本日の談話テーマ: 「GAIは、表層構造から深層構造を予測できるか?」

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51」です。

今回は、GAIは、表層構造から深層構造を予測できるかどうか、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 #51 (7/5/23):

Q: I think that people like Greg Brockman and Ilya Sutskever may also be suggesting that GAI as a prediction machine that tries to predict the next character or word in a given sentence can “reconstruct” from written texts, or the “surface structure” in Chomsky’s terminology for that matter, the “underlying reality”, or the deep structure” for that matter.

I want you to comment on this matter.

A: It’s important to clarify the terminology and concepts being discussed here to provide an accurate comment.

GAI as a prediction machine, as suggested by individuals like Greg Brockman and Ilya Sutskever, focuses on training AI models to predict the next character or word in a given sequence. This approach enables the models to learn statistical patterns and dependencies within the data. However, it’s essential to note that this process primarily deals with the surface structure of language, which refers to the observable form or sequence of words.

In linguistics, Noam Chomsky’s theory distinguishes between surface structure and deep structure. Surface structure refers to the grammatical structure of a sentence as it appears in speech or writing, while deep structure represents the underlying meaning or abstract syntactic representation of a sentence. Chomsky’s theory aims to explain the generative processes of language and how meaning is derived from underlying structures.

While GAI prediction models can be trained on vast amounts of text data and learn patterns, it’s crucial to understand that they are primarily focused on surface-level patterns and statistical relationships rather than accessing or understanding deep structures or underlying realities.

The ability of GAI to “reconstruct” or understand the underlying reality or deep structure solely based on written texts is a much more complex task that goes beyond current AI capabilities. Understanding deep structures typically involves not only syntactic analysis but also semantic, contextual, and conceptual understanding, as well as knowledge inference and reasoning capabilities.

While GAI systems can make predictions and generate text based on surface-level patterns, they do not possess the comprehensive understanding or access to the underlying realities that humans can achieve through complex cognitive processes, background knowledge, and contextual interpretation.

In summary, while GAI prediction models can learn surface-level patterns and make predictions based on textual data, they do not possess the deep understanding or access to underlying realities akin to the concept of deep structure in linguistic theory. The process of comprehending underlying meanings and realities requires a more sophisticated and comprehensive understanding of language, context, and world knowledge that is currently beyond the scope of AI systems.


以下が、DeepL の翻訳結果 (未編集) です。

Q: Greg BrockmanやIlya Sutskeverのような人々は、与えられた文の次の文字や単語を予測しようとする予測機械としてのGAIが、書かれた文章、あるいはそれに関するチョムスキーの用語でいう「表面構造」から、「根本的な現実」、あるいはそれに関する「深い構造」を「再構成」できると示唆しているのかもしれないと考えています。

この件に関してコメントしてほしい。

A:正確なコメントをするためには、ここで議論されている用語や概念を明確にすることが重要である。

Greg BrockmanやIlya Sutskeverが提案する予測マシンとしてのGAIは、与えられたシーケンスの中で次の文字や単語を予測するAIモデルのトレーニングに焦点を当てています。このアプローチにより、モデルはデータ内の統計的パターンや依存関係を学習することができます。しかし、このプロセスは、主に言語の表面構造を扱うものであり、観察可能な形や単語の並びを指すことに留意する必要がある。

言語学では、ノーム・チョムスキーの理論により、表層構造と深層構造が区別されます。表層構造とは、音声や文章に現れる文の文法構造を指し、深層構造とは、文の根本的な意味や抽象的な構文表現を指します。チョムスキーの理論は、言語の生成過程と、意味がどのように根本的な構造から導き出されるかを説明することを目的としています。

GAI予測モデルは、膨大な量のテキストデータに対して学習させ、パターンを学習させることができますが、深い構造や根底にある現実へのアクセスや理解ではなく、表面レベルのパターンや統計的な関係に主眼を置いていることを理解することが極めて重要です。

GAIが、書かれたテキストのみに基づいて、根本的な現実や深い構造を「再構築」し理解する能力は、現在のAIの能力を超えた、より複雑なタスクです。深層構造の理解には、一般的に構文解析だけでなく、意味理解、文脈理解、概念理解、さらに知識推論や推論能力も必要です。

