[英語教學資源中心寫作工作坊] 以咖啡、鹽、計算機為暗喻 細說亦敵亦友的大型語言模型（LLMs）
Check out the news!
英語教學資源中心於2024年4月25日（四），在政大指南校區達賢圖書館的達賢講堂（實體講座）及羅家倫講堂（同步直播）舉辦寫作實踐工作坊，講題為「人工智慧輔助課堂學習：好處、壞處與迷思」。此次工作坊邀請到國立臺灣大學寫作教學中心兼任講師歐墨然先生（Mr. Graham Oliver）擔任主講嘉賓，分析ChatGPT等以人工智慧為根據的大型語言模型（LLMs）運用於英語寫作中的優勢與限制。本活動由英語教學資源中心鍾曉芳主任進行開場，與會者在講座期間，可透過掃描QR code連結至大會平台提出問題及分享意見，與講者進行即時互動，此外，講者也親自與兩個講堂的與會者進行面對面交流，以加強他們對主題的理解和應用能力。
歐墨然先生認為，現今大眾所耳熟能詳的「AI」一詞，本質上過於籠統及容易混淆，因此他選擇以較適切的「大型語言模型」（LLMs），來稱呼如：ChatGPT、Google Gemini及Claude等的程式。除了明確其定義以外，歐墨然先生也開宗明義地指出這次工作坊的焦點：何時使用LLMs、何時不應使用LLMs，以及使用時需多加留意的注意事項。他也提出幾個描述LLMs的暗喻：一、咖啡——使用LLMs是否得宜，將取決於我們如何使用它們；二、鹽粒——對於LLMs生成的內容，我們務必謹慎處理；三、計算機——LLMs在日常生活中是相當實用且便利的工具，但若要有效運用LLMs，使用者必須建立堅實的知識基礎；四、亦敵亦友的編輯——通常LLMs能提供內容編輯上良好的建議，然而在某些情況下，這些建議並不如我們預期的有建設性，歐墨然先生更進一步引述「幻覺」（hallucination）一詞，來強調LLMs產生的資訊，有可能表面正確，卻其實不然。
當我們在寫作中利用LLMs時，歐墨然先生也強調某些事項需列入考量，例如：在ChatGPT和Google Gemini等常見的模型中，前者傾向產出品質較佳的作品，而後者則傾向貼近我們給予的指令。在輸入指令時，我們需要考慮的資訊如：目標受眾、寫作或口說文體、語氣及語境。在工作坊中，歐墨然先生也現場邀請觀眾參與任務，透過修改作者的性別、年齡及職位等指令，以檢視產出的作品是否有所不同，並且進一步分享其異同。而在比較不同階段產出的作品時，歐墨然先生也反覆叮嚀觀眾：「應成為主動的使用者，而非被動的使用者！」也就是說，我們應在產生內容的同時，不斷檢視不同版本中的佳句並排除相對較差的句子，並且透過修改指令，創造出產生截然不同作品的可能性。
除了介紹如何使用LLMs來生成優質的內容，歐墨然先生也討論到其工具模型可能帶來的隱憂。例如：ChatGPT可能是適合的編輯，但它不一定是好的老師，言下之意為，雖然它可以修正語法上的錯誤，但它未必能夠解釋其背後正確的原理，因此歐墨然先生說道：「當我們決定使用LLMs產出內容時，良好的修正技巧是不可或缺的。」此外，LLMs產出的文句中，即便呈現的文法皆為正確，但通常充斥過量華而不實且不自然的語彙，並且在不同主題的文章裡，句型結構會出現內容冗長且重複性高的問題，而這些情當今仍需使用者多加留意。
歐墨然先生也提及，LLMs較擅長改善文章的結構（例如：前言、總結），而在改善文章的內容方面，則可能會流於表面或誤導，甚至帶有偏見。部份與會者也觀察到了這一點：他們在自己的LLM指令中，做出性別、代名詞或形容詞方面的改變以後，顯示出來的內容也有所不同，例如：把作者的身份從女性研究者改為男性研究者時，LLMs便產出更加專業的描述。然而，歐墨然先生對此並不感到驚訝，因為網路上本身就存在著帶有偏見的內容，況且帶有偏見的內容，也不只限於性別差異，還可能存在於其他方面。他也進一步指出，就學術倫理及資訊來源方面，LLMs產出的內容更有機會導致不經意的剽竊行為，這是因為LLMs會在未經原作者同意且LLM使用者也未必知情的情況下，自動擷取網路上的資訊原文，而這方面的危機，存在於越精細的主題裡，就越有機會出現。要避免這種情況，歐墨然先生建議大家在LLMs輸入指令時，應提供更多與自身有關的內容與細節，否則每一個人產生的內容，都有可能極度相似。
即便LLMs仍有待改善的缺點，例如它們雖能在某些情況下給予幫助，卻也有可能暗藏隱憂，但若我們能夠成為「主動的使用者」，則講者對於LLMs在未來的使用上仍持樂觀態度。他在工作坊的最後，總結了一些能夠讓LLMs幫助我們寫作的實用技巧。首先，我們嘗試記錄每一次LLMs所產出之較優良的內容（通常大約10%到20%），再在下一次使用LLMs的時，貼上相關內容。第二，我們需多加留意現今的LLMs只能針對比較廣泛討論的議題作出精細的分析；若我們要求LLMs同時作出細緻的學術分析和個人反思，則有可能事倍功半。第三，我們應當嘗試從LLMs裡看出瑕疵，並對其作出評價。第四，除了廣泛應用的ChatGPT以外，我們也可以考慮其他LLM工具，例如Elicit、Bing 和 Semantic Scholar。最後，相對於例如Google翻譯（Google Translate）之類常用的翻譯網站，LLMs在翻譯不同的中文詞彙和成語時，比較能夠指出當中的細微差異與適用情況。
