記事作成日:2023年11月3日


AI

新聞記事でAIに関する重要な英語の論文が紹介されていました。 その論文は経済研究へのAIの活用(特にLLMという言語モデル)に関する内容ですが、実際のビジネスでも活用できる内容と 考え、その翻訳をしてみました。 30ページほどの論文ですが、重要な部分だけを選択して英語と日本語訳の対比という形式で数回に分けて紹介します。

今回は全11回中の5回目です。

第1回 1 Introduction:はじめに
第2回 2 What Are LLMs?:LLMとは何か?
第3回 3 Applications:応用
第4回 3.1 Ideation:アイデア出し
第5回 3.2 Writing:執筆
第6回 3.3 Background Research:背景調査
第7回 3.4 Coding:コーディング
第8回 3.5 Data Analysis:データ分析
第9回 3.6 Mathematical Derivations:数学的導出
第10回 3.7 Summary:まとめ
第11回 4 Outlook and Concluding Thoughts:展望と考察

3.2 Writing

The core competence of LLMs is to generate text. This implies that they are quite capable and useful for many tasks related to writing, from synthesizing text based on bullet points, changing the style of text, editing text, evaluating styles, and generating titles, headlines, and tweets. Let us cover each in turn.

3.2 執筆

LLMの中核を担う性能はテキストの生成であるため、執筆に関する幅広いタスクを 得意とします。箇条書きを与えて自然な文章を生成したり、テキストの口調や文体を変換したり、タイトルを生成したり、ツイートを書いたり、さまざまな用途で役に立てられます。各ケースを具体的に見ていきましょう。

3.2.1 Synthesizing text

One of the most useful capabilities of LLMs may be to translate rough bullet points into well-structured sentences that are clear and easy to read. The following example describes a typical use case for this. Note that it also asks the system to provide examples and offer solutions:

Please write a paragraph that integrates the following arguments and goes beyond. Include a topic sentence.

  • large language models automate cognitive tasks
  • the work of economists is largely cognitive [add examples]
  • we economists should therefore worry about recent progress in language models
  • in the short run, these models can help us be more productive
  • in the long run, there is a good likelihood that they will displace human economists
  • we should think how to prepare [offer solutions]

The recent progress in large language models has the potential to automate many cognitive tasks, including those of economists. Economists are responsible for analyzing data, making predictions, and providing policy advice, all of which are cognitive tasks. Therefore, economists should be concerned about the implications of these language models, both in the short and long run. In the short run, these models can help us be more productive, but in the long run, there is a good likelihood that they will displace human economists. To prepare for this eventuality, economists should think about how to use these models to their advantage and how to develop skills that are not easily automated. This could include developing expertise in areas such as communication, problem-solving, and critical thinking.

This allows researchers to concentrate their energy on the ideas in their text as opposed to the writing process itself. The text is generally high-quality (although some observe that it is, naturally, a bit sterile and lacks the idiosyncracies and elements of surprise that characterize human writing - a feature that detectors of LLM-written text zero in on).

3.2.1 テキストの合成

その中でも、最も有用なLLMの性能の1つは、箇条書きをしっかり構成され、読みやすくてわかりやすい文章に変換することです。こちらに典型的な使用例を示します。AIシステムには、例を提供することと解決策を提供することも求めていることに注意してください。

次に挙げる論点を取り入れ、さらにこれらを超えた内容で一段落の文章を書いてください。トピックセンテンスも提案してください。

  • 大規模な言語モデルは認知タスクを自動化できる。
  • 経済学者の仕事はほとんどが認知的なものである。 [例を追加]
  • したがって、私たち経済学者は言語モデルの最近の進歩について心配をすべきである.
  • 短期的には、これらのモデルは生産性の向上に役立つだろう.
  • 長期的には、彼らが人間の経済学者に取って代わる可能性が高い。
  • [解決策を]どのように準備するかを考案すべきである。