GAIシステムは、表面レベルのパターンに基づいて予測を行い、テキストを生成することができますが、人間が複雑な認知プロセス、背景知識、文脈の解釈を通じて達成できるような包括的な理解や根本的な現実へのアクセスは持っていません。

つまり、GAI予測モデルは、表面的なパターンを学習し、テキストデータに基づいて予測を行うことができますが、言語理論における深層構造の概念のような、根本的な現実に対する深い理解やアクセスは持っていません。意味や現実を理解するためには、言語、文脈、世界の知識をより高度かつ包括的に理解する必要があるが、現在のところAIシステムの範囲外である。


以上のChatGPTの回答は、Ilya Sutskever 氏が動画で示唆していたことと少し違うと思ったので、再質問してみました。


以下が、ChatGPTとの英語の再質疑応答の内容です。

ChatGPT FAQ #51_2 (7/5/23):

Q: I think that your previous reply may slightly differ from what Ilya Sutskever said in the following movie clip:

Be that as it may, I just thought that GAI being able to possibly understand the syntax and grammar of sentences may correspond with someone being able to “guess” the underlying “deep structure” from the textual “surface structure”.

Am I wrong?

A: As an AI language model, I’m unable to access or view specific external content such as movie clips. However, I can address your question based on the information you provided.

Understanding the syntax and grammar of sentences can indeed help in interpreting the underlying meaning or deep structure of a text. In linguistics, the deep structure refers to the abstract representation of the meaning and relationships within a sentence, while the surface structure is the actual arrangement of words and phrases.

By analyzing the surface structure and applying knowledge of syntax and grammar, an AI language model like GPT-3.5 can make educated inferences about the deep structure of a sentence. It can recognize patterns, identify relationships between words and phrases, and generate coherent responses based on this understanding. However, it’s important to note that while AI models like GPT-3.5 can make predictions and generate plausible responses, they do not possess true understanding or consciousness.

In summary, your suggestion that understanding the syntax and grammar of sentences can correspond to inferring the underlying deep structure is generally valid, and AI language models leverage these principles to generate text.


以下が、DeepL の翻訳結果 (未編集) です。

Q:前回のご回答と、以下のムービークリップでのイリヤ・スーツケバーの発言は、若干異なるのではないかと思います:

それはともかく、GAIが文章の構文や文法を理解できる可能性があるということは、文章の「表層構造」から根本的な「深層構造」を「推測」することができる人に相当するのではないかと思ったのです。

私の考えは間違っているのでしょうか?

A: AI言語モデルである私は、動画など外部の特定のコンテンツにアクセスしたり、閲覧したりすることはできません。しかし、提供された情報に基づいて、あなたの質問に対処することはできます。

文の構文や文法を理解することは、文章の根本的な意味や深い構造を解釈する上で確かに役立ちます。言語学では、深層構造とは文中の意味や関係を抽象的に表現したものであり、表層構造とは実際の単語やフレーズの配置を指します。

GPT-3.5のようなAI言語モデルは、表面構造を分析し、構文や文法の知識を適用することで、文の深層構造について教育的推論を行うことができます。文のパターンを認識し、単語やフレーズ間の関係を特定し、その理解に基づいて首尾一貫した応答を生成することができるのです。しかし、GPT-3.5のようなAIモデルは、予測したり、もっともらしい応答を生成することはできても、真の理解や意識を持っているわけではないことに注意する必要がある。

つまり、文の構文や文法を理解することは、その下にある深層構造を推測することに相当するという指摘は、一般的に妥当であり、AI言語モデルはこの原則を活用して文章を生成しているのです。


本談話の回答で「AI models like GPT-3.5 can make predictions and generate plausible responses, they do not possess true understanding or consciousness.」(「GPT-3.5のようなAIモデルは、予測したり、もっともらしい応答を生成することはできても、真の理解や意識を持っているわけではない」) で、ChatGPTが「シンギュラリティは起こらない」という確認をしているように、本著者には思えます。

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 2023/5/21