TL;DR
This paper investigates how and why different grammatical forms of negative imperatives (like "Don't do X" vs. "Be careful not to do X") are used in instructional texts. By analyzing a corpus of real-world instructions, the authors identify correlations between the intended function of these warnings and their linguistic realization, offering insights for both understanding and generating clearer instructions.
What Problem Does This Paper Address?
When we follow instructions, we don't just need to know what to do, but also what not to do, especially when an incorrect action could be dangerous or lead to undesirable outcomes. While positive imperative commands ("Do this") have been well-studied, the linguistic strategies for conveying negative constraints – preventing unwanted actions – have received less attention. This paper focuses on "negative imperatives," which are expressions used to warn users against performing certain actions. The core problem is understanding the linguistic choices involved in these negative instructions: why do authors choose "Don't sand it" over "Be careful not to sand it," or vice versa? This understanding is crucial for improving how instructions are written and how intelligent systems can generate them.
Methods
The authors conducted a corpus study, analyzing a collection of instructional texts. They developed a coding schema to categorize negative imperatives based on two key aspects:
- Form: This refers to the grammatical structure of the negative imperative. They distinguished between:
- DONT imperatives: Characterized by "do not" or "don't" (e.g., "Don't sand it").
- Neg-TC imperatives: Characterized by phrases like "take care," "be careful," "make sure," etc., followed by a negative infinitive (e.g., "Be careful not to crease the wallpaper").
- Function: This captures the semantic and pragmatic purpose of the imperative. They defined two functional features:
- Intentionality: Whether the author believes the user is consciously intending to perform the action to be prevented (CONscious) or if the action is more likely to be accidental or overlooked (UNConscious).
- Awareness: Whether the author believes the user is aware that the consequences of the action are bad (AWare) or unaware (UNAware).
They then used statistical analysis, specifically chi-squared tests and Kappa statistics for inter-coder reliability, to find correlations between these form and function features.
Datasets
The study used a corpus of instructional texts collected opportunistically from the internet and other sources. This corpus, approximately 4 MB in size, included:
- Recipes
- Do-it-yourself manuals
- Computer game instructions
- Sun Open-windows online instructions
- Administrative application forms
The initial corpus was filtered to create a final sample of 239 negative imperative examples that were coded for the study. No new benchmark datasets were introduced; rather, an existing collection of texts was curated and analyzed.
Key Insights and Ideas
- Intentionality Matters: The study found a strong tendency to use "DONT" imperatives when the author believes the user is likely to consciously choose the action to be prevented. Conversely, "Neg-TC" imperatives are more likely when the author perceives the action as something the user might do unconsciously or overlook.
- Awareness is Key: There's a clear preference for "DONT" imperatives when the author believes the user is unaware of the negative consequences of an action. The "Neg-TC" form is favored when the user is presumed to be aware of the bad outcomes.
- Functional Features Drive Form: The research strongly supports the hypothesis that the author's perception of the user's intentionality and awareness of negative consequences significantly influences the choice of grammatical form for negative imperatives.
- Methodological Rigor: The paper emphasizes the importance of using robust statistical measures like the Kappa statistic for assessing inter-coder reliability, especially for subjective functional features, moving beyond simple percentage agreement.
Key Results
The statistical analysis revealed significant correlations between the functional features (Intentionality and Awareness) and the form of negative imperatives.
- Intentionality: The chi-squared statistic was 51.4 with a significance level of 0.001, indicating a strong correlation. The contingency table showed that "DONT" imperatives were predominantly used for conscious actions, while "Neg-TC" were more likely for unconscious ones. Notably, there were no examples of "CONscious" actions coded as "Neg-TC."
- Awareness: The chi-squared statistic was 56.9, also with a significance level of 0.001, confirming a strong correlation. The results indicated a preference for "DONT" when the user is presumed "UNAware" of negative consequences, and "Neg-TC" when the user is presumed "AWare."
- Inter-coder Reliability: Kappa values indicated "substantial" agreement for Awareness (0.75) and "moderate" agreement for Intentionality (0.51), suggesting the coding schema was applied consistently. The "form" feature had perfect reliability (Kappa = 1.0) due to its objective, syntactic nature.
Limitations and Future Work
- Manual Feature Assignment: The authors acknowledge that currently, the functional features (intentionality and awareness) are assigned manually by users of their generation system (DRAFTER). They note that these features are "too difficult to derive automatically."
- Future Work: The authors plan to augment and refine their list of features and use them in understanding applications. They also intend to extend their analysis to "ensurative expressions" (e.g., "Make sure to...") and explore automatic derivation of these functional features.
Summary updated