Skip to content

lsc pattern interpretable lexical semantic change analysis via diachronic sense

content_hash: 2e3c226662d81c20 ---# LSC-Pattern: Interpretable Lexical Semantic Change Analysis via Diachronic Sense Distributions

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Lexuss-D/LSC-Pattern
Area: NLP Understanding
Keywords: lexical semantic change, word sense disambiguation, diachronic sense distributions, semantic dimensions, interpretability

TL;DR

LSC-Pattern proposes an interpretable framework for lexical semantic change analysis that couples dictionary-grounded diachronic sense distributions with operationalized semantic dimensions, leveraging LLM-based word sense disambiguation to ground usages into lexicographic senses and decomposing macro-trajectories across valence, semantic scope, and metaphoricity into concrete sense-frequency shifts.

Background & Motivation

Lexical semantic change (LSC) detection identifies words whose meanings evolve over time across historical corpora. Substantial computational progress has been made using diachronic word embeddings, contextualized representations, and clustering-based usage modeling. While these approaches can reliably detect whether a target word has undergone semantic drift, they leave the precise linguistic source and cognitive drivers of the shift fundamentally underspecified: producing aggregate scalar change scores or latent contextual clusters fails to pinpoint which dictionary senses are changing and what semantic property drives the observed transition.

The core tension lies in the gap between corpus-level statistical changes and inspectable linguistic evidence. A target word may exhibit an increasingly negative affective context because a pejorative sense has surged in relative frequency, or its semantic breadth may broaden because a specific denotation has been generalized. However, without grounding usages to explicit, standardized dictionary definitions, existing semantic-dimension approaches (tracking sentiment, intensity, or breadth) remain unlinked to concrete lexical units, while unsupervised clustering yields latent groupings that do not align with human lexicographic classifications.

To bridge corpus-level change detection and linguistically interpretable semantic patterns, this paper proposes a modular two-step framework that retains the lexicographic transparency of explicit dictionary senses while preserving the flexibility of continuous dimension-level analysis. Core idea: decouple lexical semantic change analysis into LLM-driven dictionary-grounded diachronic sense distribution construction and multi-dimensional trajectory tracking, enabling the fine-grained attribution of macro-level trends (amelioration, pejoration, generalization, specialization, and metaphorization) directly to the frequency redistribution of explicit dictionary senses.

Method

Overall Architecture

The LSC-Pattern framework comprises two sequential stages: constructing dictionary-grounded diachronic sense distributions (Step 1) and decomposing diachronic trajectories along selected semantic dimensions (Step 2). Given diachronic corpora usages across different historical periods and an authoritative dictionary sense inventory, Step 1 uses an LLM-based word sense disambiguation (WSD) module to map each usage to an explicit dictionary sense definition (or a fallback Sense 0), aggregating the assignments into discrete probability distributions by decade. Step 2 applies domain-specific operationalized metrics—valence, semantic scope, and metaphoricity—to evaluate usage contexts or aligned senses, and cross-inspects the dimension slopes against sense distributions to identify and verify candidate semantic change patterns.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Diachronic Usages + Dictionary Sense Inventory"] --> B["Stage 1: Dictionary-Grounded WSD with LLM<br/>Map usages to explicit senses or Sense 0"]
    B --> C["Diachronic Sense Distribution Aggregation<br/>Compute relative frequency pt(si|w) per decade"]
    C --> D["Stage 2: Multi-Dimensional Trajectory Analysis<br/>Valence / Semantic Scope / Metaphoricity"]
    D --> E["Pattern Attribution & Qualitative Verification<br/>Inspect sense redistribution for Pejoration / Generalization / Metaphorization"]

Key Designs

1. Dictionary-Grounded WSD with LLM: Replacing Latent Clusters with Explicit Lexicographic Senses

Addressing the fundamental limitation that unsupervised clustering yields uninterpretable latent clusters, LSC-Pattern explicitly grounds each usage to a standardized dictionary inventory. For a target entry \(w\), candidate senses are drawn from the Iwanami Japanese Dictionary (5th Edition), represented as \(S_w = \{s_1, s_2, \dots, s_{|S_w|}\}\), augmented with a fallback label \(s_0\) (Sense 0) for out-of-vocabulary senses, ambiguous contexts, or disambiguation failures. Each contextual usage \(u \in U_{t,w}\) in time period \(t\) is mapped via a WSD function:

\[\hat{s}(u) = f_{\text{WSD}}(u, S_w), \quad \hat{s}(u) \in S_w \cup \{s_0\}\]

