计算机能思考吗？图1专题7：“计算机能否做类比？”

Can Computers Think? The History and Status of the Debate - Map 1 of 7

问题7

Issue Area: Can computers draw analogies?

问题域： 计算机能否做类比？

The link to the part of the map this discussion is about: http://www.macrovu.com/CCTWeb/CCT1/CCTMap1Analogies.html

1. Alan Turing, 1950, Yes, machines can (or will be able to) think. A computational system can possess all important elements of human thinking or understanding.
I believe that at the end of the century ... one will be able to speak of machines thinking without expecting to be contradicted.

公共起点

66. (Disputing 1) Computers can't understand analogies. Computers cannot understand analogical comparisons or metaphors. For example, a machine could not understand the sentence. "She ran the like the wind."

Note: Analogy arguments are also discussed by George Lakeoff in the "Symbolic Data" arguments on Map 3.

（反驳1） 计算机不能理解比喻。计算机不能理解类比性比较或隐喻。例如，一个极其不能理解这样的句子“她像风一样奔跑”。

注：比喻也在图三的“符号数据”讨论区由George Lake提出并讨论。

67. (Disputing 66) Computers have understood analogy. Existing models have discovered and understood analogies.

（反驳66）计算机已经理解了比喻。有现有的模型已经做到了发现和理解类比。

68. (Supporting 67) Brian Falkenhaimer, K. Forbus, and D. Gentner, 1990 (Implemented Model) SME. SME is a structure-mapping engine that discovers analogies between domains by a set of match rules. The analogies that result are judged according to the criteria of clarity, richness, abstractness, and systematicity. SME has found mappings between heat and water flow, solar systems and atoms, and in other domains.

（支持67）Brian Falkenhaimer, K. Forbus, 和D. Gentner，1990 （已实现模型）SME. SME是一个结构映射引擎，它通过一组规则发现讨论域之间的类比关系。结果产生的类比再被根据明晰性、丰富性、抽象性和系统性准则进行评判。SME已经发现了如热流和水流，太阳系和原子结构以及一些其他的映射关系。

69. (Disputing 68) David Chambers, Robert French, and Douglas Hofstadter, 1995SME only draws analogies from prestructured representations. SME creates analogies using high-level representations that are structured with those specific analogies in mind. Its behavior provides no evidence of intelligence because the analogies it discovers are already built into the data it works with.
Supported by "The Front-End Assumption Is Dubious," Box. 74

（反驳68） David Chambers，Robert French，和Douglas Hofstadter，1995 SME只能对预先结构化好的表述进行类比。SME只是通过使用高层次的已经用在人的头脑中描画好的类比结构化了的表述来进行类比。其行为并不能为智慧的存在提供证明，因为它所发现的比喻已经被建立在它使用的数据中了。

70. (Disputing 68) David Chalmers, Robert French, and Douglas Hofstadter, 1995Objects, attributes, and relations are too rigidly distinguished by SME. In order for its analogical mappings to work, SME assumes a rigid distinction between objects, attributes, and relations. But it is unclear whether humans make such a rigid distinction (most likely not -- translator). For example, we sometimes conceptualize wealth as an object that flows between people, but at other time we conceptualize wealth as an attribute that changes with each transaction we make. (This is one of the attributes of human that are hard for computers to achieve. It's to do with social experience that develops over time -- translator)

（反驳68）David Chambers，Robert French，和Douglas Hofstadter，1995 SME对物体、属性和关系的辨识过于僵硬。为了使其类比映射能工作，SME假设了一个物体、属性和关系间的硬性区分。但不清楚的是人类是否做如此鲜明的区分（很可能不——编者）。例如，我们有时将财富想象成在人与人之间流动的东西，不过有时又将其看作一种随着我们所作的每一笔具体交易而变化的属性。

71. (Disputing 68) David Chalmers, Robert French, and Douglas Hofstadter, 1995SME's treatment of relations is too rigid. In SME, relations are treated as n-place predicates that can only be mapped to other n-place predicates. For example, attraction is a 2-place predicate that could be represented as "attracts (sun, planet)" and then mapped to "attracts (neucleus, electron)." But it is unlikely that the human mind is so rigid in its treatment of relational mappings.

（反驳68）David Chambers，Robert French，和Douglas Hofstadter，1995 SME对关系的处理过于僵硬。在SME中，关系被处理成n元谓词形式，并只能被映射到其他的n元谓词关系。例如，“吸引力”是一个二元关系谓词，它能被表示成关系式“吸引（恒星，行星）”，并被映射到例如“吸引（原子核，电子）”。不过人类的思维是如此处理关系关联的可能性很小。

72. (Supporting 67) Keith Holyoak and Paul Thagard, 1989 (Implemented Model) ACME.ACME is a connectionist network that discovers cross domain analogical mappings. The ACME network uses structural, semantic, and pragmatic constraints to seek out those mappings.

