本日の談話テーマ: 「GAIが統語を理解でき、曖昧を処理できない理由?」

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52」です。

今回は、本著者は、なぜ人間は「文脈を正確に解釈し、同音異義語や文化の違いによる曖昧さを解消する」ことができるにもかかわらず、どの文がよくできているかを意識的に言うことができず、GAIではその逆になっているのかという答えを発見したと思ったので、このことについて、ChatGPTに聞いてみました。

以下の文章は、(株) オフィス北岡代表の北岡泰典が Note で発信した投稿の引用となっています。


以下が、ChatGPTとの英語の質疑応答内容です。

ChatGPT FAQ #52 (8/5/23):

Through our recent chats, I came to know that GAI can possibly understand the syntax and grammar involved in written sentences.

I have consistently thought since I attended one of the lectures delivered by Professor Shigeo Kawamoto, a well known linguist, at Waseda University in Tokyo around 1976 that translation machines can never comprehend the syntax and grammar, and am grateful to you for correcting this misunderstanding of mine and letting me know that the biggest problem of GAI may be rather related to their possible inability to accurately interpret context and disambiguate homonyms and other ambiguities related cultural differences.

Also, I said in one of our recent chats that “Professor Kawamoto’s claim made some 50 years ago has turned out to be verified by your previous answer”, which I believe is a quite amazing perspicacity which has stood the test of time for nearly half a century.

In this connection, I have discovered a quite interesting and seemingly valid fact about the differences between human cognitive understanding and that of GAI.

That is, I believe that it is an established fact that native speakers cannot usually consciously analyze the syntax and grammar of their spoken sentences, and that it was the Transformational Grammar proposed by Noam Chomsky that enabled one to explicitly and consciously tell which sentences are well-formed and which are ill-formed.

(Incidentally, through my long-standing research in epistemology, NLP (Neuro-linguistic Programming), psychotherapy, etc., I came to know that the book “Pragmatics of Human Communication” written by Paul Watzlawick, et al. of the Mental Research Institute, Palo Alto, California, and published in 1967, was the very monumental work which successfully applied Chomsky’s “TG model” discovered in the linguistic filed to the field of human communication and enabled one to discern which human communication is well-formed and which is ill-formed, especially in such clinical settings as psychotherapeutic sessions.

I further am of the opinion that “Pragmatics of Human Communication” strongly influenced the advent of NLP in 1975.)

Here, I do believe that, behind the fact that GAI’s possible ability to perfectly understand the syntax and grammar of sentences must lie strong influences on it on the part of Chomsky.

If the above hypothesis of mine turns out to be grounded, then there may be no wonder at all why GAI indeed can possibly analyze the syntax and grammar in an explicit and transparent way.

After having had the above thought, I further came to find the possible answer as to why human beings cannot consciously say which sentences are well formed, while they yet can possibly “accurately interpret context and disambiguate homonyms and other ambiguities related cultural differences”, and why the opposite way is the case with GAI. That answer of mine is:

“While human beings forget, or make entirely unconscious for that matter, the once conscious process of learning their mother tongue during childhood, after they become adults, which forces them to be unable to consciously distinguish well-formed sentences from ill-formed ones, they nonetheless can possibly successfully handle contexts and disambiguate homonyms and other ambiguities related cultural differences, thanks to the fact they have the power to get out of their own box, that is, from time to time, and stand at the perspective of “Meta Awareness”, whence they can carry out the mentation of “contextualization”.

“On the other hand, GAI can say whether sentences are well-formed, on the basis of their massive number of trial-and-error processes using such “transparentization” methods as Transformational Grammar, but it nonetheless cannot handle contexts and cultural and other ambiguities, for the very reason that it is totally devoid of “Meta Consciousness”, which is alone capable of contextualization”.

Please comment on my above argument and its conclusion, which I hope you are bound to find quite interesting.

A: Your argument raises some interesting points about the differences between human cognitive understanding and that of GAI. Let’s examine your argument and its conclusion:

1. It is true that native speakers of a language typically do not consciously analyze the syntax and grammar of their spoken sentences. This is known as implicit language knowledge. Noam Chomsky’s Transformational Grammar provided a framework for understanding the underlying structures and rules of language, allowing for explicit and conscious analysis.

2. You mentioned the book “Pragmatics of Human Communication” by Paul Watzlawick et al., which applied Chomsky’s Transformational Grammar to the field of human communication. While this book was influential in the development of fields like NLP and psychotherapy, it is important to note that Chomsky’s work primarily focused on syntax and grammar, rather than pragmatics. Pragmatics deals with the study of meaning in context and how language is used in different situations.