The disambiguation module is implemented using Qwen3-8B in thinking mode. The model receives the target sentence with marked brackets and candidate glosses, outputting the predicted sense index in brackets (e.g., [2]) followed by a concise rationale. After stripping <think>...</think> reasoning tokens, the parsed sense indices are aggregated over decade \(t\) to compute the diachronic sense distribution:

\[p_t(s_i \mid w) = \frac{|U_{t,s_i}|}{|U_{t,w}|}\]

When a scalar divergence measure between periods \(t\) and \(t'\) is needed, the Jensen–Shannon divergence (JSD) between their corresponding discrete probability distributions is directly calculated.

2. Operationalized Semantic-Dimension Trajectories: Formalizing Classical Semantic Change Types

To interpret the cognitive directions of semantic shifts, Step 2 formalizes three interpretable semantic dimensions corresponding to classical linguistic change typologies: - Valence (Amelioration / Pejoration): Measures the evaluative sentiment environment surrounding the target word. Using a Japanese sentiment polarity lexicon, the usage-level valence score \(\text{val}(u)\) averages the polarity of context words \(x \in C(u)\) (strictly excluding the target word itself). The word-level score for period \(t\) is \(\text{Val}_t(w) = \frac{1}{|U_{t,w}|} \sum_{u \in U_{t,w}} \text{val}(u)\). An upward slope marks an amelioration candidate, whereas a downward slope signals pejoration. - Semantic Scope (Specialization / Generalization): Quantifies conceptual specificity via taxonomic depth in the Japanese WordNet hierarchy. Senses from the Iwanami dictionary are mapped to WordNet synsets using Sarashina-Embedding-v2 representations. The specificity \(\text{scope}(s_i)\) is defined as the distance from the aligned synset to the WordNet root. The period-level semantic scope score is the distribution-weighted expectation:

\[\text{Scope}_t(w) = \sum_{i} p_t(s_i \mid w) \cdot \text{scope}(s_i)\]

An increasing slope indicates specialization (shift toward deeper, narrower senses), while a decreasing slope indicates generalization (broadening toward shallower, more general senses). - Metaphoricity (Metaphorization): Assesses figurative extension using a Japanese metaphor detection model. The binary label \(m(u) \in \{0, 1\}\) evaluates whether the target usage is metaphorical, yielding the period-level proportion \(\text{Meta}_t(w) = \frac{1}{|U_{t,w}|} \sum_{u \in U_{t,w}} m(u)\). A rising trajectory signals candidate metaphorization.

3. Trend Discovery and Sense Redistribution Attribution: Explaining Macro Shifts via Micro Sense Dynamics

Rather than treating continuous dimensional trajectories as definitive proof of semantic change, LSC-Pattern utilizes them as exploratory discovery signals. For target words observed across at least ten decades, ordinary least-squares (OLS) linear regressions \(y_{w,d,t} = \beta_{0,w,d} + \beta_{1,w,d} t + \epsilon_t\) are fitted across decades. Target words with the steepest slopes \(\beta_1\) are flagged as candidate change items. Researchers then jointly inspect the underlying sense distributions and attested historical quotes. This verification confirms whether the macro property shift is caused by a genuine historical reallocation of sense frequencies—such as a pejorative sense replacing an original neutral sense—rather than topical noise or detector artifacts.

