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: Interpretability
Keywords: Lexical Semantic Change, Word Sense Disambiguation, Large Language Models, Semantic Dimensions, Interpretability
TL;DR¶
Addressing the black-box nature of conventional lexical semantic change detection that outputs only uninterpretable scalar change scores, LSC-Pattern connects LLM-based dictionary-grounded word sense disambiguation with multi-dimensional semantic properties (valence, semantic scope, metaphoricity), achieving high correlation with human judgments and transparent linguistic explanations across 150 years of Japanese corpora.
Background & Motivation¶
Lexical semantic change detection (LSCD) aims to identify words whose meanings evolve across historical periods. Recent computational approaches have evolved from early static diachronic embeddings to contextualized representations and unsupervised usage clustering with models such as BERT and RoBERTa, establishing standardized benchmarks like DURel and SemEval-2020 Task 1. These methods excel at estimating whether a word has changed and quantifying its overall change magnitude. However, in historical linguistics, lexicography, and cognitive semantics, detecting an aggregate change score is insufficient. Linguists need to understand the underlying cognitive motivations and specific taxonomic change types: whether a word has undergone pejoration because a negative sense became more prominent, or generalization because an initially restricted technical sense broadened into general usage.
Existing representation-based methods primarily model usage distributions through unsupervised clustering in latent embedding spaces. Unfortunately, these latent clusters are ungrounded and often do not correspond to lexicographically established word senses, making their boundaries ambiguous and difficult to trace. Conversely, emerging dimension-based approaches describe change along interpretable properties such as sentiment, intensity, or breadth. Yet, these dimensions are typically computed directly over bag-of-words contexts without anchoring to explicit word senses. As a result, it remains impossible to tell whether a shift in a property trajectory arises from the birth of a novel sense, the loss of an obsolete sense, or the redistribution of existing dictionary senses. This fundamental disconnect between corpus-level statistical signals and inspectable linguistic evidence limits the utility of computational LSCD for empirical linguistic research.
To bridge this gap, this paper introduces LSC-Pattern, an interpretable two-step framework designed to tie diachronic sense distributions directly to multi-dimensional semantic properties. The key insight is that anchoring contextual usages to an explicit, fixed dictionary sense inventory enables transparent tracking of sense redistribution over time, which subsequently explains macro-level trajectories across semantic dimensions. Core idea: decompose lexical semantic change into dictionary-grounded diachronic sense distribution construction and multi-dimensional trajectory decomposition, using sense redistribution as an inspectable mechanism to explain classical change patterns including pejoration, generalization, and metaphorization.
Method¶
Overall Architecture¶
LSC-Pattern follows a modular two-step pipeline. Step 1 takes as input historical usage examples of a target word across historical periods and an authoritative dictionary sense inventory; a large language model (Qwen3-8B in thinking mode) performs dictionary-grounded word sense disambiguation (WSD) to map each usage to an explicit dictionary sense (with a dedicated fallback Sense 0 for unlisted or ambiguous contexts), aggregating assignments by decade to construct period-specific sense distributions. Step 2 selects orthogonal semantic dimensions aligned with targeted semantic change patterns—operationalizing valence for amelioration/pejoration, WordNet taxonomic depth for specialization/generalization, and classifier-detected metaphoricity for metaphorization—to decompose scalar trajectories by explicit dictionary senses and validate candidates through prototypical usage analysis.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Diachronic Usages & Sense Inventory<br/>CHJ/SHC/BCCWJ Corpora (1870s-2010s)"] --> B["Dictionary-Grounded Diachronic WSD<br/>Qwen3-8B Thinking Mode maps context to senses"]
B --> C["Diachronic Sense Distribution Construction<br/>Decade-level sense probabilities & JSD change scores"]
C --> D["Multi-Dimensional Trajectory Decomposition<br/>Valence, WordNet Scope Depth, & Metaphoricity"]
D --> E["Change Pattern Discovery & Prototypical Verification<br/>Pejoration in Sabetsu / Generalization in Meshi"]
Key Designs¶
1. Dictionary-Grounded Diachronic WSD: Grounding Latent Usages into Fixed Lexicographic Sense Inventories
Unsupervised contextualized clustering yields ungrounded, unstable clusters across time periods that resist lexicographic validation. This design anchors the sense space to the authoritative Iwanami Japanese Dictionary (5th Edition), defining for each target entry a discrete candidate set \(S_w = \{s_1, s_2, \dots, s_{|S_w|}\}\) augmented with an explicit fallback sense \(s_0\) for out-of-vocabulary senses, insufficient contexts, or severe ambiguities. To robustly handle historical orthographic shifts and archaic grammatical structures spanning 150 years, Qwen3-8B in thinking mode is employed as the disambiguation function \(f_{\text{WSD}}(u, S_w)\), taking the target occurrence context and definition glosses to predict the bracketed index \([i]\) along with a concise justification. For each period \(t\), usages assigned to sense \(s_i\) form subset \(U_{t,s_i} = \{u \in U_{t,w} \mid \hat{s}(u) = s_i\}\), producing a stable discrete probability distribution:
Because the sense inventory remains constant across periods, diachronic shifts reflect changes in the relative prominence of known dictionary senses rather than latent cluster shifts. The aggregate change score between any two periods \(t\) and \(t'\) is measured directly via the Jensen–Shannon divergence (JSD) between their sense distributions.
