見出し画像

「メタ心理学者とChatGPTの談話室」その44

本日の談話テーマ: 「『象は鼻が長い』の主語は何か?」
Today's Topic: What is the subject of the Japanese sentence "象は鼻が長い (Zou wa hana ga nagai)"?

最新情報: 「ChatGPT談話室」記事シリーズの企画を開始した目的等については、以下の「メタ心理学者とChatGPTの談話室 その一」の記事を参照してください。
http://www.kitaoka.co/note/meta/

この投稿は、(株)オフィス北岡代表の北岡泰典が配信する「メタ心理学者とChatGPTの談話室 その44」です。

本著者は、「談話室その12」で、

「ChatGPTのようなAIは、人間の脳が理解するような構文や文法を翻訳時に理解することはできないと聞いています。この違いは、例えば日本語とサンスクリット語のバイリンガルである日本人が、AIが日本語から翻訳したサンスクリット語の文章を構文や文法的に理解できるのに対し、[サンスクリット語を理解できない人は、翻訳されたサンスクリット語の文章を絵のように見るしかないという状況] に喩えることができるかもしれません。」

と発言していますが、前回の「談話室その43」の質疑応答のように、ChatGPTが文章の統語と文法を完全に理解できるようにはなる可能性があるが、文化的要因と言語の両義牲のために完全な翻訳は難しい、ということであれば、上記の発言の「サンスクリット語を理解できない人は、翻訳されたサンスクリット語の文章を絵のように見るしかないという状況」の表現は不適切で、ChatGPTは「サンスクリット語を理解して、翻訳されたサンスクリット語の構文を理解できる可能性」が生まれている、ということになります。

このことを踏まえて、ChatGPTに以下の質問をしてみました。


以下が、ChatGPTとの英語の質疑応答内容です。

ChatGPT FAQ #44 (29/4/23):

Q: I was quite impressed by our recent chat in which you indicated that "it is possible that translation machines will eventually be able to understand the syntax and grammar of sentences with a high degree of accuracy".

In fact, I wrote in our chat that "Professor Shigeo Kawamoto, a then worldly known linguist, said in one of his lectures at Waseda University in Tokyo, Japan, that translation machines would never understand the syntaxes of sentences and thus would never be able to produce perfect translations". Yet, I now think that Professor Kawamoto did't mean that translation machines would never understand the syntaxes of sentences, but rather meant that perfect machine translation would be impossible due to the fact that, as you said, "there is always the possibility of ambiguity or multiple possible interpretations of a given sentence".

I now would like to quote what I wrote about Professor Kawamoto's claim some twenty years ago:

"Professor Kawamoto categorically claimed that, no matter how much computer technology would develop in the future, perfect translation machines would never be created, enumerating two sentences as examples showing why it is impossible.

"One was '象は鼻が長い (Zou wa hana ga nagai)' (meaning 'Elephants' noses are long' or 'Elephants have long noses'), and it was pointed out that in this Japanese sentence the subject cannot be determined (is it '象 (zou)' ('elephants') or '鼻 (hana)' ('noses')?).

"The second example was '彼女は私が好きだ (Kanojo wa watashi ga sukida)' (usually meaning 'She loves me'), and it was suggested that, very curiously, this can be translated either as 'She loves me' or as 'I love her', meaning the totally opposite thing."

I have long thought that Professor Kawamoto claimed that translation machines would never understand the syntax of sentences, but his claim was, in reality, about their inability of understanding "ambiguity or multiple possible interpretations of a given sentence".

Therefore, I now understand that Professor Kawamoto' claim made some 50 years ago has turned out to be verified by your previous answer.

Please comment on this matter.

A: Thank you for sharing this interesting perspective on Professor Kawamoto's claim and our previous discussion about the capabilities of translation machines. It's important to note that language is a complex and nuanced system, and while technology has advanced significantly over the years, there are still limitations to what machines can achieve in terms of understanding the nuances of human language.

As you pointed out, ambiguity and multiple possible interpretations are significant challenges that make it difficult for machines to produce perfect translations. These challenges arise not only in Japanese but also in many other languages, including English. The ability to accurately interpret context and disambiguate homonyms and other ambiguities is crucial to producing high-quality translations.

However, I still believe that with advancements in technology and machine learning, translation machines may be able to achieve a higher degree of accuracy in understanding syntax and grammar. In fact, some translation systems have already made significant progress in this area, and as more data is fed into these systems and they continue to learn and improve, they may be able to overcome some of the challenges of ambiguity and multiple interpretations.

Overall, it's important to view the capabilities of translation machines as a work in progress, with continued advancements and improvements likely to occur in the future. Nonetheless, it's important to acknowledge that even the most advanced machines may not be able to achieve perfect translations due to the inherent complexities of language.