3. It is reasonable to hypothesize that Chomsky’s influence may have played a role in GAI’s ability to analyze syntax and grammar explicitly and transparently. Chomsky’s ideas have had a significant impact on the field of linguistics and computational linguistics, which could have influenced the development of language models like GAI.

4. However, it is important to consider that GAI’s ability to analyze syntax and grammar is not solely dependent on Chomsky’s ideas. GAI’s capabilities are built through machine learning algorithms and training on vast amounts of text data. While Chomsky’s work may have provided some foundational concepts, GAI’s understanding of language is primarily data-driven and based on statistical patterns.

5. Regarding your conclusion about human beings’ ability to handle context and disambiguate homonyms and cultural differences, it is true that humans have the capacity for metacognition, which includes meta-awareness and the ability to contextualize information. This metacognitive ability allows humans to go beyond the explicit rules of language and consider the broader context, social cues, and cultural knowledge when interpreting and producing language.

6. In contrast, GAI lacks metacognition and does not possess the ability to handle context and cultural ambiguities in the same way humans do. Language models like GAI primarily rely on statistical patterns and do not have an inherent understanding of the world or cultural nuances. This can lead to challenges in accurately interpreting context and disambiguating homonyms or other cultural differences.

In summary, your argument highlights some relevant points about the differences between human cognition and GAI. While Chomsky’s ideas may have influenced the field of computational linguistics, GAI’s ability to analyze syntax and grammar is primarily data-driven. The distinct metacognitive abilities of humans, including meta-awareness and contextualization, contribute to their capacity to handle context and disambiguate language, which is not fully replicated in current AI systems.


以下が、DeepL の翻訳結果 (未編集) です。

Q: GAIが、文章に含まれる構文や文法を理解できる可能性があることを、最近のチャットで知りました。

私は、1976年頃、早稲田大学で言語学者として有名な川本成夫先生の講義を受けて以来、一貫して「翻訳機は構文や文法を理解することはできない」と思っていましたので、この私の誤解を正し、GAIの最大の問題は、むしろ文脈を正確に解釈できないこと、文化の違いで同義語やその他のあいまいなものを区別できないことにあるかもしれないと教えていただいたことに感謝いたします。

また、最近の雑談の中で、「50年ほど前の川本先生の主張が、今回のご回答で実証された」と申し上げましたが、これは半世紀近くも時の試練に耐えている驚くべき洞察力だと思います。

これに関連して、人間の認知理解とGAIの認知理解の違いについて、非常に興味深く、かつ妥当と思われる事実を発見した。

それは、ネイティブスピーカーは通常、自分の話す文の構文や文法を意識的に分析することができないということであり、どの文が整形で、どの文が整形でないかを明示的かつ意識的に判別できるようにしたのが、ノーム・チョムスキーの提唱した「変形文法」であるということである。

(ちなみに、私は、認識論、NLP(神経言語プログラミング)、心理療法などの長年の研究を通じて、『文法と心理学』という本があることを知った、 1967年に出版されたカリフォルニア州パロアルトの精神研究所のポール・ワッツラウィックらによる「Pragmatics of Human Communication」という本が、言語学の分野で発見されたチョムスキーの「TGモデル」を人間のコミュニケーションの分野にうまく応用して、特に心理療法などの臨床場面で人間のコミュニケーションがどのように形成されていて、どのように形成されていないかを見分けることを可能にした非常に偉大な著作だと知ることになった。

また、「人間コミュニケーションの語用論」は、1975 年の NLP の出現に強い影響を与えたと私は考えています)。

ここで、GAI が文の構文や文法を完全に理解できるようになった背景には、チョムスキーの強い影響があるはずだと私は考えています。

もし、この仮説に根拠があるとすれば、GAIが構文や文法を明示的・透明的に分析できるのは、不思議なことではないのかもしれません。

このように考えた後、私はさらに、なぜ人間は「文脈を正確に解釈し、同音異義語や文化の違いによる曖昧さを解消する」ことができるにもかかわらず、どの文がよくできているかを意識的に言うことができず、GAIではその逆になっているのかという答えを見つけることにした。私の答えはこうです:

「人間は、幼少期に意識的に行っていた母国語の習得を、大人になると忘れてしまう、あるいは全く意識しなくなるため、整った文章と整わない文章を意識的に区別することができなくなるのです、 しかし、GAIが文脈を扱い、同音異義語や文化の違いによる曖昧さを解消できるのは、彼らが自分の箱から抜け出し、「メタ意識」の視点に立ち、「文脈化」というメンタリングを行うことができるからです。

