Source-linked AI summary
CoAuthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities
Mina Lee, Percy Liang, Qian Yang
TL;DR
Because LM capabilities vary across interaction contexts and are difficult to interpret, the paper proposes curating and analyzing large interaction datasets for interaction design. It presents CoAuthor, which captures writing interactions with GPT-3 and supports analysis of language, ideation, and collaboration capabilities under varied definitions of good collaboration.
Problem
LM capabilities are highly context-dependent and difficult to grasp, limiting understanding of what they can and cannot do across interaction contexts.
Method
The paper curates and analyzes CoAuthor, a large interaction dataset capturing writers’ creative and argumentative writing interactions with GPT-3.
Results
CoAuthor addresses questions about GPT-3’s language, ideation, and collaboration capabilities and its contribution as a writing collaborator under various definitions of good collaboration.
Takeaways & Limitations
Large interaction datasets can support more incisive examinations of LM generative capabilities and more principled discussions of their promises and pitfalls in interaction design.
Takeaways & Limitations
Prior interaction-logging studies often use restricted settings, adapted models, or specific tasks, making generalization to other tasks and configurations difficult.
Abstract
from arXiv · showhide
Large language models (LMs) offer unprecedented language generation capabilities and exciting opportunities for interaction design. However, their highly context-dependent capabilities are difficult to grasp and are often subjectively interpreted. In this paper, we argue that by curating and analyzing large interaction datasets, the HCI community can foster more incisive examinations of LMs' generative capabilities. Exemplifying this approach, we present CoAuthor, a dataset designed for revealing GPT-3's capabilities in assisting creative and argumentative writing. CoAuthor captures rich interactions between 63 writers and four instances of GPT-3 across 1445 writing sessions. We demonstrate that CoAuthor can address questions about GPT-3's language, ideation, and collaboration capabilities, and reveal its contribution as a writing "collaborator" under various definitions of good collaboration. Finally, we discuss how this work may facilitate a more principled discussion around LMs' promises and pitfalls in relation to interaction design. The dataset and an interface for replaying the writing sessions are publicly available at https://coauthor.stanford.edu.
1 INTRODUCTION
The paper argues that interaction datasets can make context-dependent LM capabilities more accessible for interaction design. It exemplifies this approach with CoAuthor, which records GPT-3-assisted creative and argumentative writing and supports analysis of language, ideation, and collaboration.
- Motivation: LM capabilities vary with interaction context, making it difficult to determine when and how GPT-3 can support writers.Relevant factors include preceding text and decoding parameters, while interactive writing adds iterative turns and edits that complicate attribution of the model’s contribution.
- Contribution: CoAuthor is a dataset for revealing GPT-3’s generative capabilities in interactive creative and argumentative writing.It is presented as an HCI-oriented approach to curating and analyzing large interaction datasets.
- Dataset: 63 writers and four GPT-3 instances produced rich interactions across 1445 writing sessions.Each session begins with a prompt and includes free writing, requests for suggestions, acceptance or dismissal, and editing in flexible order.
- Results: CoAuthor supports questions about GPT-3’s language, ideation, and collaboration capabilities.The analysis also examines GPT-3’s contribution as a writing collaborator under different definitions of good collaboration.
- Availability: The dataset and replay interface are publicly available for examining the writing interactions.The interface is intended to give designers a felt understanding of the sessions.
- Implications: The paper frames interaction datasets as a basis for more principled discussion of GPT-3’s promises and pitfalls in interaction design.This is presented alongside contributions concerning holistic capability understanding and dataset-based HCI research.
2 RELATED WORK
Related work identifies a gap between holistic, process-sensitive understanding of LM capabilities and methods or benchmarks focused on restricted contexts or automated results. It motivates datasets that cover diverse contexts, preserve interactions, and support multiple interpretations.
- Language Models’ Generative Capabilities: LMs can generate fluent text across many domains but still produce linguistically flawed, factually incorrect, or ethically problematic text.The paper therefore states that guardrails are necessary when using these models.
- Challenges: Understanding interactive-writing capabilities is difficult because LM behavior is highly context-dependent and capabilities can be subjectively interpreted.The paper notes that covering all contexts and characterizing success or failure is effectively intractable.
- Traditional Methods: Contextual inquiry offers rich, specific insights, but findings from one model and writing context may not generalize to others.The cited example involved four professional novelists writing fictional stories with GPT-2.
- Traditional Methods: Interaction logging covers relatively diverse tasks and writers while enabling fine-grained analysis, but prior studies often used restricted settings, adapted models, or task-specific interfaces.These choices make generalization to other tasks and configurations of the same model difficult.
- Dataset Approach: The dataset approach can cover diverse contexts and accommodate subjective interpretations of what constitutes good interaction.Examples include datasets designed for varied user populations and datasets that avoid imposing a single definition of good interaction.