本中心萬分感謝歐墨然先生精彩的分享，讓身為老師或學生的與會者對將LLMs運用在寫作及課堂教學前所需要具備的基礎知識有了更深入的瞭解。在2024年5月30日（四），歐墨然先生將會主持另外一場關於同儕評論寫作實踐工作坊，還有英語教學資源中心介紹網路修改指引軟體Writefull之使用（報名連結：https://moltke.nccu.edu.tw/Registration/registration.do?action=conferenceInfo&conferenceID=X22917）。我們期待他的下一次工作坊！
Metaphors of Coffee, Salt, and Calculator: Exploring Strengths, Weaknesses and the Frenemy of Large Language Models (LLMs)
A hands-on writing workshop, titled “Strengths, Weaknesses and Confusion: AI-Assisted Learning in the Classroom”, featuring Mr. Graham Oliver, Adjunct Instructor of the Academic Writing Education Center at National Taiwan University, was organized by the EMI Resource Center at National Chengchi University on April 25th 2024 (Thursday). This event, led by Professor Siaw-Fong Chung, the center director, was held in Dah Hsian Lecture Hall (face-to-face) and Luo Jialun Lecture Hall (through live streaming), Dah Hsian Library on the NCCU Zhinan campus, providing an insightful account of the advantages and limitations of using large language models (LLMs), such as ChatGPT, in English writing. Interaction was facilitated through QR codes, enabling participants to interact with the speaker by raising questions. The speaker further interacted in person with participants in both lecture halls, further enhancing their understanding and application of the topic.
Instead of the commonly-known term “AI”, which he regarded as overly broad and confusing, Mr. Oliver preferred a more precise one “large language models” (LLMs) when referring to programs such as ChatGPT, Google Gemini and Claude. Along with the definition, Mr. Oliver started the workshop with clear focuses: when to use LLMs, when not to, and what to watch out for while using them. He also proposed several metaphors describing LLMs: first, coffee – whether LLMs are good to use depends very much on how we are using them; second, a grain of salt – we need to be careful believing what we are told by LLMs; third, calculator – just like a calculator, LLMs are a useful and convenient tool in our daily life, but we need a lot of foundational knowledge to use them well; fourth, a frenemy editor – most of the time LLMs give us good advice, but in some situations they may not be so constructive as we imagine. Mr. Oliver further cited the word “hallucination” to highlight the possibility of the emergence of untrue, while true on the surface, information by LLMs.