大規模言語モデルの最近の進歩により、経済学者の仕事を含む多くの認知タスクを自動化する可能性が示されています。経済学者はデータの分析、予測、政策に関するアドバイスの提供をになっていますが、これらはすべて認知タスクです。したがって、経済学者は、短期的にも長期的にも、これらタスクに対する言語モデルの影響を懸念すべきです。短期的には、これらのモデルは私たちの生産性の向上に貢献することができますが、長期的には人間の経済学者に取って代わられる可能性が高くなります。この不測の事態に備えるために、経済学者はこれらのモデルを有益に利用する方法とともに、簡単にAIによって自動化されないような高度なスキルを開発することに重点を置くべきです。これらの高度スキルには、コミュニケーション、問題解決能力、批判的思考などの専門分野における専門知識が挙げられます。

これにより、研究者は執筆の作業そのものではなく、テキスト内のアイデアにエネルギーを集中することができます。テキストは全体的に高品質です (ただし、予想通りにやや役に立たない部分もあり、人間が書いた文章の特徴である特異性や驚きの要素が欠落しているとの見方もあります。「AI生成コンテンツの検出器」はこのような特徴を手がかりにしています)。

LLMs are also responsive to requests to write in a particular style. For example, the example above will read differently if we add Write in an academic style , “colloquial style”, “style that is comprehensible to non-economists” or “in the style of [enter your favorite politician].” When I added the sentence “Write in the style of a gangster.” the system started with “Yo, economists, it’s time to wake up and smell the coffee.” Alternatively, when I wrote “Can you add a joke?” it included the following as the penultimate sentence: “After all, we don’t want to be replaced by a machine that can’t even tell a good joke!” Of perhaps greater relevance for our daily writing, LLMs can also write text in LaTeX format, which automatically surrounds variables by the $ signs required for LaTeX, and it can draft beamer slides in LaTeX code based on bullet points. Writing is of course a skill that is not only useful for research - cognitive workers in many areas report using, e.g., ChatGPT to draft emails, memos, performance reviews, or summaries based on a few bullet points. Likewise, professors report generating the first draft of letters of recommendation or syllabi for standard courses using LLMs. The writing tool https://lex.page incorporates several related functions in an easy-to-use web interface.

LLMはさらに、特定の文体やスタイルで文章を執筆する指示に従う事ができます。 例えば、上で示した例に関して、もしプロンプトに「学術的なスタイル」「口語的 なスタイル」、「経済学者以外にも理解できるスタイル」、または「[好きな政治家を入力] のスタイル」と指定を付け加えた場合に、全く異なる雰囲気の出力が得られるでしょう。仮に「暴力団員のような口調で書いて」とでも指示した場合、おそらく「おい!経済学者らは皆起きてこい…」のようなものが出てくるのでしょう。代わりに「冗談も1つくらい付け加えて」と指示を出した場合、「冗談さえ言えないような機械に置き換えられたくないですよね」と出てくるかもしれません。これらは私たちの日常的な執筆への気づきを与えてくれる可能性もあります。 LLMはLATEX形式で書くこともできます。その場合、自動的に変数を$記号で囲みます。また、箇条書きに基づいて LaTeX コードでビーマースライドを作成できます。もちろん、文章の執筆は研究者に限って有用というわけではなく、さまざまな分野の認知労働に携わっている人々は、ChatGPTをメールの下書きの作成、メモの作成、人間が書いた文章のレビュー、箇条書きに基づく文章作成などに使用することを報告しています。 同様に、教授たちでさえ、LLM を使用して推薦状またはシラバスの初稿を作成していると報告しています。https://lex.page にある執筆ツールは、使いやすいWeb インターフェイスにいくつかの関連機能が組み込まれています。

3.2.2 Editing text

Another useful capability is editing. LLMs can edit text for grammatical or spelling mistakes, style, clarity, or simplicity. This set of capabilities are perhaps most useful for non-native speakers who want to improve their writing. In the following example, I have bolded each word with a mistakes as well as the corrections made by the system for emphasis:

Can you correct the following sentence?

These paper describe six domain in which large language models (LLMs) displays emergant capbilities that are useful to the economic research.