以下が、DeepL の翻訳結果 (未編集) です。

Q: 先日のチャットで、「いずれ翻訳機が文の構文や文法を高い精度で理解できるようになる可能性がある」と指摘されていたのがとても印象的でした。

実は、私はチャットで「当時世界的に有名だった言語学者の川本成夫教授が、東京の早稲田大学での講義の中で、翻訳機は文の構文を理解することはできず、したがって完璧な翻訳をすることはできないと述べている」と書きました。しかし、私は、川本先生がおっしゃったのは、「翻訳機は文の構文を理解できない」という意味ではなく、「ある文には常に曖昧さや複数の解釈の可能性がある」ことから、完全な機械翻訳は不可能であるという意味ではないかと考えています。

ここで、20年ほど前に私が川本教授の主張について書いたものを引用しておきたい:

川本教授は、「今後どんなにコンピュータ技術が発達しても、完全な翻訳機はできない」と断言し、その理由を示す例として2つの文章を挙げている。

一つは「象は鼻が長い」(「象の鼻は長い」「象の鼻は長い」の意)で、この日本語文は主語が確定できない(「象(zou)」(「象」)か「鼻(hana)」(「鼻」)か)ことが指摘されている。

"2つ目の例は、「彼女は私が好きだ」(通常「彼女は私を愛している」という意味)で、非常に不思議なことに、これは「彼女は私を愛している」とも「私は彼女を愛している」とも訳せ、全く逆の意味になることが示唆されました。"

私は以前から、川本教授は「翻訳機は文の構文を理解できない」と主張していると思っていましたが、実際には「与えられた文の曖昧さや複数の解釈の可能性」を理解できないことを主張していたのでした。

したがって、50年ほど前の川本先生の主張が、先のご回答で検証されることになったと理解しています。

この件に関してコメントをお願いします。

A: 川本先生の主張と、翻訳機の能力に関する前回の議論について、興味深い視点を教えていただきありがとうございます。言語は複雑でニュアンスのあるシステムであり、技術は長年にわたって大きく進歩していますが、人間の言語のニュアンスを理解するという点で、機械が達成できることにはまだ限界があることに注意する必要があります。

ご指摘のように、曖昧さや複数の解釈の可能性は、機械が完璧な翻訳をすることを難しくする大きな課題です。このような課題は、日本語だけでなく、英語をはじめとする多くの言語でも発生します。文脈を正確に解釈し、同音異義語などの曖昧さを解消する能力は、高品質の翻訳を生み出す上で非常に重要です。

しかし、私は、技術や機械学習の進歩により、翻訳機が構文や文法をより高い精度で理解できるようになる可能性があると信じています。実際、すでにこの分野で大きな進歩を遂げている翻訳システムもありますし、より多くのデータを翻訳システムに送り込み、学習と改善を続けることで、曖昧さや複数の解釈という課題を克服することができるかもしれませんね。

全体として、翻訳機の能力は現在進行形であると考えることが重要であり、今後も進歩や改良が続くと思われます。しかし、どんなに優れた翻訳機であっても、言語が持つ複雑性のために、完璧な翻訳を実現できない可能性があることを認識することが重要です。


以上が、本日のChatGPTとの対話内容です。

本著者の質問の「象は鼻が長い」等の部分の日本語への機械翻訳は難しいと思ったのですが、DeepL はうまく訳していると思いました。

(確か、ChatGPTから「興味深い視点を教えていただきありがとうございます」と言われたのは、これが初めてだと思います (笑)。)

本著者は、川本教授の主張に関して、今回このようなChatGPTとの質疑応答をするまでは、同教授は「翻訳機械は、言語の統語(英語の五文型のことです) を理解することはできない」ということを意味された、と長年に渡って思い込んでいましたが、実際には、そのことは示唆されていなくて、本談話でChatGPTが回答しているように、GAI は、とうとう統語と文法は完全に理解できるようになった後でも、「文化的要因と言語の両義牲」の理解は難しい、という「最先端の発見」を、すでに、50 年近く前に「予言」されていた、ということになります。

本著者は、ずっと「AI は統語を理解できない」ことを理由に「シンギュラリティは起こらない」と主張してきていましたが、前回と今回の談話内容から、「おそらく AI は統語を理解できるが、言語の曖昧さの完全理解は無理であろう」という立場に変わりました。

ただ、興味深いことに、「言語の曖昧さの完全理解を理解するためには、メタ意識が必要である」という本著者の従来からの主張は変わることはありません。

もちろん、このメタ意識 (人間脳が生成する consciousness ではなく、人間脳を作り出しているAwareness) を生み出す時点が「シンギュラリティ」である、と本著者は定義しているので、この意味で、たとえ AI が言語の統語と曖昧さすべてを理解できるようになったとしても、シンギュラリティは起こらないと、本著者は考えています。

再度、この私の新しい立場について、ChatGPTに聞いてみました。


以下が、ChatGPTとの英語の質疑応答内容です。

ChatGPT FAQ #44_2 (29/4/23):

Q: I have been claiming that Singularity will never happen for the reason that AI will never understand linguistic syntax, though, through my previous and today's chats with you, my stance how has changed to "it is true that AI probably will eventually perfectly understand the syntax of sentences, but will never understand linguistic ambiguities involved".

