Xref: utzoo sci.lang:5266 comp.ai:4799 Path: utzoo!attcan!uunet!ginosko!brutus.cs.uiuc.edu!psuvax1!rutgers!eddie!uw-beaver!fluke!ssc-vax!bcsaic!rwojcik From: rwojcik@bcsaic.UUCP (Rick Wojcik) Newsgroups: sci.lang,comp.ai Subject: Re: What's the Chinese room problem? Message-ID: <15336@bcsaic.UUCP> Date: 29 Sep 89 01:03:35 GMT References: <822kimj@yvax.byu.edu> Reply-To: rwojcik@bcsaic.UUCP (Rick Wojcik) Organization: Boeing Computer Services AI Center, Seattle Lines: 76 Celso Alvarez (CA) writes: me>. . . The trick to translation is to construct expressions in the me>target language that evoke the same thoughts as those in the source me>language. CA> Much more than thoughts are evoked by language. How do you translate CA> the signalling of identity, roles, and social relationships? I think that such concepts have to be represented as thought structures, since they have an impact on language structure. But your question may be filed under my general question: Just what do 'Chinese Room' debaters think a translation is? What criteria do you use to judge that a translation from one language to another is successful? My position is that there is no such thing as translation in an absolute sense. A seemingly trivial example is the translation of expressions that refer to language-specific grammatical structure. Thus, there is no way to translate French 'tutoyer' directly into English. You must rely on circumlocution. It means roughly 'use the intimate 2nd person singular form of the verb'. But practical translators might take an equivalent French expression to 'Don't tutoyer me' into English as 'Don't use that tone of voice with me', or some such thing. But it is difficult to say what makes one such translation better than another. People can get into heated arguments over such questions. N. Boubaki (NB) writes: >...Those who deal with real >language to language translation know that there is no one-to-one match >between expressions in one language and those in another. NB> But this difficulty would affect the native Chinese speaker and the NB> Chinese Room Demon equally. That is one premise of Searle's NB> argument - the "mechanical" system is presumed to be just as competent NB> (not necesarily perfect) at translation as the "understanding" system. I know, but I think that Searle, like most of us, has implicitly adopted the conduit metaphor in his conceptualization of the problem. He really believes that there is some absolute sense whereby an expression in one language corresponds to one in another. This seems clear from his insistence that the translation itself be 'mechanical'--in other words, symbol manipulation. Those involved in translation understand that the translation process requires editing and revision. Who determines that the "mechanical" system is "just as competent" if there is no mechanical basis for judging competence? But that is just what you need to do in order to bring about translation. You need mechanize the ability to judge and revise. That would be tantamount to mechanizing the understanding process, since it is only by understanding expressions in two different languages that you can judge their equivalence. I want to be careful to distinguish modern Machine Translation efforts, which do not attempt to automate the revision process (rather they attempt to facilitate it), from an ideal MT system, which would require mechanized understanding to do its job properly. So I agree with you that Searle ultimately begs the question. The question is whether or not 'understanding' is a mechanizable process. He either assumes that it is not, or he doesn't have a proper conception of the nature of translation. Ray Allis (RA) writes: RA>It seems to me your position is in fact very close to Searle's. The problem RA>I have with his little parable is that he pretends that the output from RA>the Chinese room is satisfactory (or rather lets us assume so). I believe RA>that if the room does not "understand" Chinese, and he argues that it does RA>not, the output will not be satisfactory... From my above remarks, you should see that I am closer to your viewpoint than Searle's. In fact, I find myself largely in agreement with most of what you said. I would only quibble on the issue of whether or not modern NLP efforts, including MT, are futile. The pragmatic purpose of such work is to increase human efficiency in language-intensive work on computers. There are many good things you can do without addressing the need for full language understanding. MT (really Machine-Assisted Translation) can improve the output of a human translator, even though the MAT system may produce some pretty bad translations. Our grammar-checking system is proving useful in the writing of aircraft maintenance manuals. But this takes us away from the philosophical question of whether or not you can mechanize language understanding. -- Rick Wojcik csnet: rwojcik@atc.boeing.com uucp: uw-beaver!bcsaic!rwojcik