A Worked Example

Consider the target word 「差別」 (sabetsu, distinction/discrimination) spanning the 1870s to the 2000s: 1. Dictionary Senses: Sense 1 is defined as "discriminatory treatment" (WordNet depth 6); Sense 2 is defined as "neutral distinction/differentiation" (WordNet depth 7). 2. Disambiguation and Sense Distribution: In the 1880s and 1890s, the neutral Sense 2 dominated the corpus, accounting for 84.0% and 69.4% of occurrences (e.g., in a 1957 magazine example: "With no distinction between frontal and surprise attacks, nor between strategic and tactical attacks"). Beginning in the 1920s, Sense 1 surged to 64.2%, and remained consistently above 80% after the 1950s, reaching 100% in the 1990s and 98.0% in the 2000s (e.g., a 2001 usage: "A struggle to win freedom from prejudice and discrimination in America"). 3. Dimensional Trajectories: The valence trajectory exhibits a marked negative trend (slope \(-0.0124\) per decade, \(R^2 = 0.564\)), and the semantic scope trajectory also decreases (slope \(-0.0533\) per decade, \(R^2 = 0.734\)). 4. Linguistic Attribution: Cross-inspection confirms that the negative valence trajectory is fundamentally driven by the replacement of the neutral distinction sense with the pejorative discrimination sense, providing an inspectable account of historical pejoration.

Key Experimental Results

The framework is evaluated on three historical and contemporary corpora from NINJAL: the Corpus of Historical Japanese (CHJ, 1870s–1920s, 12.3M words), the Showa-Heisei Corpus (SHC, 1930s–2010s, 31.1M words), and the Balanced Corpus of Contemporary Written Japanese (BCCWJ, 2000s, 4.3M words), all strictly filtered to the magazine register. The main analysis processes 150 polysemous entries from the Iwanami Japanese Dictionary spanning 213,650 usages. The intrinsic WSD evaluation subset comprises 976 usages sampled from 50 target words, independently labeled by three native-speaker linguists.

Main Results

Intrinsic WSD accuracy evaluated against 976 human-annotated usages across five baselines and Qwen3-8B:

Method Mean annot. Majority vote
Random baseline 0.281 0.282
WordNet BERT \(\rightarrow\) Iwanami 0.476 0.435
Target token–gloss BERT 0.577 0.591
Usage–gloss embedding (Sarashina) 0.663 0.681
MFS (Most Frequent Sense, optimistic) 0.695 0.699
Qwen3-8B (LSC-Pattern, Ours) 0.793 0.815

Note: Majority vote accuracy is computed on the 961 usages where at least two human annotators agreed (mean pairwise Cohen's \(\kappa = 0.798\), mean Krippendorff's \(\alpha = 0.689\)). Across individual historical corpora, Qwen3-8B maintains consistent majority-vote accuracy: 0.803 on CHJ, 0.813 on SHC, and 0.844 on BCCWJ.

Extrinsic correlation with human change judgments on the JaSemChange benchmark:

Setting Corpus pair Spearman's \(\rho\) \(p\)-value
Overlap (7 polysemous entries) SHC–BCCWJ 0.821 0.023
Overlap (7 polysemous entries) CHJ–BCCWJ 0.750 0.052
Full benchmark (20 entries) SHC–BCCWJ 0.580 0.007
Full benchmark (20 entries) CHJ–BCCWJ 0.367 0.112

Ablation Study and Candidate Discovery

Across the 130 target entries observed in at least ten decades, linear regression slope statistics identify top semantic change candidates across all three dimensions:

Dimension Pattern candidate Top 3 entries Slope & \(R^2\) examples
Valence (+) Amelioration 現地 (genchi), 陽気 (youki), 前衛 (zen'ei) 現地: slope +0.0538, \(R^2 = 0.531\)
Valence (−) Pejoration ケース (keesu), 下流 (karyuu), カード (kaado) ケース: slope -0.0328, \(R^2 = 0.233\)
Semantic scope (+) Specialization 面 (men), 得意 (tokui), 足元 (ashimoto) 面: slope +0.3104, \(R^2 = 0.677\)
Semantic scope (−) Generalization プラス (purasu), 品種 (hinshu), 縁 (en) プラス: slope -0.2347, \(R^2 = 0.468\)
Metaphoricity (+) Metaphorization 回転 (kaiten), プラス (purasu), 土俵 (dohyou) 回転: slope +0.0763, \(R^2 = 0.694\)
Metaphoricity (−) Less metaphorical 陽気 (youki), 天国 (tengoku), 手間 (tema) 陽気: slope -0.0674, \(R^2 = 0.801\)

Robustness checks show that OLS slopes strongly correlate with non-parametric Theil–Sen estimator slopes (Valence \(\rho = 0.787\), Scope \(\rho = 0.839\), Metaphoricity \(\rho = 0.908\), all \(p < 10^{-27}\)).

