AIによる心理学英文の校閲におけるプロンプトの比較

はじめに

英語が得意ではないので，昔から色々なツールを使って論文や学会発表の英語を書いてきました。例えば，論文管理ソフトの横断検索を使ってお手本となる英語文章を探してそれを元に書く（いわゆる英借文）やSpringer Examplarのような学術的な英語表現の頻度検索サイトの活用など。そして，最近はDeepL WriteやchatGPTに代表されるAIによる英文校閲が強力なツールだと感じています。

chatGPTによる英文校閲ではプロンプトが重要

chatGPTによる英文校閲はどうやら，英語でお願いする方が質が高いようです。そのための英文校閲用プロンプトが色々と紹介されてます。例えば，こちらのページで紹介されているプロンプトとして以下のようなものがあります。

I want you to act as an English translator, spelling corrector and improver. I will speak to you in any language and you will detect the language, translate it and answer in the corrected and improved version of my text, in English. I want you to replace my simplified A0-level words and sentences with more beautiful and elegant, upper level English words and sentences. Keep the meaning same, but make them more scientific and academic. I want you to only reply the correction, the improvements and nothing else, do not write explanations. My sentences are “ここに文章を入れる”

たしかにこのプロンプトを使ってみると，自分の英文がかなり洗練された文賞になるようです。「ようです」と推測しているのは，知らない単語やあまり論文ではない単語に書き換わることもあり，意味としてはたしかに合っているが，心理学論文として書き替えられた単語は適切なのかの判断をしなければいけない場合も多いです（これは英文校閲会社に出した場合も同様ですが）。おそらく，これは「I want you to replace my simplified A0-level words and sentences with more beautiful and elegant, upper level English words and sentences.」という指示に起因するようです。なので，この指示は外した方がいいかもしれません。

そこで，もう少し自分でプロンプトを修正してみました。

I want you to act as an cognitive psychologist, spelling corrector and improver. I will speak to you in Einglish, answer in the corrected and improved version of my text, in English. Keep the meaning same, but make them more scientific and academic. I want you to only reply the correction, the improvements and nothing else, do not write explanations. My sentences are “ここに文章を入れる”

あまり変えてないのですが，翻訳家ではなく認知心理学者かつ英文校閲者として振る舞ってもらうことにし，よりハイレベルな英文に書き替えるという指示は取り除きました。この指示だと，大きな修正はあまり行われないようです。

プロンプト別の出力例の比較

実際に，それぞれのプロンプトで英文を校閲してもらいました。元の文章は自分が前に書いた論文のアブストラクトです。この文章自体は英文校閲会社の英文校閲が入っているので，修正する必要はあまりないはずなので，参考程度にしてください。実際に，自分が書いてそのままの英文では修正はもっと入りました。特に紹介されていたプロンプトでは書き替えられる量も多かったです。

原文

The retrieval of episodic memory can cause forgetting of other related episodic memories, a phenomenon known as retrieval-induced forgetting (RIF). The retrieval of semantic memory also leads to RIF. Although previous studies demonstrated unsuccessful RIF of negative words by episodic retrieval (Dehili & Brennen, 2008; Kobayashi & Tanno, 2012), no study clarified whether semantic retrieval can lead to successful forgetting of negative words. Against this background, we examined RIF of negative words by semantic retrieval and directly compared RIF of negative words by episodic and semantic retrieval. In each experiment, participants learned neutral and negative words and then engaged in retrieval practice. During retrieval practice, participants were required to recall words that associated with learned words from episodic memories (experiment 1) or semantic memories (experiment 2). Finally, a stem-cued recall test for learned words was administered. The results showed unsuccessful RIF of negative words by episodic retrieval and successful RIF of negative words by semantic retrieval. Thus, semantic retrieval can cause forgetting of negative words, but episodic retrieval not. We propose that episodic integration, which a high level of inter-relationship between negative words might encourage, reduce or eliminate RIF of negative words by episodic retrieval.

紹介された元のプロンプトによる修正

The retrieval of episodic memory can induce forgetting of other related episodic memories, which is commonly known as retrieval-induced forgetting (RIF). It is worth noting that RIF can also occur during the retrieval of semantic memory. Although previous research has demonstrated unsuccessful RIF of negative words through episodic retrieval (Dehili & Brennen, 2008; Kobayashi & Tanno, 2012), none have investigated whether semantic retrieval can lead to successful forgetting of negative words. Against this background, we aimed to examine RIF of negative words through semantic retrieval and to compare it directly with RIF of negative words through episodic retrieval. In each experiment, participants were taught neutral and negative words and then required to engage in retrieval practice. During retrieval practice, participants were instructed to recall words associated with the learned words from either episodic memories (Experiment 1) or semantic memories (Experiment 2). Finally, a stem-cued recall test for the learned words was conducted. The results indicated unsuccessful RIF of negative words through episodic retrieval and successful RIF of negative words through semantic retrieval. As a result, semantic retrieval can lead to the forgetting of negative words, whereas episodic retrieval cannot. We postulate that episodic integration, which is influenced by a high level of inter-relationship among negative words, might reduce or eliminate RIF of negative words through episodic retrieval.

こちらで少し手を入れたプロンプトによる修正

The retrieval of episodic memory can result in forgetting of other related episodic memories, known as retrieval-induced forgetting (RIF). Similarly, semantic retrieval can lead to RIF. Although previous studies have demonstrated unsuccessful RIF of negative words through episodic retrieval (Dehili & Brennen, 2008; Kobayashi & Tanno, 2012), no study has clarified whether semantic retrieval can successfully lead to forgetting of negative words. Against this backdrop, we investigated RIF of negative words through semantic retrieval and directly compared it to RIF of negative words through episodic retrieval. In each experiment, participants learned neutral and negative words, after which they engaged in retrieval practice. During retrieval practice, participants were required to recall words associated with the learned words from episodic memories (Experiment 1) or semantic memories (Experiment 2). Finally, a stem-cued recall test for learned words was administered. Results indicated unsuccessful RIF of negative words through episodic retrieval and successful RIF of negative words through semantic retrieval. Thus, semantic retrieval can lead to forgetting of negative words, but episodic retrieval cannot. We suggest that episodic integration, which may be fostered by high interrelationships between negative words, reduces or eliminates RIF of negative words through episodic retrieval.

どのような活用が可能か

このようにAIによる英文校閲は非常に便利です。さらに，校閲してもらった英文をDeepL Writeなどで再チェックするなどの方法も有効だと思います。ただ，言うまでもないことですが，修正された英文がそのまま使えるわけでもないです。分野やテーマ特有の表現もあるはずですので，最低限，それらの英語を自分でチェックして修正できる程度の英語能力は必要でしょう。

注意点

雑誌によってはAIによって生成された文章の利用が禁止されていたり，利用した場合は著者に含めたり，利用したことをAcknowledgementsなどに表記したりする必要があるそうです。ここらへんのポリシーは英語が第一言語ではない研究者の苦労に対する理解が欠けているような気もしますが，今後のポリシーがどう展開するのかに注意を払いつつ，うまく活用していきたいですね。