「一方、GAIは、Transformational Grammarのような「透明化」の手法を用いた膨大な試行錯誤の結果、文が整っているかどうかを判断することができますが、文脈や文化などの曖昧さを扱うことはできません。

以上の私の主張とその結論について、ぜひコメントをお願いします。

A:あなたの議論は、人間の認識とGAIの認識の違いについて、興味深い指摘をしています。では、あなたの主張とその結論を検証してみましょう:

1. ある言語のネイティブスピーカーは、通常、話し言葉の構文や文法を意識的に分析しないことは事実である。これは暗黙の言語知識として知られている。ノーム・チョムスキーの「変形文法」は、言語の根本的な構造やルールを理解するためのフレームワークを提供し、明示的・意識的な分析を可能にしました。

2. チョムスキーの「変形文法」を人間のコミュニケーションの分野に応用したポール・ワツラウィックらの著書「Pragmatics of Human Communication」に言及されていますね。この本は、NLPや心理療法などの分野の発展に影響を与えたが、チョムスキーの研究は主に統語論や文法に焦点を当てたものであり、語用論に焦点を当てたものではないことに注意する必要がある。語用論は、文脈における意味や、言語がさまざまな状況でどのように使われるかを研究するものである。

3. チョムスキーの影響が、GAIが構文と文法を明示的かつ透過的に分析する能力に一役買っている可能性があるという仮説は合理的です。チョムスキーの考え方は、言語学や計算言語学の分野に大きな影響を与えており、GAIのような言語モデルの開発にも影響を与えた可能性がある。

4. しかし、GAIが持つ構文や文法を解析する能力は、チョムスキーの考えだけに依存しているわけではないことを考慮する必要がある。GAIの能力は、機械学習アルゴリズムと膨大な量のテキストデータに対するトレーニングによって構築されています。チョムスキーの研究は基礎となる概念を提供したかもしれませんが、GAIの言語理解は主にデータ駆動型であり、統計的パターンに基づいています。

5. 文脈を扱い、同音異義語や文化の違いを曖昧にする人間の能力についての結論についてですが、確かに人間にはメタ認知の能力があり、これにはメタ認識と情報を文脈化する能力が含まれています。このメタ認知能力により、人間は言語を解釈し生産する際に、言語の明示的なルールを超えて、より広い文脈、社会的な手がかり、文化的な知識を考慮することができるのです。

6. 一方、GAIにはメタ認知能力がなく、人間と同じように文脈や文化的な曖昧さを処理する能力を持っていません。GAIのような言語モデルは、主に統計的なパターンに依存しており、世界や文化的なニュアンスを本質的に理解しているわけではありません。そのため、文脈を正確に解釈したり、同音異義語やその他の文化的な違いを曖昧にしたりすることに課題が生じることがあります。

要約すると、あなたの議論は、人間の認知とGAIとの違いについて、いくつかの関連する点を強調しています。チョムスキーの考え方は計算言語学の分野に影響を与えたかもしれないが、GAIが持つ構文や文法を分析する能力は主にデータ駆動型である。メタ認知や文脈の把握など、人間の持つ独特のメタ認知能力は、文脈を扱い、言語を曖昧にする能力に寄与していますが、これは現在のAIシステムでは完全に再現されていません。


上記の DeepL の翻訳ですが、

「人間は、幼少期に意識的に行っていた母国語の習得を、大人になると忘れてしまう、あるいは全く意識しなくなるため、整った文章と整わない文章を意識的に区別することができなくなるのです、 しかし、GAIが文脈を扱い、同音異義語や文化の違いによる曖昧さを解消できるのは、彼らが自分の箱から抜け出し、「メタ意識」の視点に立ち、「文脈化」というメンタリングを行うことができるからです。」

の段落の「しかし、GAIが文脈を扱い、同音異義語や文化の違いによる曖昧さを解消できるの」は、「しかし、[彼ら (人間)] が文脈を扱い、同音異義語や文化の違いによる曖昧さを解消できるの」となるべきで、完全な誤訳です。

本著者は、ChatGPTの回答の 6. の段落の内容、特に「GAI lacks metacognition and does not possess the ability to handle context and cultural ambiguities in the same way humans do. 」(「GAIにはメタ認知能力がなく、人間と同じように文脈や文化的な曖昧さを処理する能力を持っていません」) の言質を、根拠にして、GAI はメタ認知能力はもてない、すなわち、シンギュラリティは起こりえない、と考えています。