Key Findings

  • Sense 0 is rare and structurally concentrated: Overall, Sense 0 accounts for only 1.42% of the 213,650 model predictions. Over 39% of these occurrences stem from just two entries: loanword 「イン」 (in, 93.17% Sense 0, representing a genuine inventory omission where Iwanami only lists sports senses while corpus usages reflect innerwear or the preposition), and 「北」 (kita, 68.42% Sense 0, caused by model misinterpretation of parenthetical antonym definitions). For the remaining 148 entries, the pooled Sense 0 rate is merely 0.87%.
  • Dictionary granularity bounds detection resolution: The correlation drop between the 7-entry polysemous subset (\(\rho = 0.821\)) and the 20-entry full JaSemChange set (\(\rho = 0.580\)) highlights that when a dictionary assigns only a single broad sense to an evolving word, discrete sense redistribution cannot capture fine-grained sub-sense drift, leading to tied change scores.
  • Multidimensional trajectories capture complementary shift phenomena: Case studies reveal that single dimensions can be blind to valid changes: for 「土台」 (dodai), both the concrete foundation sense and the abstract basis sense sit at WordNet depth 8 (yielding a constant scope score of 0), yet the metaphoricity trajectory surges (slope \(+0.0404\)), successfully uncovering an abstract metaphorical mapping that taxonomic depth fails to detect.

Highlights & Insights

  • Synergistic coupling of discrete senses and continuous dimensions: Replaces opaque vector clustering with a transparent framework that directly traces macroscopic property shifts to microscopic dictionary sense distributions.
  • Empirical rigour over 150 years of historical data: Constructs and evaluates an end-to-end historical NLP pipeline across 15 decades of Japanese magazine corpora, providing an inspectable blueprint for computational historical linguistics.
  • Chain-of-thought robustness for historical disambiguation: Zero-shot Qwen3-8B in thinking mode effectively navigates archaic orthography and mixed syntactic registers, exceeding fine-tuned BERT baselines by 11.6% to 22.4% in majority-vote accuracy.

Limitations & Future Work

  • Dictionary coverage and granularity dependency: Fixed lexicographic inventories may omit obsolete historical senses, specialized terms, or emerging slang. Integrating word sense induction (WSI) or context-generated definitions could address dictionary coverage gaps.
  • Proxy noise in operationalized semantic dimensions: Context-based valence can reflect discourse domain rather than intrinsic sentiment; WordNet hierarchy depth is an imperfect proxy for cognitive semantic specificity; and modern metaphor detectors systematically undercount conventionalized historical metaphors.
  • Corpus heterogeneity across epochs: Orthographic, editorial, and genre variances between CHJ, SHC, and BCCWJ introduce register noise that warrants further controlled historical corpus alignment.
  • vs. Unsupervised Embedding Clustering (Hamilton et al., 2016; Kutuzov et al., 2022): Prior LSCD methods rely on latent usage clusters that cannot be audited against standardized lexicographic categories; LSC-Pattern grounds usages to explicit dictionary glosses with high human agreement and interpretable attributions.
  • vs. Continuous Semantic-Dimension Tracking (Baes et al., 2024): Purely dimensional methods track shifts in property scores but cannot identify which specific meanings are responsible for the trend; LSC-Pattern provides the exact sense-level explanatory mechanism.

Rating

  • Novelty: ⭐⭐⭐⭐ [Pioneering framework connecting dictionary-grounded LLM disambiguation with multi-dimensional diachronic semantic trajectories]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive 150-year corpus coverage, 213k disambiguated usages, rigorous human agreement validation, and deep qualitative case analyses]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Exceptionally clear architectural exposition, rigorous mathematical and linguistic formulations, and highly coherent case studies]
  • Value: ⭐⭐⭐⭐⭐ [Provides a standardized, inspectable methodology and empirical foundation for computational linguistics and historical semantics]