- Datasets in NLP: NLP datasets are reusable and expandable for evaluating models across tasks, but typically assume full automation rather than user-guided augmentation.Consequently, they often capture final results instead of the writing process.
3 DESIGNING DATASETS FOR HCI
The paper proposes four desiderata for large interaction datasets that reveal LM capabilities for interactive writing and make them more accessible to interaction design.
- Desiderata: Datasets should cover diverse writing tasks, prompts, and writers because LM capabilities depend strongly on context.This breadth is intended to account for context-dependent behavior.
- Desiderata: Datasets should support subjective interpretations by avoiding a single imposed label or metric.Designers should be able to extract meaning and analyze interactions according to their own design goals.
- Desiderata: Datasets should capture writing processes rather than only final results.The passage presents process capture as a core requirement for interactive-writing analysis.
4 DESIGNING COAUTHOR
CoAuthor was designed as a reusable dataset for studying interactive writing across diverse contexts while preserving rich interaction processes and supporting multiple interpretations of collaboration.
- Cover diverse contexts: CoAuthor covers creative and argumentative writing across multiple prompts and writers to capture context-dependent language-model capabilities.Creative writing emphasizes imagination, character development, narrative, and plot; argumentative writing involves evidence and establishing a position.
- Support subjective interpretations: The dataset supports multiple definitions of good collaboration through measures of writing outcomes, writer perceptions, and writer behavior.Measures include session time, length, queries, acceptance rate, writer-written proportion, surveys, and interaction patterns.
- Capture processes, not just results: CoAuthor records interactions as detailed events and deterministic, non-overlapping event blocks for subsequent analysis.Events include text edits, cursor movement, suggestions, and suggestion acceptance or dismissal; event blocks abstract event sequences such as text-insert(a) text-insert(b) into insert(ab).
- Allow for possible reuse and expansion: Each writing session pairs a prompt with a model instance and decoding parameters, allowing subsets or additional sessions to be collected for future design goals.GPT-3 was used without fine-tuning, while decoding parameters controlled suggestion randomness.
- System and interface: The data-collection interface was a text editor supporting ordinary editing and cursor interactions, with writers requesting five suggestions by pressing Tab.Suggestions could be accepted with the keyboard or mouse, or dismissed by continuing to type, clicking outside the popup, or pressing Escape.
- Study procedure: The study recruited 63 main-round writers, required at least ten minutes of writing per prompt, and collected post-session surveys about capabilities and experience.Writers could respond to each prompt up to five times or skip it; surveys addressed fluency, idea pooling, enhanced quality, perceived capabilities and limitations, ownership, satisfaction, and reuse.
- Dataset overview: 830 stories and 615 essays were collected; sessions averaged 418 words, 11.8 system queries, a 72.3% acceptance rate, and 72.6% writer-written text.These statistics summarize the resulting creative and argumentative writing sessions.
5 DEMONSTRATING USES OF COAUTHOR
CoAuthor demonstrates how interaction data can reveal GPT-3’s language, ideation, and collaboration capabilities in creative and argumentative writing. The analyses show strong variation across prompts, writers, and decoding settings, while linking interaction patterns to productivity and ownership.
- Language Capabilities: Ability to Generate Fluent Text.: GPT-3-generated sentences had fewer spelling and grammar errors than writer-authored sentences, although grammaticality varied across prompts.Errors per word were 0.037 ± 0.001 for writers, 0.033 ± 0.001 for GPT-3, and 0.032 ± 0.001 for both.
- Language Capabilities: Ability to Generate Fluent Text.: Combined writer-and-GPT-3 sentences achieved higher vocabulary diversity than sentences written by either writers or GPT-3 alone.Distinct-2 scores were 0.884 ± 0.001 for writers, 0.876 ± 0.001 for GPT-3, and 0.923 ± 0.001 for both.
- Ideation Capabilities: Ability to Generate New Ideas.: 13% of accepted creative-writing suggestions and 7% of accepted argumentative-writing suggestions contained new named entities.Among those entities, writers subsequently reused 20% in creative writing and 14% in argumentative writing.
- Ideation Capabilities: Ability to Generate New Ideas.: Higher randomness increased the frequency of new named entities in accepted suggestions, but reuse rates were similar under high and low randomness.In creative writing, entity frequency was 15 ± 1% versus 10 ± 0%; reuse was 19 ± 2% versus 22 ± 2%.
- Collaboration Capabilities: Ability to Work Jointly.: Collaboration equality and mutuality varied greatly across writers but less across prompts.For example, Writer #1 scored around 0.45 and 0.75, while Writer #2 scored below 0.1 and 0.2 for equality and mutuality.