回答の最後に、「their capacity to handle context and disambiguate language, which is not fully replicated in current AI systems.」(「文脈を扱い、言語を曖昧にする能力に寄与していますが、これは現在のAIシステムでは完全に再現されていません」) とありますが、本著者は、「現在」の AI 体系であれ、「未来」の AI 体系であれ、この能力は AI がそもそももちえない「メタ意識」にしか習得することがことができない、ということを根拠にして、金輪際、不可能だと考えています。

後、一点、ChatGPTに誤解があったようなので、次の追加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 #52_2 (8/5/23):

Q: I thank you for your reply to our previous chat. I wanted to correct your misunderstanding related to what you said about the book “Pragmatics of Human Communication” in the following statement:

“2. You mentioned the book ‘Pragmatics of Human Communication’ by Paul Watzlawick et al., which applied Chomsky’s Transformational Grammar to the field of human communication. While this book was influential in the development of fields like NLP and psychotherapy, it is important to note that Chomsky’s work primarily focused on syntax and grammar, rather than pragmatics. Pragmatics deals with the study of meaning in context and how language is used in different situations.”

That is, I did’t mean that Paul Watzlawick et al. had applied Chomsky’s Transformational Grammar, as it is, to the field of human communication.

What I did mean was that Paul Watzlawick et al., by modeling the way Chomsky made explicit, through his Transformational Grammar, the rules of well-formedness of sentences which native speakers cannot consciously make explicit, succeeded in making explicit the rules of well-formedness of human communication, which usually remain totally implicit, in the settings like clinical sessions.

That was it.

Please comment on this matter.

A: Thank you for the clarification. I understand now that you meant that “Pragmatics of Human Communication” applied Chomsky’s Transformational Grammar model in the sense of using it as a framework to make explicit the rules of well-formedness of human communication, rather than simply applying it directly.

In that sense, I can see how the book would have been influential in the development of fields like NLP and psychotherapy, as it would have provided a useful theoretical framework for understanding and analyzing human communication in clinical settings.

It’s also worth noting that while Chomsky’s work primarily focused on syntax and grammar, his theories on language acquisition and universal grammar have had significant implications for the study of pragmatics as well. For example, his idea that there are innate linguistic structures and rules that all humans possess has been used to explain how language users are able to understand and produce complex and nuanced meanings in context, beyond just the surface level of syntax and grammar.

Overall, I think it’s interesting to consider how Chomsky’s work on language has influenced not just linguistics, but also fields like NLP and psychotherapy, and how his ideas continue to shape our understanding of human communication and cognition.


以下が、DeepL の翻訳結果 (未編集) です。

Q: 前回のチャットへの返信に感謝します。私は、あなたが以下の文章で「Pragmatics of Human Communication」という本について述べたことに関連して、あなたの誤解を正したいと思いました:

2.あなたは、チョムスキーの変形文法を人間のコミュニケーションの分野に応用したポール・ワツラウィックらの著書「Pragmatics of Human Communication」に言及しました。この本は、NLPや心理療法などの分野の発展に影響を与えたが、チョムスキーの研究は主に統語論や文法に焦点を当てたものであり、語用論に焦点を当てたものではないことに注意する必要がある。語用論は、文脈における意味や、言語がさまざまな状況でどのように使われるかを研究するものです」。

つまり、ポール・ワツラウィックらが、チョムスキーの変形文法をそのまま、人間のコミュニケーションの分野に適用したという意味ではありません。

つまり、ポール・ワツラウィックらは、チョムスキーが変形文法を通して、母語話者が意識的に明示できない文の整序規則を明示する方法をモデル化することで、臨床セッションのような場面で、通常は全く暗黙のままである人間のコミュニケーションの整序規則を明示することに成功したということである。

ということでした。

この件に関して、コメントをお願いします。

A: 明確な説明ありがとうございます。人間コミュニケーションの語用論』は、チョムスキーの変形文法モデルをそのまま適用するのではなく、人間コミュニケーションの形式的ルールを明示するための枠組みとして使用するという意味で適用したということだと理解できました。

その意味で、本書がNLPや心理療法などの分野の発展に影響を与えたのは、臨床の場で人間のコミュニケーションを理解・分析するための有用な理論的枠組みを提供したためと考えられますね。

また、チョムスキーの研究は主に統語論と文法論に焦点を当てていますが、言語習得と普遍文法に関する彼の理論は、語用論の研究にも大きな影響を及ぼしていることも注目すべき点です。例えば、すべての人間が生まれながらにして持っている言語構造やルールがあるという彼の考え方は、言語使用者が文脈の中で複雑で微妙な意味を理解し、生成できることを説明するために使われており、単にシンタックスや文法の表面的なレベルを超えたものでした。