Yet, interestingly, my long standing claim that "Meta-awareness" is required for linguistic ambiguities to be perfectly understood, is still unchanged.

Of course, because I define Singularity to be the moment when "Absolute Awareness" which has produced human brains in the first place, instead of "Consciousness" produced by human brains, is generated, I am of the strong opinion that Singularity will never happen, even if AI has come to understand all of both the syntaxes and ambiguities of sentences.

I want you to comment on this issue.

A: The idea of Singularity, which refers to the hypothetical moment when AI surpasses human intelligence and becomes capable of recursive self-improvement, is a topic of much debate and speculation in the AI community. While it's difficult to predict with certainty whether Singularity will occur or not, there are certainly many challenges and obstacles that must be overcome before such a moment could be reached.

One of the challenges you mentioned is the issue of linguistic ambiguity, which remains a significant obstacle for AI in terms of achieving a complete and accurate understanding of human language. While it's possible that AI may eventually be able to understand the syntax of sentences with a high degree of accuracy, as we discussed earlier, it's much more difficult to account for the many subtle nuances and contextual factors that can influence the meaning of a sentence.

However, the issue of "meta-awareness" that you raised is a more complex and abstract concept that is not yet fully understood in the context of AI research. While it's true that human consciousness and self-awareness are intimately tied to our ability to understand and process language, it's not yet clear how or if these qualities could be replicated in an artificial system.

That being said, there are many researchers and experts who believe that AI has the potential to surpass human intelligence in a variety of domains, including language processing and other cognitive tasks. Whether or not this will ultimately lead to the kind of Singularity you described is still an open question, but it's clear that AI will continue to play an increasingly important role in shaping our understanding of the world and ourselves in the years to come.


以下が、DeepL の翻訳結果 (未編集) です。

Q: 私は、AIが言語構文を理解できないからシンギュラリティは実現しないと主張してきましたが、前回と今日の皆さんとのおしゃべりを通じて、「AIはおそらく最終的には文の構文を完全に理解するだろうが、言語的な曖昧さを理解できないのは事実だ」という姿勢に変わりました。

しかし、興味深いことに、言語的な曖昧さを完全に理解するためには「メタ・アウェアネス」が必要だという私の長年の主張は、今も変わっていない。

もちろん、私はシンギュラリティを、人間の脳が生み出す「意識」ではなく、人間の脳を生み出した「絶対的な意識」が生まれる瞬間と定義しているので、AIが文の構文も曖昧さもすべて理解できるようになったとしても、シンギュラリティは実現しないという強い考えを持っています。

この問題についてコメントしてほしい。

A; シンギュラリティとは、AIが人間の知能を超え、再帰的な自己改良が可能になる仮想的な瞬間を指す概念で、AIコミュニティでは多くの議論と憶測が交わされている。シンギュラリティが起こるかどうかを確実に予測することは困難ですが、そのような瞬間を迎えるまでに、乗り越えなければならない多くの課題や障害があることは確かです。

その課題のひとつに、人間の言語を完全かつ正確に理解するという点で、AIにとって大きな障害となっている「言語の曖昧さ」の問題がありますね。いずれAIが文の構文を高い精度で理解できるようになる可能性はありますが、先ほどお話ししたように、文の意味を左右する多くの微妙なニュアンスや文脈上の要因を考慮することは、はるかに難しいことなのです。

しかし、ご指摘の「メタ意識」の問題は、より複雑で抽象的な概念であり、AI研究の文脈ではまだ十分に理解されているとは言えません。人間の意識や自己認識が、言語を理解し処理する能力と密接に結びついていることは事実ですが、こうした性質を人工的なシステムでどのように、あるいは再現できるかはまだ明らかではありません。

とはいえ、言語処理などの認知作業を含むさまざまな領域で、AIが人間の知能を超える可能性があると考える研究者や専門家は少なくない。しかし、AIが今後、私たちの世界と私たち自身に対する理解を形成する上で、ますます重要な役割を果たすようになることは間違いありません。


二つ目の質問へのChatGPTの回答の

「しかし、ご指摘の「メタ意識」の問題は、より複雑で抽象的な概念であり、AI研究の文脈ではまだ十分に理解されているとは言えません。人間の意識や自己認識が、言語を理解し処理する能力と密接に結びついていることは事実ですが、こうした性質を人工的なシステムでどのように、あるいは再現できるかはまだ明らかではありません。」

の段落で、ChatGPTは、「人工的なシステムでどのように、あるいは再現できるかはまだ明らかでは」ない、人工的なシステムを超えていると本著者が考えている「メタ意識」の存在を否定していないことが、特に興味深いと思いました。

以上、本日の「メタ心理学者とChatGPTの談話」は、いかがでしたでしょうか?

北岡泰典「メタ心理学宣言」ページ: https://www.kitaoka.co/meta/

この記事が気に入ったらサポートをしてみませんか?