TL;DR
This paper investigates the "concept creep" hypothesis, which suggests that psychological concepts related to harm have broadened their meanings over the past few decades. Using computational linguistic models on a large corpus of psychology abstracts, the authors provide evidence for this broadening in concepts like "addiction" and "harassment," while others like "trauma" remained relatively static.
What Problem Does This Paper Address?
The field of psychology, like many social sciences, relies on precise definitions of concepts to understand and address complex human behaviors and experiences. However, there's a growing concern, articulated by the "concept creep" hypothesis, that many terms related to harm (e.g., addiction, bullying, prejudice, trauma) have become increasingly broad over time. This "semantic inflation" could lead to an over-pathologizing of everyday experiences and potentially obscure genuine harm. This paper aims to empirically test this hypothesis by analyzing semantic changes in these concepts within academic psychology discourse over the last four decades.
Methods
The authors employ two computational linguistic models to analyze semantic change:
- LSA-based Model (Sagi et al., 2009): This approach uses Latent Semantic Analysis (LSA) to create term-document co-occurrence matrices. By factorizing these matrices and contextualizing word embeddings for different decades, they can estimate the "semantic breadth" of a concept by calculating the average cosine similarity between its contextual representations. A lower similarity indicates a broader meaning.
- Diachronic Adaptation of word2vec (Hamilton et al., 2016): This prediction-based neural model is trained on separate word2vec models for each decade. They then align these models using orthogonal Procrustes to measure:
- Semantic Displacement: Quantifies how much a word's meaning has changed between time periods (cosine distance between aligned embeddings).
- Pair-wise Similarity Time-Series: Examines how the cosine similarity between a target concept and its nearest neighbors (related words) evolves over time, revealing shifts in associated meanings.
Datasets
The primary dataset used is a corpus of abstracts from psychology journals, spanning from 1930 to 2019. This corpus was collected from the E-Research and PubMed databases, totaling 871,340 abstracts and 133,082,240 tokens. For their experiments, they focused on abstracts from 1970 onwards due to insufficient data in earlier periods.
Key Insights and Ideas
- "Concept Creep" is not uniform: While the hypothesis suggests a general broadening, the study reveals that not all harm-related concepts have expanded equally. Some, like "trauma," have maintained a relatively stable semantic breadth.
- Shifting meanings, not just broadening: The analysis goes beyond just measuring breadth to showing how meanings have shifted. For example, "addiction" has moved from substance-related meanings to encompass behavioral dependencies like gaming and internet use.
- Temporal patterns of change: The study highlights that semantic changes were most substantial between the 1980s and 1990s, suggesting this period was particularly critical for the evolution of these concepts.
- Horizontal vs. Vertical Creep: The methods primarily capture "horizontal creep" (expansion into new contexts), but the authors acknowledge the need for future work to specifically address "vertical creep" (inclusion of milder instances).
Key Results
- Broadening: Concepts like "addiction," "bullying," and "harassment" showed evidence of broadening in their semantic scope.
- "Addiction" expanded to include psychological dependencies (gaming, internet).
- "Harassment" broadened to include non-sexual forms of unwanted attention and moved towards more psychological meanings.
- "Bullying" transformed into a more victim-centered concept and extended to new domains like workplaces.
- Static/Narrowing: "Trauma" remained relatively static in its semantic breadth, while "prejudice" showed some narrowing.
- Semantic Displacement: All five concepts (addiction, bullying, harassment, prejudice, trauma) underwent the most significant semantic displacement between the 1980s and 1990s.
- Specific Meaning Shifts:
- "Addiction" shifted from "drug" and "narcotic" to "gaming," "internet," and "smartphone."
- "Bullying" became more associated with "victim" and "workplace," while retaining links to "school" and "child."
- "Harassment" expanded to include "cyber" and "online" contexts.
- "Trauma" shifted from primarily "physical" to "psychological" meanings and increased its association with "stress."
- "Prejudice" strengthened its links to "discrimination" and "racial" but weakened connections to "black" and "woman."
Limitations and Future Work
- Methodological Limitations: The authors acknowledge that their methods primarily capture "horizontal" semantic creep and do not adequately account for "vertical" creep (inclusion of milder phenomena). Future work should develop methods to distinguish these.
- Generalizability: The findings are specific to harm-related concepts in psychology. Future research should explore other domains and languages to see if similar expansionary trends exist.
- Cultural Factors: A fundamental challenge is to uncover the cultural factors driving this semantic inflation and understand its societal implications.
- Replication and Refinement: The authors suggest replicating these findings with additional concepts (e.g., "mental illness," "safety") and refining the analytic parameters of the models used. More focused analyses on individual concepts are also recommended.
Summary updated