- Design Implications: Queries and accepted suggestions correlated more strongly with total text production than time spent writing, while writer-authored text correlated with ownership.Correlations with production were 0.42 for queries, 0.48 for accepted suggestions, and 0.29 for writing time; edits showed no meaningful ownership correlation.
6 DISCUSSION
The paper positions large interaction datasets as boundary objects that connect HCI and NLP perspectives on LM capabilities. CoAuthor supports replay-based hypothesis formation and interactive-writing model research while retaining diverse, human-centered interpretations.
- Datasets can bridge HCI and NLP by enabling shared understanding of LMs’ generative capabilities.
- CoAuthor makes interaction dynamics tangible through replay, helping researchers formulate hypotheses about writers’ changing preferences.Replay can reveal whether writers prefer suggestions at the beginning or throughout writing.
- Researchers can use CoAuthor to check whether linguistic accommodation occurs before pursuing targeted empirical investigation.The proposed phenomenon concerns writers’ and GPT-3’s style, voice, or tone influencing each other over time.
- CoAuthor can support interactive-writing models because accepted suggestions provide in-situ generation and evaluation data.The paper proposes that fine-tuning on accepted suggestions may produce suggestions writers find more desirable.
7 CONCLUSION
The conclusion identifies a need for better understanding of LMs in interaction design and presents CoAuthor as one viable dataset-based approach. It reports the dataset’s scale and encourages researchers to reuse and extend it for different design goals.
- Curating and analyzing large interaction datasets is presented as a viable approach to studying LMs for interaction design.
- CoAuthor contains interactions between 63 writers and four GPT-3 instances across 1445 writing sessions.
- The authors encourage researchers to use, analyze, and extend CoAuthor according to their design goals and research perspectives.
A.1 Server
The server is implemented in Python with Flask and supports suggestion requests and event saving for writing sessions. It runs on the authors’ institutional infrastructure.
- The server uses Python and Flask to handle suggestions and save writing-session events.
- The system was deployed on the authors’ institutional infrastructure.
- The frontend communicates with the server both to request suggestions and to save session events.
A.2 Decoding parameters
GPT-3 suggestions were generated with the davinci engine using fixed decoding settings and several temperature, frequency-penalty, and response-length configurations. The listed settings include a 30-word-piece response length and top-p of 1.
- The experiments used GPT-3’s davinci engine with a 30-word-piece response length and top P of 1.
- Frequency penalty varied across 0, 0.5, and 1, while presence penalty and best-of remained 0 and 1.
A.3 Suggestions
The study used ten prompts each for creative and argumentative writing, with prompts sourced from WritingPrompts and The New York Times and minor modifications. Writers could request repeated sessions on prompts, skip prompts, and receive filtered first-sentence suggestions from GPT-3.
- Suggestions: GPT-3 outputs were parsed into sentences, and only the first sentence was shown as a suggestion after filtering swear words, duplicates, and empty strings.Although GPT-3 was queried five times, filtering sometimes left fewer suggestions for writers.
- Prompt selection: Ten prompts were used for both creative and argumentative writing, sourced from WritingPrompts and The New York Times with minor modifications.Argumentative prompts added “In my opinion,” to signal that GPT-3 should begin responding rather than continue the prompt.
- Prompt selection: Each writer could work on a prompt up to five times, skip prompts, or move to another prompt after one or two sessions.Some writers repeatedly used prompts, while others skipped particular prompts entirely.
- Suggestions: The survey examined writer information, collaborative-writing benefits, perceived LM capabilities and limitations, and overall experiences.Questions addressed competence, creative and persuasive writing ability, understanding of writers’ intentions, and aspects of suggestions needing improvement.
D QUALIFICATION ROUND
The qualification round recruited and screened crowd workers for collaborative AI writing, requiring a clear ending or conclusion and at least ten minutes of interaction. Participants used an AI writing interface, completed self-evaluations and surveys, and received payment for participation.
- D QUALIFICATION ROUND: Qualification candidates were recruited through Amazon Mechanical Turk and screened for a greater-than-97% approval rate, United States location, and more than 10,000 approved HITs.The authors manually rated 201 submissions and qualified writers whose work received a 4 or 5.
- D QUALIFICATION ROUND: Candidates had to produce a story with a clear ending or an essay with a clear stance and conclusion while collaborating with the system for at least ten minutes.Writers who did not interact with the system were automatically disqualified.
- E INSTRUCTIONS: The qualification task taught participants to write collaboratively with AI by accepting or editing suggestions and producing a short, interesting story with a clear ending.The main round used an identical story-writing task with potentially different prompts.
- E INSTRUCTIONS: Participants could request suggestions with the Tab key, select them with clicks or keyboard controls, and revise their story or suggestions.The interface displayed suggestions at the end of the story and allowed repeated requests for more ideas.