On the negativity of negation

Christopher PottsProceedings From Semantics and Linguistic Theory 2010

Abstract

Natural language negation is persistently negative in the pragmatic sense, and emphatic and attenuating negative polarity items modulate this effect in systematic ways. I use large corpora of informal texts with meta-data approximating features of the context to characterize this pragmatic negativity, and I attempt to explain it in terms of the ways in which negative sentences engage the questions under discussion. The discussion highlights some of the ways in which quantitative corpus methods can be used to achieve novel results in linguistic pragmatics.

Introduction

In formal semantics, the traditional Fregean view treats negation as a "benign reverser": a neutral operator that simply flips a truth value. Under this view, "It is not raining" is merely the logical opposite of "It is raining," carrying no inherent emotional weight. However, linguistic intuition suggests otherwise. Negation often feels "unpleasant," frequently associated with falsity, absence, or even "evil." Christopher Potts addresses this discrepancy by quantifying "pragmatic negativity"—the tendency for negation to be statistically biased toward negative sentiment and reactive contexts. By analyzing millions of online reviews and social media posts, this research demonstrates that negation is a "downer" in real-world conversation, a phenomenon modulated by Negative Polarity Items (NPIs).

Methodology and Data

To move beyond intuition, Potts employs "quantitative corpus pragmatics," correlating linguistic tokens with contextual meta-data across several large-scale corpora.

Analytical Techniques

  • Conditional Probability Distributions: Rather than simple frequency counts, the study calculates the probability of a specific rating category given the presence of a word ($Pr(c|w)$). This allows for direct comparison between words with vastly different raw frequencies.
  • Clause-Level Parsing: Using the Stanford Parser and Tregex, negation patterns are extracted within their local syntactic contexts to ensure precise analysis (e.g., distinguishing "not good" from "good" appearing elsewhere).
  • Dialogue Act Analysis: The Switchboard corpus is used to correlate negation with specific conversational moves, such as "rejecting" or "accepting" proposals.

Datasets

The study utilizes diverse, meta-data-rich corpora, most notably the IMDB dataset containing 1.36 million movie reviews with 1–10 star ratings. Other sources include the Experience Project (confession-style texts with reader reactions), Multi-domain reviews (Amazon, TripAdvisor, etc.), and the Convote corpus of Congressional speech transcripts. Cross-linguistic consistency is tested using German and Japanese Amazon reviews.

Conditional probabilities for negation across rating categories in the IMDB dataset.

Pragmatic Negativity and the Question Under Discussion

Potts argues that negation is pragmatically negative because it is inherently reactive. It typically serves as a response to a "Question Under Discussion" (QUD). One rarely asserts that something isn't the case unless the context or a previous speaker suggested it was. This makes negation fundamentally "clashy" or corrective.

Furthermore, positive statements are generally more "resolving" by narrowing down possibilities, whereas negative statements leave many alternatives open. This lack of resolution contributes to the sense that negation is used when expectations are not met.

The NPI "Volume Knob"

Negative Polarity Items (NPIs) modulate this pragmatic negativity:

  • Emphatic NPIs (e.g., not... at all, not... whatsoever) amplify negativity, significantly increasing the likelihood of appearing in low-star reviews.
  • Attenuating NPIs (e.g., not... all that, not... much) soften the statement. These shift the distribution toward mild positives (6–7 stars), often functioning as politeness markers.

Empirical Findings

The quantitative analysis reveals a persistent bias across domains and languages. In the IMDB corpus, the probability of negation is highest for low-star reviews and declines steadily toward 10-star reviews.

Probability of a rating category given the presence of negation in the IMDB corpus.

This trend is mirrored in dialogue acts: in the Switchboard corpus, 59.4% of "reject" moves contained negation, compared to only 1.8% of "accept" moves. In the Experience Project, negation correlates highly with reactions of solidarity (hugs, I understand) and shock (just wow), but rarely with positive excitement (rocks). Similarly, in the Convote corpus, speakers who eventually voted "No" on a bill used significantly more negation in their floor speeches than those who voted "Yes."

Conclusion and Future Directions

The "negativity of negation" appears to be a stable cross-linguistic phenomenon. However, several avenues for future research remain. While the QUD framework provides a compelling explanation, large-scale corpora specifically annotated for QUDs are needed for direct testing. Additionally, genre-specific effects warrant further investigation. Future studies should also employ full multi-level regression modeling to better account for individual speaker and topic variance.

Summary updated

Abstract

Natural language negation is persistently negative in the pragmatic sense, and emphatic and attenuating negative polarity items modulate this effect in systematic ways. I use large corpora of informal texts with meta-data approximating features of the context to characterize this pragmatic negativity, and I attempt to explain it in terms of the ways in which negative sentences engage the questions under discussion. The discussion highlights some of the ways in which quantitative corpus methods can be used to achieve novel results in linguistic pragmatics.
BibTeX
@inproceedings{potts2010negativity,
  author = {Christopher Potts},
  title = {{On the negativity of negation}},
  booktitle = {Proceedings From Semantics and Linguistic Theory},
  year = {2010},
}
Figure