Negation Mitigates Rather Than Inverts the Neural Representations of Adjectives

Neurology Medicine Humanities and Others Clinical Psychology Biophysics Life Sciences
negation adjectives neural representation behavioral study MEG experiment semantics

Background Introduction

A notable characteristic of human language processing is our ability to combine stored lexical elements, i.e., words, as needed to flexibly generate or alter current meanings. At the core of this process is how we construct semantic representations in real time. Although research on the generation of syntactic structures has made steady progress and sparked productive discussions, there has been relatively less study on how new semantic configurations are represented over time. The research team from Johns Hopkins University, based on existing literature, specifically examined how negation in language affects the semantic representation of words, especially adjectives. This study provides a cornerstone for understanding how the human brain represents changing meanings in real-time.

Paper Source

The paper titled “Negation mitigates rather than inverts the neural representations of adjectives” was co-authored by Arianna Zuanazzi, Pablo Ripollés, Wy Ming Lin, Laura Gwilliams, Jean-Rémi King, and David Poeppel from institutions such as the Department of Psychology at New York University, New York University Music and Audio Research Laboratory (MARL), Hector Institute of Education Science and Psychology at the University of Tübingen in Germany, Stanford University’s Department of Psychology, and the École Normale Supérieure at PSL University in Paris, France. This paper was published in the PLOS Biology journal with the DOI 10.1371/journal.pbio.3002622 on May 30, 2024.

Research Content and Methods

The research work was divided into two parts: a behavioral experiment and a neurophysiological experiment (MEG).

Experiment 1: Behavioral Experiment (Continuous Mouse-tracking)

The research team designed an online behavioral experiment to track changes in the representation of adjectives when reading negated and non-negated phrases. The experiment consisted of two parts: an initial test with 78 participants and a no-feedback retest with 55 participants. In the experiment, participants read phrases composed of two modifiers and an adjective, and rated the overall semantics of each phrase on a scale.

The modifiers included in the research design were “not” and “really,” while the adjectives were selected as antonym pairs, such as “bad-good,” “cold-hot,” etc. The participants’ mouse trajectory data provided dynamically changing data, reflecting the process of understanding words and phrases.

  • Reaction Time Analysis: The experiment results showed that participants responded faster to high-polarity adjectives (such as “good”) compared to low-polarity adjectives (such as “bad”), and reacted faster to non-negated phrases (such as “really good”) than negated phrases (such as “not good”). This result supports the previous research viewpoint that processing negated phrases is more challenging.

  • Continuous Mouse-tracking Analysis: The results showed that the final interpretation of negated phrases was located in the middle of the semantic scale but never completely overlapped with the interpretations of the affirmative polarity words. Moreover, the final interpretation of negated phrases was more variable than that of affirmative phrases, indicating that the negation operation moved the final interpretation of adjectives towards the direction of the antonym but never fully reversed it.

Experiment 2: MEG Experiment

In the MEG experiment, participants read phrases and decided whether a probe number reflected the overall meaning of the phrase. MEG data provided a process to track the evolution of adjective representation, mainly using time and space decoding methods.

  • Temporal Decoding Analysis: The temporal decoding results of word characteristics showed significant decoding of “really” and “not” within a time window of 120 to 740ms after the first modifier. The decoding of polarity pairs (such as “bad-good”) was significant within a 90 to 410ms time window after the onset of the adjective.

  • Temporal and Spatial Decoding Analysis: The decoding accuracy of distinguishing low-polarity and high-polarity words was significant in the left temporal lobe, superior temporal gyrus, and middle temporal gyrus, indicating that these regions are involved in lexical-semantic representation and combination.

  • Analysis of the Impact of Negation on Adjective Representation: The study evaluated the function of negation in real-time semantic construction by testing four hypotheses (no effect, mitigation, inversion, and change) using two complementary decoding methods (generalization and prediction). Results showed that the representation of adjectives in negated phrases is only weakened during the preliminary representation stage, rather than reversed or changed. This result supports the behavioral data’s conclusion from a neural level that “negation does not reverse the meaning of adjectives but weakens their representation.”

Beta Wave Dynamics

To explore whether negation operates through inhibitory mechanisms, the research team conducted a time-frequency analysis of beta wave power. Results showed that the power of low and high beta waves in the left sensorimotor cortex was significantly higher during the processing of affirmative phrases compared to negated phrases, further supporting the hypothesis that negation might operate through a general inhibitory system.

Significance and Value of the Study

This research, employing time-resolved behavioral and neuroimaging methods together with multivariable decoding techniques, provides new insights into how negation affects lexical-semantic representations. The study demonstrated neural evidence that negation functions as a weakening mechanism during the early semantic processing stages of adjectives. Moreover, this research lays a foundation for understanding how the human brain processes meaning changes in combinatorial structures, offering important implications for language comprehension, cognitive model construction, and AI applications in natural language processing.

In summary, this paper provides a reasonable explanation for the function of negation in lexical-semantic representation, revealing that negation weakens lexical representations through a general inhibitory system during early semantic processing stages. Future research could focus more on how negation in different contexts affects more complex semantic and syntactic representations, supplementing the current research findings.