One may ask why a frame semanticist would use a chatty AI instead of the other, established tools— such as [Word Sketch](https://www.sketchengine.eu/guide/word-sketch-collocations-and-word-combinations/), or any systems developed with the goal of inferring lexical units, or LLMs available in repositories like the one offered by [Hugging Face](https://huggingface.co/)? The answer is that new technologies, such as [ChatGPT](https://chat.openai.com) and [OpenAssistant](https://open-assistant.io), produces very chatty, intuitive discourse. Because it works as a chatbot, those AIs requires no coding knowledge from the linguist. Elaborate set-up and extensive querying of those chatbots are both completely code-free. The linguist who wants to expand FrameNet can directly prompt one of those tools to work as a co-pilot, as an eliciting interlocutor, as an assistant in the task, and can instruct the AI in how to answer. Consider the dialogue sequence in (1-8) with ChatGPT as an example. [1] User: ``` The conceptual frame for "running a risk" contains the following core frame elements: asset, action, bad outcome, risky situation, perpetrator. It is evoked by words such as the verb "risk," the noun "risk," and the verb "dare." What would be the core elements in the conceptual frame for "protecting" and which words would evoke the conceptual frame for "protecting"? ``` [2] ChatGPT: ``` The conceptual frame for "protecting" might have the following core frame elements: object or person being protected, threat or danger, protector. Words that might evoke this conceptual frame could include the verb "protect," the noun "protection," and the adjective "protective." ``` [3] User: ``` What other words evoke the protecting frame? ``` [4] ChatGPT: ``` Other words that might evoke the "protecting" frame could include words such as "defend," "shield," "guard," "safeguard," and "preserve." These words all convey the idea of protecting someone or something from harm or danger. ``` [5] User: ``` Give a sentence that evokes the "protecting" frame and say which phrases in the sentence evoke which parts of the frame. ``` [6] ChatGPT: ``` "The mother bear fiercely guarded her cubs from the approaching wolf." In this sentence, the phrases "mother bear," "fiercely guarded," and "cubs" all evoke the core frame elements of the "protecting" frame. The mother bear is the protector, the cubs are the objects being protected, and the approaching wolf is the threat or danger. The verb "guarded" also evokes the frame, as it conveys the idea of protecting something. ``` [7] User: ``` What other frames can you think of that resemble “running a risk” and “protecting”? ``` [8] ChatGPT: ``` Other frames that might resemble "running a risk" and "protecting" could include "taking a chance" and "safeguarding." The frame for "taking a chance" might contain core frame elements such as the opportunity or situation being pursued, the potential reward or outcome, and the individual or entity taking the chance. The frame for "safeguarding" might have core frame elements such as the thing being safeguarded, the potential danger or threat, and the individual or entity responsible for safeguarding it. These frames would be evoked by words such as "chance," "safeguard," and "safeguarding." ``` In the preceding sequence, we asked ChatGPT to propose a frame that already exists in FrameNet, so as to be able to compare its answer to one provided by trained frame semanticists. The ***Protecting*** frame, as defined in FrameNet, is shown here: ![[protecting.png]] We found ChatGPT's answers in [2], [4] and [6] impressive, when compared to the expert answers presented in the figure. However, the remainder of the dialogue, in which we asked ChatGPT to suggest more frames resembling ***Running_a_risk*** and ***Protecting***, was merely a paraphrase of the previous frames already present either in our prompt or in its answer to the prompt.  After our first, intuitive attempts to use ChatGPT as a talkative co-pilot for linguists, we developed a more structured and controlled test of whether ChatGPT (and by extension conversational AIs in general) can help the linguist. The LLM had built computational representations in response to immense training on linguistic data. We were curious to see whether it could turn those representations into information useful to the linguist. In other words, we tried to engineer prompts, and even script prompts, that would lead ChatGPT to generate text that the linguist would judge to be not only intelligible but also helpful for expanding FrameNet. If you are interested in another analysis using chatty AIs, visit [[LLMs on the English Comparative Correlative]]. To check other prompts, visit the [[Prompt Repo]].