Source-linked AI summary
Cultural Shift or Linguistic Drift? Comparing Two Computational Measures of Semantic Change
William L. Hamilton, Jure Leskovec, Dan Jurafsky
TL;DR
Semantic-change research needs to distinguish cultural shifts from regular linguistic drift. The paper compares global vector-distance and local-neighborhood measures, finding that they show different sensitivities: the local measure favors noun changes associated with cultural shifts, while the global measure favors verb changes associated with regular drift. This comparison supports choosing measures according to the research task.
Problem
Existing distributional work does not disentangle semantic changes caused by cultural shifts from those caused by regular linguistic processes.
Method
The paper compares a local-neighborhood measure with a traditional global measure using six historical corpora, mixed-model analyses, and case studies.
Results
The local measure is more sensitive to noun changes, whereas the global measure is more sensitive to verb changes, matching the proposed cultural-versus-linguistic distinction.
Takeaways & Limitations
Researchers should choose semantic-change measures according to task, using local neighborhoods for cultural change and global distance for grammaticalization.
Takeaways & Limitations
The cultural-versus-linguistic interpretation relies on noun–verb mapping as a proxy for the two types of change.
Abstract
from arXiv · showhide
Words shift in meaning for many reasons, including cultural factors like new technologies and regular linguistic processes like subjectification. Understanding the evolution of language and culture requires disentangling these underlying causes. Here we show how two different distributional measures can be used to detect two different types of semantic change. The first measure, which has been used in many previous works, analyzes global shifts in a word's distributional semantics, it is sensitive to changes due to regular processes of linguistic drift, such as the semantic generalization of promise ("I promise." -> "It promised to be exciting."). The second measure, which we develop here, focuses on local changes to a word's nearest semantic neighbors; it is more sensitive to cultural shifts, such as the change in the meaning of cell ("prison cell" -> "cell phone"). Comparing measurements made by these two methods allows researchers to determine whether changes are more cultural or linguistic in nature, a distinction that is essential for work in the digital humanities and historical linguistics.
1 Introduction
The paper addresses the lack of methods for disentangling cultural shifts from regular linguistic processes in semantic change. It compares a local-neighborhood measure with a traditional global measure to detect these different types of change.
- Previous distributional studies detect diachronic semantic change but do not distinguish its underlying causes.
- The distinction between cultural and regular linguistic change matters differently across digital humanities and historical linguistics.Cultural shifts are valuable signals for historical and digital-humanities research but noise for studies of regular linguistic processes.
- The local neighborhood measure targets changes in a word’s nearest semantic neighbors and is intended to capture cultural shifts in nominal meanings.
- The traditional global measure tracks shifts in a word’s overall distributional semantics and is intended to capture regular linguistic drift, especially in verbs.
- The study evaluates the two measures using six historical corpora across multiple languages and illustrative case studies.
2 Methods
The study measures semantic change with aligned diachronic word embeddings using global vector distance and local nearest-neighbor change. It evaluates their differing sensitivities through regression analyses across historical corpora and case studies.
- Diachronic word2vec vectors represent each word’s co-occurrence statistics by time period and are aligned across years for comparison.The embeddings use SGNS and measure changes between consecutive decades.
- The analysis uses historical Google N-gram and COHA vectors spanning multiple languages and decades.COHA is smaller but genre-balanced and includes lemmas as well as surface forms.
- Global measure: The global measure computes cosine distance between a word’s vectors in two consecutive decades.
- Local neighborhood measure: The local measure compares second-order similarity vectors built from each word’s k nearest neighbors across adjacent decades.These vectors encode the word’s cosine similarities with neighbors, and local change is the cosine distance between the resulting vectors.
- Local neighborhood measure: The local measure is less sensitive to global shifts in syntagmatic contexts while capturing strong changes in paradigmatic relations.
- Statistical methodology: Noun–verb differences are tested with linear mixed models that account for correlations among repeated measurements of words across time.The regressions include frequency, decade, and noun-versus-verb status as predictors.
3 Results
The two measures yield different patterns of semantic change: the local measure is more sensitive to nouns, while the global measure is more sensitive to verbs and related parts of speech. Case studies show that regular linguistic shifts favor the global measure, whereas cultural shifts favor the local measure.
- Nouns vs. verbs: The local neighborhood measure assigns higher semantic-change rates to nouns, while the global distance measure assigns higher rates to verbs across languages and datasets.
- Nouns vs. verbs: Including adjectives and adverbs produces analogous significant trends alongside verbs.
- Case studies: Six case studies divide into regular linguistic shifts that changed more under the global measure and cultural changes that changed more under the local measure.
- Case studies: Set A includes actually, must, and promise, representing subjectification, modal semantic change, and pragmatic change in performative speech acts.
- Case studies: Set B includes gay, virus, and cell, whose new meanings arose through community-specific vernacular or technological advances.
4 Discussion
The findings support using different semantic-change measures for different research goals. Comparing the measures can also help assess whether a change is more linguistic or cultural in nature.
- The local measure aligns with nominal and cultural shifts, while the global measure aligns with verbal and regular linguistic drift.
- Cultural change-point detection would favor the local measure, whereas grammaticalization research would favor the traditional global distance approach.
- Comparing both measurements allows researchers to assess the extent to which semantic changes are linguistic or cultural.