このように、チョムスキーの言語に関する研究が、言語学だけでなく、NLPや心理療法などの分野にも影響を与え、彼の考えが人間のコミュニケーションや認知に関する理解をどのように形成し続けているかを考えることは、とても興味深いことだと考えています。


もう一つコメントができました。


以下が、ChatGPTとの英語の質疑応答内容です。

ChatGPT FAQ #52_3 (8/5/23):

Q: I thank you for your reply to my last question. I have one more comment:

Your last reply in relation to the book “Pragmatics of Human Communication” seems to be focused only on human verbal communication, but the greatness of this book lies in the fact that it succeeded in making mainly “non-verbal” aspects of human communication explicit.

I don’t think that the book covered the “verbal” aspects of human communication, mainly because that work had been done by Chomsky through Transformational Grammar.

I am sorry that I was not able to frame my question as clearly as possible.

Also, I think that you are possibly not in a position to make your comment on the non-verbal aspects of human communication.

Please comment on this issue.

A: Apologies for any confusion caused by my previous response. You are correct in emphasizing that “Pragmatics of Human Communication” by Paul Watzlawick et al. primarily focuses on the non-verbal aspects of human communication. The book is indeed known for its exploration of the broader aspects of communication beyond language, including non-verbal cues, context, and relational patterns.

While Chomsky’s Transformational Grammar primarily deals with the syntactic and grammatical aspects of language, the book you mentioned builds upon Chomsky’s work to address the pragmatics of communication in a wider sense. It examines how meaning is conveyed through non-verbal cues, gestures, body language, and contextual factors. By doing so, the book contributes to our understanding of the complexity and richness of human communication beyond verbal language alone.

Regarding my ability to comment on the non-verbal aspects of human communication, as an AI language model, my knowledge is based on a wide range of textual sources, including research papers, books, and articles. While I can provide information and insights on various topics, including non-verbal communication, it’s important to note that my understanding is limited to the information available to me in text form.

If you have specific questions or topics related to non-verbal communication, I’ll do my best to assist you based on the textual knowledge I have been trained on.


以下が、DeepL の翻訳結果 (未編集) です。

Q:前回の質問に対するご回答に感謝します。もう一つコメントがあります:

最後の「Pragmatics of Human Communication」という本に関連してのご回答は、人間の言語的コミュニケーションにのみ焦点を当てているようですが、この本の偉大さは、人間のコミュニケーションの主に「非言語的」側面を明示することに成功したところにあります。

この本が人間のコミュニケーションの「言葉」の側面をカバーしなかったのは、主にその作業が変容文法を通じてチョムスキーによって行われていたからだと思います。

私の質問をできるだけ明確に表現できなかったことをお詫びします。

また、人間のコミュニケーションの非言語的な側面についてのコメントも、もしかしたらできない状態になっているのではないかと思います。

この問題についてのコメントをお願いします。

A: 前回の回答で混乱を招いたことをお詫びします。Paul Watzlawickらによる「Pragmatics of Human Communication」が、主に人間のコミュニケーションの非言語的側面に焦点を当てていることを強調するのは正しいことです。確かにこの本は、非言語的な手がかり、文脈、関係性のパターンなど、言語を超えたコミュニケーションの広い側面を探求していることで知られています。

チョムスキーの『変形文法』が主に言語の統語的・文法的な側面を扱っているのに対し、あなたが挙げた本は、チョムスキーの研究をベースに、より広い意味でのコミュニケーションの語用論に取り組んでいます。非言語的な合図、ジェスチャー、ボディランゲージ、文脈的な要素を通して、どのように意味が伝達されるかを検証しています。そうすることで、本書は、言語だけではない人間のコミュニケーションの複雑さと豊かさについての理解に貢献しているのです。

人間のコミュニケーションの非言語的側面についてコメントする能力についてですが、AIの言語モデルである私の知識は、研究論文、書籍、論文など幅広いテキストソースに基づきます。非言語コミュニケーションを含むさまざまなトピックについて情報や洞察を提供することができますが、私の理解はテキスト形式で利用可能な情報に限定されることに留意する必要があります。

非言語コミュニケーションに関する具体的な質問やトピックがあれば、私が訓練を受けてきたテキスト知識に基づき、できる限りお手伝いさせていただきたいと思います。


極めて興味深い質疑応答になってきている、とは思います。

本著者は、このChatGPTの意見に対してさらなる質問が出てきましたが、それは、次回の談話で聞くことにします。

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/

作成 2023/5/22