This paper describes six domains in which large language models (LLMs) display emergent capabilities that are useful for economic research.

Can you explain the main corrections you made?

I changed “These” to “This” to make the sentence singular, and I changed “useful to” to “useful for” to make the phrase more grammatically correct.

As illustrated by the follow-up question, the system can even even explain its edits afterwards so that students can learn from any corrections it has made. Another useful application is that the system can convert text to be comprehensible to readers of different levels, for example, using a prompt like “Rewrite the following text so that it is comprehensible to an eight-year-old” or “comprehensible to an eighth-grader.”

3.2.2 文章の編集

人間が書いた文章の編集はLLMのもう1つの有用な機能です。文法やスペルの間違いを見つける、スタイルやわかりやすさの改善、簡潔な言い方を提案などが含まれます。これらの機能は英語ネイティブ以外の方がライティングスキルを改善したいときに使いたいものかもしれません。次の例では、私は間違いの箇所およびLLMによる修正の結果の箇所を強調するために太字にしています。

次の一文を修正してくれますか?

これらの論文(単数形)は大規模言語モデル(LLM)が経済学の研究にとって性能を発揮する6つの領域(単数形)を提示します。

この論文は大規模言語モデル(LLM)が経済学の研究に有用な性能を発揮する6つの領域を提示します。

どのような修正を行なったのかを説明してください。

論文が単数形だったので「これらの」を「この」に変更しました。また「有用」の文法を調整しました。

私が行なった追加の質問に見えるように、AIのシステムはさらに、修正を行なった根拠を解説できます。そうすると生徒は修正された箇所から学びを得ることができます。もう1つの有用な応用としては、AIシステムは文章をあらゆるレベルの読者の理解度に合わせた形に変換することができます。例えば、「この文章を8歳の子供でも理解できるようなレベルに書き換えてください」とプロンプトで指示することが可能です。

3.2.3 Evaluating text

LLMs can also evaluate text for its style, its clarity, or similar. The following question about a draft abstract of this paper turned out to be a bit of a defeating experience:

What are the main stylistic shortcomings of the following text?

This paper describes six domains in which large language models (LLMs) display emergent capabilities that are useful for economic research: ideation, writing, background research, data analysis, coding, and mathematical derivations. I provide general instructions as well as specific use cases for how to take advantage of each of these and classify the LLM capabilities that are displayed on a range from “experimental” to “highly useful.” I hypothesize that ongoing advances will improve the performance of LLMs across all of these domains, and that economic researchers who take advantage of LLMs to automate micro tasks will become signi cantly more productive. Finally, I speculate on the longer term implications of cognitive automation via LLMs for economic research.

The main stylistic shortcomings of the text are that it is overly wordy and lacks a clear structure. The sentences are too long and complex, making it difficult to follow the author’s argument. Additionally, the text lacks transitions between ideas, making it difficult to follow the flow of the argument. Finally, the text does not provide any concrete examples or evidence to support the author’s claims.

Can you rewrite the paragraph to correct for these shortcomings?

This paper examines six domains in which large language models (LLMs) can be used for economic research: learning, ideation, writing, data analysis, coding, and mathematical derivations. I provide instructions and use cases for each domain, and classify the LLM capabilities from “experimental” to “highly useful.” I hypothesize that LLMs will continue to improve in all of these areas, and that economic researchers who use them to automate micro tasks will benefit from increased productivity. Finally, I discuss the implications of cognitive automation via LLMs for economic research.

After some deliberation, I agree with all the shortcomings identified by the LLM. The rewritten version alleviates some of them but does not do so perfectly. The system can also respond to question such as Which argument is the hardest to follow in the following text?