（支持67）Keith Holyoak and Paul Thagard，1989 （已实现模型）ACME。ACME是一个单元连接网络，它能发现跨区域的类比映射。ACME网络使用结构化，语义且使用的限制条件来找出这些映射。

73. (Disputing 72) David Chalmers, Robert French, and Douglas Hofstadter, 1995ACME doesn't understand analogy. ACME's claim to understand analogies is overblown. All ACME does is take algebraic sentences in predicate logic notation and compare them. For example, it only understands that "Socrates is like a midwife" to the extent that it understands that "(a(b)),(c(d)) ... is similar to (A(B)),(C(D))."

"That ( (a(b)),(c(d)) ... is similar to (A(B)),(C(D)) ) is not the same as understanding that Socrate is like a midwife.

（反驳72） David Chalmers，Robert French，and Douglas Hofstadter，1995 ACME不能理解类比。ACME对类比的理解的说法被夸大了。ACME所作的一切只是将谓词逻辑概念中的代数语句取出并进行比较。例如，它只能以理解“(a(b)),(c(d)) ... 类似于(A(B)),(C(D))”这样的程度理解“苏格拉底就像是接生婆”这样的概念。（但很显然，对于人类，这两着是完全不同的）

74. (Supporting 73) David Chalmers, Robert French, and Douglas Hofstadter, 1995The front-end assumption is dubious. Models that use preconfigured representation and hand-tailored data assume that a separate front-end module could be built that would filter sensory data into the model's representational form.

（支持73）David Chalmers，Robert French，and Douglas Hofstadter，1995 前端假设是不可靠的。使用预先配置的表述和手工裁剪的数据的模型假设了一个能为之过滤数据并构造成这个单元所期望的表述的前侧模块的存在。

75. (Supporting 74) David Chalmers, Robert French, and Douglas Hofstadter, 1995All-encompassing representations could not be processed. The all-purpose representation that a front-end module would provide to a computer model would have to encode a vast amount of information, enough for it to adapt to all the various contexts and analogies it might be used in. Such a representation would be too bulky (no less than what nowadays google takes hold of plus a full digestion as educated humans do which google is far from being capable of -- translator) for efficient processing. (Criticizing an existing implemented model is almost always much easier than building it, however, arguments such as this are very prudent and constructive -- translator)

（支持74）David Chalmers，Robert French，and Douglas Hofstadter，1995 全包含的表述难以处理。前侧模块应该提供给类比模型的全用途的表述需要包含大量的信息，以足够让它适应所有不同的语境和类比要求。如此的表述将会无比庞大，因而无法有效地实现。

76. (Supporting 74) David Chalmers, Robert French, and Douglas Hofstadter, 1995Perception depends on analogy. How we see things depends in part on what high-level analogical processes we use. For example, Saddam Hussein will be perceived quite differently depending on whether he is viewed as analogous to Adolf Hitler (a ruthless aggressor) or to Robin Hood (a generous crusador).

（支持74）David Chalmers，Robert French，and Douglas Hofstadter，1995 观察依赖于类比。我们如何看待事物部分依赖于我们使用的高级别的类比过程。例如，依据是将其看成阿道夫希特勒还是罗宾汉，我们对萨达姆侯赛因的看法会有不同。

77. (Supporting 67) Douglas Hofstandter and Melanie Mitchell, 1995 (Implemented Model) COPYCAT. COPYCAT is a model that discovers analogies using 3 components: (1) a "slipnet" of abstract Platonic concepts whose relations can change as the model runs, (2) a "workspace" of perceptual activity that acts like a short-term memory, and (3) a "coderack" of agents that are probabilistically selected to carry out tasks in the workspace. COPYCAT is neither a symbol manipulator nor a connectionist network, though it draws on both paradigms. Representations are not delivered hand-tailored to the model, but are built up through fluid interactions between low-level and high-level components.

（支持67） Douglas Hofstandter和Melanie Mitchell，1995 （已实现模型）模仿者（COPYCAT）。模仿者是一个使用三个组件来发现类比的模型：(1) 抽象的柏拉图式观点的集合，它能随着模型运行而改变；(2) 一个类似于短时存储的观察活动的“工作区间”，以及(3) 一组可随机随机选择在工作区间执行任务的工作者。模仿者既不是符号操作器，也不是单元连接网络，尽管它受这两者的启发。它所使用的表述并非手工裁剪后馈送给它的，而是由其中低级别和高级别组件之间的流动交互产生。