📄 IMHO: An Exploratory Study of Hedging in Web Forums

Summary

The paper looks at hedging (language that softens or qualifies how firmly a claim is asserted) in a technical customer-support web forum rather than in scientific articles, where most automatic hedge detection has been done. The authors ask whether hedged posts are more likely to receive "kudos" (community usefulness votes) and whether hedge use differs across user ranks. Drawing on Hyland's functional taxonomy of hedges, they separate first-person singular epistemic phrases (e.g. "I think", "IMHO"), which they link to reader-oriented hedging, from non-phrasal lexical hedges ("seem", "suggest"). They tag both types automatically with two lexicons and maximal matching, then label each post with three discretization functions and four measures. Posts containing hedges attract more kudos on average, and posts without any hedge attract the fewest. Gurus and ranked users use epistemic phrases most often, while employees use them least. The authors conclude that first-person epistemic phrases are a reliable, automatable signal of stance, and that annotation work should prioritise them.

Key points

• Data: a software vendor's support forum with 172,253 posts (1,044,263 sentences), cleaned of HTML, URLs, emoticons and IPs, with user names anonymised. There are four user categories: employee, guru, notranked and ranked.
• Lexicons: 76 first-person epistemic phrases (drawn from Kärkkäinen and Wierzbicka, plus acronyms such as IMHO, IMO and AFAIK) and 109 non-phrasal hedges (from Rubin). Tagging uses longest match, so "I would suggest" wins over "suggest".
• Coverage: 107,134 posts contain at least one hedge. Of these, 34,301 contain an epistemic phrase, 101,086 a non-phrasal hedge, and 28,253 both.
• Method: posts are labelled with three discretization functions (presence, majority, and a combined "only/mostly" scheme), giving four measures m1–m4 based on raw counts, lexicon-size normalisation or word-count normalisation.
• Results:
  • Mean kudos are highest for posts dominated by epistemic phrases under m2, and lowest for posts with no hedges.
  • Differences in hedge-type proportions between user categories are significant (two-sample proportion test, p<0.05).
  • The frequency of "I" and "my" correlates negatively with kudos, so the effect is not simply first-person usage.

Interesting ideas

• I1. Restricting analysis to the subset of a fuzzy phenomenon that can be detected reliably (first-person epistemic phrases), instead of chasing every hedge, yields trustworthy counts from a cheap method.
• I2. Maximal (longest) matching of multiword hedge lexicons avoids double-counting a phrase and its inner word.
• I3. The same labelling question is run through several alternative discretizations and normalisations, keeping only conclusions that hold across all of them and treating the rest as measure-dependent.
• I4. A confound is checked directly: plain use of "I" and "my" is tested to rule out that first-person usage alone explains the kudos effect.
• I5. Hedges are mapped onto Hyland's functional classes (reader-oriented vs. writer-oriented vs. accuracy-related), so stance is typed by function rather than kept as a single uncertainty score.

For Mova

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

• I1 (frames): Speech-act and stance detection in prag could first anchor on high-precision patterns such as an "I think/believe" matrix clause with a 1sg subject, then mark the embedded clause's commitment level explicitly.
• I2 (tokens): Multiword expressions like "I would suggest" or "as far as I know" could be recognised as one unit before tagging, so later stages see a single stance marker.
• I3 (learn): Evaluations could report results under several scoring variants and flag findings that flip between them, as a robustness check alongside negative controls.
• I4 (learn): When a feature seems predictive, a matched control feature (the surface pronoun without the epistemic verb) can be built in to show the effect comes from the construction itself.
• I5 (world): Asserted facts in the text world could carry a commitment tag (speaker-held belief, disclaimed, or approximate), letting the answerer explain how certain the source was.