3.2.3 文章の評価

LLMは、文章を文体、わかりやすさなどの観点から評価することが可能です。本論文に関する要約の草案に対して質問をした時に、少し難しそうでした。

次の文章の文体に関する主な欠点を挙げてください。

この論文では、大規模言語モデル(LLM)が経済学の研究に役立つ創発的な機能を発揮する6つの領域 (アイデア出し、執筆、背景調査、データ分析、コーディング、数学的導出) について説明しています。それぞれのユースケースについて、「実験的」から「非常に有用」にわたるLLM機能を、一般的な手順と具体的な使用例を示します。今後も継続するのであろう技術の進歩により、これらすべての領域にわたってLLMの性能が向上し、LLMを利用して細かいタスクを自動化することによって、経済学の研究者の生産性が大幅に向上すると仮説を立てています。最後に、LLMによる認知タスクの自動化が、経済学の研究に及ぼす長期的な影響も推測しています。

この文章の文体に関する主な欠点としては、文体が冗長であり、明確な構造が見当たらないことが挙げられます。文章が長すぎて複雑なので、読者にとって著者の主張を掴みにくくなっています。さらに、この文章ではアイデア間の関連性が不足しているため、議論の流れを追うことが難しいです。最後に、著者の主張を裏付ける具体的な例や証拠が提供されていません。

これらの欠点を修正した文章を書き直してくれますか?

この論文では、経済学の研究に大規模言語モデル (LLM) を活用できる 6つの領域(学習、アイデア出し、執筆、データ分析、コーディング、数学的導出) について考察をしています。各領域の手順と使用例を提供し、LLMの機能を「実験的」から「非常に有用」まで分類しています。私は、LLMはこれらすべての分野において今後も改善され、マイクロタスクを自動化するために LLMを使用する経済学の研究者は生産性向上の恩恵を受けるだろうと仮説を立てています。最後に、LLM による認知の自動化が経済学の研究に与える影響についても説明します。

しばらく考えた後に、私はLLMによって指摘されたこれらの欠点に同意するようになりました。リライトされたバージョンは欠点の一部を改善していますが、完全にというわけではありません。AIのシステムは「次の文章の中で、どの論点が一番ついていくのが大変ですか?」と言った質問にも答えることができます。

3.2.4 Generating catchy titles & headlines

Another useful application is to come up with catchy headlines or paper titles, as illustrated by the following example based on the abstract of a recent paper of mine (Korinek and Juelfs, 2022):

Please suggest two or three catchy titles for a paper with the following abstract:

We analyze how to set up institutions that future-proof our society for a scenario of ever-more-intelligent autonomous machines that substitute for human labor and drive down wages. We lay out three concerns arising from such a scenario, culminating in the economic redundancy of labor, and evaluate recent predictions and objections to these concerns. Then we analyze how to allocate work and income if these concerns start to materialize. As the income produced by autonomous machines rises and the value of labor declines, we find that it is optimal to phase out work, beginning with workers who have low labor productivity and job satisfaction, since they have comparative advantage in enjoying leisure. This is in stark contrast to welfare systems that force individuals with low labor productivity to work. If there are signi cant wage declines, avoiding mass misery will require other ways of distributing income than labor markets, whether via sufficiently well-distributed capital ownership or via benefits. Recipients could still engage in work for its own sake if they enjoy work amenities such as structure, purpose, and meaning. If work gives rise to positive externalities such as social connections or political stability, or if individuals undervalue the bene ts of work because of internalities, then there is a role for public policy to encourage work. However, we conjecture that in the long run, it would be more desirable for society to develop alternative ways of providing these benefits.

  1. “Future-Proofing Society: Preparing for Autonomous Machines and the Decline of Labor”
  2. “The End of Work? Navigating the Impact of Autonomous Machines on Labor”
  3. “The Impact of Autonomous Machines on Labor: How to Allocate Work and Income”

All three suggested titles would fit the described work quite well.