2. Multi-Dimensional Trajectory Decomposition: Decomposing Macro Shifts into Inspectable Semantic Dimensions
Detecting overall distribution divergence fails to reveal which semantic properties drive the change. This design operationalizes three orthogonal semantic dimensions mapped to classical linguistic change typologies. For Valence, evaluating amelioration and pejoration, the score is calculated by averaging sentiment polarity scores of surrounding context words \(C(u)\) using a validated sentiment dictionary while excluding the target word itself:
For Semantic Scope, evaluating generalization and specialization, Iwanami senses are aligned to Japanese WordNet synsets via Sarashina-Embedding-v2-1B, and the specificity \(\text{scope}(s_i)\) is defined by the hierarchical depth from the WordNet root; the word-level depth is the probability-weighted sum \(\text{Scope}_t(w) = \sum_i p_t(s_i \mid w) \cdot \text{scope}(s_i)\), where rising depth indicates specialization and falling depth indicates generalization. For Metaphoricity, a contemporary Japanese metaphor detector assigns a binary label \(m(u) \in \{0, 1\}\), computing period-level metaphoricity \(\text{Meta}_t(w) = \frac{1}{|U_{t,w}|} \sum_{u \in U_{t,w}} m(u)\). Joint inspection of the dimension trajectory and the underlying sense distribution enables fine-grained attribution of property changes to individual senses.
3. Change Pattern Discovery & Prototypical Verification: Filtering Trend Candidates via Verified Usage Evidence
To distinguish genuine diachronic trends from sampling noise and genre skew across historical corpora, a rigorous candidate discovery and qualitative verification protocol is established. For all 130 target entries observed in at least ten decades, linear regression models \(y_{w,d,t} = \beta_{0,w,d} + \beta_{1,w,d} t + \epsilon_t\) are fitted over decadal observations (cross-checked with Theil–Sen robust estimators and Mann–Kendall monotonic tests) to rank candidates by slope \(\beta_{1,w,d}\). Crucially, statistical slopes serve strictly as discovery signals rather than definitive proof: every candidate pattern must be substantiated through the tripartite alignment of "sense distribution redistribution \(\leftrightarrow\) dimension trajectory drift \(\leftrightarrow\) attested prototypical corpus examples." This safeguards the analysis against misleading artifacts such as register bias or dictionary coverage gaps.
A Worked Example¶
Consider the case of the Japanese noun sabetsu (差別, 'distinction / discrimination'):
1. Usage Disambiguation: In an 1880s political essay context ("without distinction between frontal and surprise attacks..."), Qwen3-8B compares the text with Iwanami definitions (Sense 1: discriminatory treatment; Sense 2: distinction, difference) and outputs Sense [2]. In a 2001 essay ("a struggle to win freedom from prejudice and discrimination..."), the model detects the human rights context and outputs Sense [1].
2. Distribution Redistribution: In the 1880s and 1890s, Sense 2 accounts for 84.0% and 69.4% of usages, respectively. By the 1920s, Sense 1 rises to 64.2%, and from the 1950s onward, it remains consistently above 80% (reaching 100% in the 1990s and 98.0% in the 2000s).
3. Multi-Dimensional Trajectory Attribution: The valence trajectory displays a steep negative slope (\(\beta = -0.0124, R^2 = 0.564\)), reflecting a sharply deteriorated evaluative environment. Because Sense 1 encodes pejorative social discrimination while Sense 2 is neutral, the sense distribution shift directly explains the pejorative shift from a neutral cognitive distinction to social discrimination.
Key Experimental Results¶
Main Results¶
The framework is evaluated on three Japanese historical and contemporary corpora from NINJAL: the Corpus of Historical Japanese (CHJ, 1870s–1920s), the Showa-Heisei Corpus (SHC, 1930s–2010s), and the Balanced Corpus of Contemporary Written Japanese (BCCWJ, 2000s), encompassing 47.7M words (213,650 usages across 150 target words).
Table 1 presents the intrinsic evaluation of the dictionary-grounded WSD module on 50 target words and 976 manually annotated usages (annotated by 3 native Japanese linguistics experts with mean pairwise Cohen's \(\kappa = 0.798\)), compared against trivial and retrieval baselines.
| Method | Mean Annotator Accuracy (Macro) | Majority Vote Accuracy |
|---|---|---|
| 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 |
| Most Frequent Sense (MFS, upper bound) | 0.695 | 0.699 |
| Qwen3-8B Thinking (Ours) | 0.793 | 0.815 |
Table 2 evaluates the extrinsic correlation (Spearman's rank correlation \(\rho\)) between LSC-Pattern sense distribution divergence (JSD) and human semantic change judgments from JaSemChange.
