📄 Extracting Imperatives from Wikipedia Article for Deletion Discussions

Summary

Wikipedia editors decide the fate of contested articles in open "Article for Deletion" (AfD) discussions, where many comments push back against other participants. The authors treat one kind of pushback, directive speech acts from Searle's taxonomy, as attacks in Dung's abstract argumentation sense. For example, a request for a source challenges a claim, and the addressee can undercut that challenge by supplying the evidence. They focus on directives expressed as imperatives. These typically have an action-verb predicate, a second-person subject that is usually omitted, and optional politeness words or adverbs. They also mention cohortatives ("let's …"). AfD examples show how varied directive phrasing is, including modal obligations, hedged suggestions and "how about" questions. For this reason the task is limited to direct imperatives, and indirect ones are left for later work. The extracted text ends at the heading of the detection section, so the actual method and any evaluation numbers are not available here. The paper's value is as an early step toward mining argumentative attacks from collaborative online deliberation.

Key points

• Two-page paper (pp. 106–107) published in the proceedings of the inaugural argumentation mining workshop at ACL 2014, held in Baltimore. It is supported by NSERC Discovery.
• Theoretical grounding:
  • Austin's speech acts.
  • Searle's five illocutionary subgroups, with Directives as the target.
  • Ervin-Tripp's six directive types, one of which is the imperative.
  • Dung's attacks and Pollock's undercutting.
• The data domain is Wikipedia AfD discussions. Related AfD studies cited are Schneider et al. (2013) on novice vs. experienced arguers and Xiao and Askin (2014) on rationale types.
• Scope is restricted to direct imperatives. Indirect or polite forms are deferred because they are hard to capture with rules.
• Marsi's (1997) imperative-mood definition is considered too narrow. Feltrim et al. (2006), working on Brazilian Portuguese abstracts, give no detail on how imperatives were recognised.

Interesting ideas

• I1. Treat a directive as an argumentative move: a request for evidence counts as an attack, and complying with it (e.g., giving sources) counts as an undercut, so dialogue turns form an attack graph.
• I2. Recognise imperatives by structure rather than keywords: an action verb as predicate, a missing or implicit second-person subject, and possibly politeness markers or adverbial modifiers.
• I3. Separate cohortatives (first-person plural, "let's") from ordinary imperatives, since the speaker includes themselves in the requested action.
• I4. Deliberately split direct from indirect directives (modal obligations, "how about…", hedged suggestions) and solve the easy direct case first.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): A speech-act frame for directives could carry "challenges claim X" and "satisfied by act Y" slots, letting Mova explain how a reply resolves a request.
• I2 (tree): On UD trees this maps to a root VERB in base form without nsubj, possibly with "please" as a discourse/advmod dependent, which gives an explainable imperative detector.
• I3 (frames): The agent role of a "let's" directive should be resolved to speaker plus addressees rather than addressee only, which affects who-does-what in the frame.
• I4 (frames): Mova could tag the surface imperative reliably and pass indirect forms to a separate implicature layer that reasons about them explicitly.