3.2.4  キャッチーなタイトルや見出しを生成

LLMをキャッチーなタイトルや見出しを生成するにも活用することができます。今の例は私の最近の論文の概要に基づいています (Korinek and Juelfs, 2022):

次の概要をもつ論文のために、キャッチーなタイトルを2、3個提案してください。

本論文では、人間の労働力を代替し、賃金を引き下げる、ますますインテリジェントな自律的な機械が登場するというシナリオに対して、社会の将来性を保証する制度をどのように設立するかを分析しています。私たちは、このようなシナリオから生じ、最終的には労働力の経済的余剰に至る3つの懸念を示し、これらの懸念に対する最近の予測と反論を評価します。次に、これらの懸念点が現実になりはじめた場合に、仕事と収入をどのように配分すべかを考察します。自律機械によって生み出される所得が増加し、労働価値が低下するにつれて、余暇を楽しむという点で比較優位を持つ労働生産性と仕事満足度の低い労働者から始め、段階的に仕事を廃止することが最適であることがわかります。これは、労働生産性の低い個人に労働を強いる福祉制度とは対照的といえます。賃金が大幅に低下した場合、大勢の苦悩を回避するには、十分に分散された資本所有権または福利厚生を通じて、労働市場以外の所得分配方法が必要となります。受領者は、構造、目的、意味などの仕事の快適さを享受できれば、それ自体のために仕事に取り組むことができます。仕事が社会的つながりや政治的安定などのプラスの外部性を生み出す場合、または個人が内部性のために仕事の利点を過小評価している場合、公共政策には仕事を奨励する役割があります。しかし、長期的には、これらの利益を提供する別の方法を開発することが社会にとって望ましいと私たちは推測しています。

  1. 将来を検証する社会:自律的な機械と労働力の減少の準備
  2. 労働の終了?自律的な機械が労働力に及ぼす影響を解析
  3. 自律的な機械が労働力に及ぼす影響:労働と収入の配分

提案された3つのタイトルはどれも、この論文にかなり相性がいいと思っています。

3.2.5 Generating tweets to promote a paper

The following exercise is similar but converts the abstract I pasted into five tweetable chunks that would work well for Twitter:

Write five tweets that summarize the following text. Write simple sentences and make them as catchy as possible: We analyze how to set up institutions […same abstract as above…]

  1. Worried about the future of work? We analyze how to set up institutions to future-proof our society. #FutureOfWork #Automation #IncomeRedistribution
  2. As autonomous machines rise and labor value declines, it’s optimal to phase out work starting with those with low labor productivity. #FutureOfWork #Automation #IncomeRedistribution
  3. If wage declines occur, avoiding mass misery will require other ways of distributing income than labor markets. #FutureOfWork #Automation #IncomeRedistribution
  4. If work gives rise to positive externalities, there’s a role for public policy to encourage work. #FutureOfWork #Automation #IncomeRedistribution
  5. In the long run, it’s more desirable to develop alternative ways of providing benefits. #FutureOfWork #Automation #IncomeRedistribution

We can only speculate on what this capability will imply for the future of content on Twitter. Goldstein et al. (2023) provide an overview of the resulting threats and potential mitigations.

3.2.5 論文を宣伝するためのツイートの生成

次の演習は先程のと類似しているが、今度は私が貼り付けた論文の概要をTwitterでのツィートに使えるような5つの塊に分割するという用件です。

次の文章のサマリーとなる5つのツィートを書いてください。簡単かつ可能の限りキャッチーな文を書いてください。(上と同じ概要)

  1. 労働の将来を心配?我々の社会の健全性を証明 する #仕事の将来 #自動化 #収入再配分
  2. 自律的な機械の登場によって労働力が減るにつれて、まずは生産性の低い仕事から消すのが良さそう #仕事の将来 #自動化 #収入再配分
  3. 収入が減少した場合、大勢の苦悩を回避するためには労働市場と収入の再配分の新しい方法を要する。 #仕事の将来 #自動化 #収入再配分
  4. 仕事によって有益な外部効果が生じる場合、公共政策には仕事を支持する役割がある #仕事の将来 #自動化 #収入再配分
  5. 長期的に、利益を得られるための代替的な手段を開発するのが良さそう #仕事の将来 #自動化 #収入再配分

このような結果から、Twitterで投稿されるコンテンツの未来が思い知らされます。 Goldstein et al. (2023) はこの結果として生じうる脅威と潜在的な軽減対策の概要を発表しています。

参照



    この記事をシェア👉🏻



カテゴリー
タグ