in a database, the superior power of the computer can draw inferences much faster than the expert. That this has not been accomplished cannot be blamed on a lack of computing power, memory, research effort or cooperation of experts. An explanation gaming credence is that experts themselves do not work from complete theories, and much of their expertise cannot be articulated in language. The advocates of neural networks claim they have found a way to overcome the inability to articulate expertise. Neural networks mimic the biological structure of the brain and therefore the expert's approach to gathering and organizing information; once the networks have been trained, their advocates say, they will be able to acquire the knowledge experts have that cannot be articulated as rules.

At the ACM meeting Jay Forrester argued that all human decisions are taken with respect to (possibly subconscious) mental models and that computers should be used to augment human mental-modeling powers. He is interested in models that make predictions about the future behaviour of large organizations and societies. He maintains that human beings are notoriously inept at understanding the dynamics of systems that contain feedback control loops. Feedback loops, which are familiar features of mechanical systems and biological organisms, also permeate organizations and social systems. The modeling approach Forrester calls system dynamics is aimed at giving us a tool to aid in understanding the operation of systems for which we have only a static description. He claims that many organizations can be successfully modeled because the members of the organization follow policies that are either explicit or are part of their habitual behaviour; hence they can be stated as precise static rules that can be embodied as interacting functions in the model. Forrester has a good deal of optimism that socio-economic systems can ultimately be modeled and that system dynamics is a powerful general approach.

Limits of Modeling

Stuart Dreyfus, a long-time advocate of modeling and critic of expert systems, is concerned that we understand the limits of modeling so that our claims about models can be well grounded. He argues that in most socio-economic domains, neither conventional mathematical modeling (including rule-based artificial intelligence) nor neural-network modeling is as trustworthy as the judgments of impartial, experienced experts. He calls the actions of experts in a domain a form of "skillful coping," about which there are four extant theories: 1. Expert behavior is an unconscious application of a conventional model. 2. It is uninterpretable neuronal and biochemical activity. 3. It is a process of recalling memories that match the current situation. 4. It is uninterpretable brain activity evolved from a domain theory learned during an initial formal encounter with the domain through some teacher.

Dreyfus says that Forrester bases system dynamics on the first theory, whereas Dreyfus himself finds the fourth theory much more credible and consistent with evidence about skillful coping. He concludes that computers that provide facts and suggest decisions can improve the judgment of experienced people. In the hands of inexperienced people, however, such computers may actually degrade coping skill. Education that equates expertise with models can inhibit the development of good judgment.

Steve Kline draws a sharp distinction between physical systems and systems that include human beings. He uses a simple complexity index to demonstrate the qualitative differences between these two kinds of systems. His measure counts the number of variables, parameters and feedback loops in the system being modeled. Physical systems modeled by differential equations (e.g., fluid flows) have low model complexity (on the order of 10') and may have high computational complexity- Hardware systems (e.g., airplanes and computer networks) have moderate model complexity (on the order of 10) and moderate to high computational complexity. But models for "human systems" — brains, personalities, organizations, economies and societies — all have extremely high model complexity (on the order of 10¹³ and beyond).

In Kline's analysis physical systems and hardware systems have three characteristics that lead to low model complexity: they operate under invariant rules, their parts are context-independent, and they are not self-observing. In contrast, human systems have changing rules and are context-dependent and self-observing. Kline ends up questioning the "science-based" approach to modeling these systems, an approach rooted in the Newtonian (mechanistic) tradition, which assumes that all of the universe is governed by fixed laws.

Eleanor Wynn continues the skepticism toward computer models of human activities by questioning whether the perspective of information processing itself is sufficient to understand human systems. Noting the widespread agreement that we do not know how to design complex software systems that are dependable, she observes that most of the discussion about software occurs within the paradigm of software engineering that begins with a formal specification and ends with an acceptance test. She argues that this paradigm completely misses how good designs are made because it is context-independent and cannot take into account the perspectives of users.

Description, Computation, Prediction

These authors share the conclusions that models involving human behavior are unavoidably complex, that such models may not work except in limited cases, and that even then they will be made to work by ongoing development rather than by prior analysis. They suggest that one's trust in the reliability of such models depends on one's assumptions about how biological organisms and societies learn and act. But they diverge on this claim. Models can produce greater understanding of complex human phenomena, lead us to wise decisions and guide us to effective actions. Forrester is optimistic about this claim. Kunz implicitly accepts it in the domain of knowledge-based systems. But Dreyfus, Wynn and Kline express serious doubts. The divergence of views on this question is at the heart of the debate over computer modeling of human realities.

In what follows I offer my own analysis of this claim, and I suggest ways that computers can assist us effectively in the domain of human actions.

What is a model? We usually understand a model to be a symbolic representation of a set of objects, their relationships and their allowable motions. We use models in three

Description. We sometimes use a model to describe how a system works. Examples are a blueprint, a scale model of a railroad, the equations of motion of a planet, the scientific method, and the software-design process.

Computation. We sometimes use a model to guide, to reproduce or to calculate action in the domain. Examples are following directions from an inertial guidance system (guiding), a flight simulator (reproducing) or computing a measurement (calculating).

Prediction. We sometimes use a model to predict the future state of a system with tolerable certainty. Examples are models that predict the lift of a wing in flight, the position of a star or me future state of the weather or the world economy. A model is useful for prediction only if the future state can be calculated much more rapidly than in real time and only if its users agree that the assumptions about parameter values and governing laws will hold at the future time.

These three aspects are hierarchical in the sense that prediction relies on a model to compute a future state given future values of parameters, and computation relies on a precise description of the allowable motions of a system. Models are of universal interest because of our unavoidable concern to anticipate and prepare for future action, and because they make the world seem simpler and more understandable.

I. Answer the following questions:

1. Why have computer models excited great concern among scientists, managers and government leaders in the recent years?

2. What is a computer model?

3. What are computer models used for? Give some examples.

4. Computer models are of universal interest, aren't they? Why? Will you give any examples of computer models?

5. What is important for obtaining reliable KBSs?

6. What kinds of tests are recommended for analyzing simple realistic situations?

7. The advocates of neural networks claim that they have found a way to overcome the inability to articulate expertise. Do you mink they are right or wrong by saying this? Argue your point of view.

8. What distinction is there between physical systems and models for "human systems"?

II. Write a summary in English.

Stage A

1. Look through the text (scheming reading).

2. Divide it into introduction, principle part, and conclusion.

3. Find and write down the main idea(s) of the text.

Stage В

1. Read the text again but now attentively (close reading).

2. Give detailed answers to the questions.

3. Write the items of the plan.

Stage С

1. Write out kernel sentences to illustrate the items of the plan.

2. Join kernel sentences together; use connective words if necessary.

3. Re-read your summary and make sure that the sentences are presented in a logical order.

Make any changes that you think are necessary.

III. Read the text. Agree or disagree with what is told in the text. Write about your attitude to the problem. Give arguments in favour of your point of view.

However there is a sharp distinction between physical systems and systems that include a human being. Models for "human systems" — brains, personalities, organizations, economies and societies — all have extremely high model complexity. The complexity measure counts the number of variables and parameters in the system being modeled. Physical systems and hardware systems have three characteristics that lead to low model complexity: they operate under invariant rules, their parts are context-independent, and they are not self-observing. In contrast, human systems have changing rules and are context-dependent and self-observing.

Характеристики

Список файлов книги