Mr. Oliver emphasized some considerations when we need to make use of LLMs in writing. Comparing common models like ChatGPT and Google Gemini, he concluded that the former tends to produce works of a slightly better quality while the latter tends to better follow our instructions. When it comes to what to include in the prompts, details about the target audience, the spoken or written mode, the tone and the context have to be taken into account. Throughout the entire workshop, Mr. Oliver invited the audience to engage in various tasks and to share their production, through making adjustments to the prompts in terms of writer’s gender, age, and job title, etc., purposefully to see how the results differed. During the comparisons of production in various stages, Mr. Oliver repeatedly reminded the listeners to become active users, instead of passive ones – saving especially good sentences and discarding those not working from later versions. Great differences in writing products may emerge once adjustments to instructions and guidance are administered.
Besides introducing means to achieve better writing, Mr. Oliver, on the other hand, discussed some pitfalls of LLMs to improve one’s production. For example, ChatGPT may be a suitable editor, but may not be a good teacher: it could make changes to language errors, but might not be able to explain the principles behind its edits consistently. Therefore, Mr. Oliver deduced that good proofreading skills are indispensable when one decides to rely on LLMs to produce writing. In addition, LLMs are found to exhibit overuse of grandiose and unnatural vocabulary in almost every single sentence, repetition of the same long sentence structures throughout a discourse (although all grammatically correct) and similarities across a variety of topics.
LLMs, according to Mr. Oliver, are better in improving the structure (such as introduction and conclusion), instead of the content of written works – superficial, incorrect, misleading and even biased information might arise. Some members in the audience expressed having noticed such a feature as well when they made changes in writer’s gender, pronouns or adjectives in their instructions to LLMs – a result showing more professionalism was exhibited when a participating professor changed the identity of the writer from a female researcher to a male one. This was not surprising, responded by Mr. Oliver. Prejudiced content is prevalent online, so bias in the generation of LLMs may not be surprising. Even worse situations in content might occur in ethics and sourcing: accidental plagiarism might be committed since LLMs can lift exact sentences from various sources online without creator’s consent and user’s awareness. The more specific the writing topic, the higher the possibility of this danger may arise. To avert this problem, Mr. Oliver suggested students provide more personalized content and details, especially the ones about themselves, or else overly similar writing is produced for everybody, resulting in overly similar writing being produced for everybody.
Despite potential weaknesses and imperfections which might help us in some situations but hurt us in others, Mr. Oliver felt optimistic about the use of LLMs in the future if we can become active editors, and concluded the workshop with a couple of tips and tricks when deploying LLMs in a helpful way to achieve our writing goals. First, we should build upon what we tell LLMs in a single session, through saving the good results generated (probably around 10–20%) and pasting them each time we start a new session of using LLMs. Second, we have to notice that current LLMs can only do solid close analysis of very widely discussed texts; our effort might be reduced if LLMs are required to do specific tasks as well as process both academic analyses and personal reflection. Third, we have to be accountable to see and critique flaws from LLMs. Fourth, we might consider other LLM-powered tools, namely Elicit, Bing and Semantic Scholar, in addition to common ones like ChatGPT. Finally, LLMs might be a good means to provide subtleties, nuances and appropriateness in the translations of some Chinese phrases and idioms, compared to commonly-used sources like Google Translate.
We express our greatest appreciation for Mr. Oliver’s stimulating sharing, which let us, as teachers and students, understand what prerequisites are involved in using LLMs in writing and in the classroom. On May 30th 2024 (Thursday), Mr. Oliver will conduct another hands-on writing workshop in NCCU on the strategies of peer- and self-editing, along with an introduction of Writefull, an online software application which provides guidance on editing, by the EMI Resource Center (link for registration: https://moltke.nccu.edu.tw/Registration/registration.do?action=conferenceInfo&conferenceID=X22917). We look forward to his next workshop!
The EMI Resource Center：https://www.eminccu.com/