| Evaluation Subset | Corpus Pair / Setting | Spearman's \(\rho\) | \(p\)-value |
|---|---|---|---|
| Polysemous Overlap Subset (7 words) | SHC–BCCWJ (1930s–2000s) | 0.821 | 0.023 |
| Polysemous Overlap Subset (7 words) | CHJ–BCCWJ (1870s–2000s) | 0.750 | 0.052 |
| Full Benchmark (All 20 words) | SHC–BCCWJ (1930s–2000s) | 0.580 | 0.007 |
| Full Benchmark (All 20 words) | CHJ–BCCWJ (1870s–2000s) | 0.367 | 0.112 |
Ablation & Case Study Analysis¶
Table 3 summarizes the decadal linear regression slopes and \(R^2\) values across all three semantic dimensions for the three representative case studies (bold numbers denote the foregrounded primary pattern).
| Target Entry (Lemma) | Illustrated Pattern | Valence Slope | Semantic Scope Depth Slope | Metaphoricity Slope |
|---|---|---|---|---|
| sabetsu (差別) | Pejoration (恶化) | -0.0124 (\(R^2=0.564\)) | -0.0533 (\(R^2=0.734\)) | 0.0121 (\(R^2=0.478\)) |
| meshi (飯) | Generalization (泛化) | 0.0123 (\(R^2=0.215\)) | -0.0193 (\(R^2=0.132\)) | 0.0038 (\(R^2=0.018\)) |
| dodai (土台) | Metaphorization (隐喻化) | 0.0079 (\(R^2=0.021\)) | 0.0000 (constant) | 0.0404 (\(R^2=0.576\)) |
Key Findings¶
- LLM Disambiguation Exhibits Robust Historical Generalization: Qwen3-8B achieves 0.815 majority-vote accuracy, outperforming the strongest retrieval baseline (0.681) and the optimistic MFS baseline (0.699). Crucially, accuracy remains high across distinct historical periods: 0.803 on CHJ (1870s–1920s), 0.813 on SHC (1930s–2010s), and 0.844 on BCCWJ (2000s), demonstrating that historical orthographic variation and archaic styles do not compromise LLM reasoning.
- Inventory Granularity Governs Benchmark Correlation: On words with multiple dictionary senses in Iwanami, sense distribution JSD correlates strongly with human change judgments (\(\rho = 0.821\)). On the full 20-word benchmark, correlation drops to 0.580 because several target words are listed as monosemous in Iwanami, making sense redistribution impossible and yielding tied JSD scores of 0. Fixed inventories thus provide excellent interpretability when sense distinctions exist, but cannot track sub-sense semantic shifts within a single definition.
- Sense 0 Concentrates in Identifiable Lexical Gaps and Disambiguation Pitfalls: Across all 213,650 usages, Sense 0 accounts for only 1.42%. Two entries account for nearly 40% of all Sense 0 cases: in (93.17% Sense 0) represents a genuine dictionary coverage gap (corpus usages refer to 'inner wear' or the English preposition, while Iwanami only lists sports senses), whereas kita ('north', 68.42% Sense 0) represents a prompt parsing error where the model misinterprets a parenthetical gloss mentioning 'south' and rejects the correct sense. Excluding these two anomalies, the Sense 0 rate across the remaining 148 entries is just 0.87%.
Highlights & Insights¶
- Transforming Unsupervised Change into Supervised Sense Redistribution: By leveraging general-purpose LLM reasoning to map usages into lexicographically grounded senses, the framework converts opaque latent embedding shifts into transparent probability shifts across documented dictionary entries.
- Bi-directional Decoupling of Semantic Dimensions and Senses: Macro property trajectories (valence, breadth, metaphoricity) are anchored to underlying dictionary senses. This prevents false conclusions—such as mistaking a shift in discourse topic for genuine lexical pejoration—by requiring physical confirmation from sense distribution shifts.
- Principled Triangulation between Geometry, Lexicon, and Real Text: The framework mandates human inspectability by presenting aligned diachronic sense curves, property trajectories, and attested historical sentences, establishing a reliable methodology for computational lexicography.
Limitations & Future Work¶
- Coverage Constraints of General-Purpose Modern Dictionaries: The Iwanami dictionary represents modern usage and inevitably lacks obsolete historical senses from the Meiji era as well as emerging contemporary slang. Integrating dynamic novel sense induction (WSI) with fixed inventories represents an important next step.
- WordNet Depth as an Imperfect Proxy for Semantic Scope: WordNet taxonomic depth is an approximate proxy that does not always reflect nuanced cognitive specificity (e.g., both physical foundation and abstract basis of dodai share depth 8 in WordNet, rendering semantic scope blind to this metaphorization).
- Corpus and Genre Heterogeneity: While restricting texts to the magazine register minimizes genre variation, differences in editorial standards and temporal overlaps between SHC and BCCWJ in the 2000s introduce residual corpus noise that warrants further controlled normalization.
Related Work & Insights¶
- vs. Contextualized Embedding Clustering (Kutuzov et al., 2022; Schlechtweg et al., 2020): Unsupervised clustering produces unlabelled clusters that vary across runs and lack lexicographic boundaries. LSC-Pattern anchors all usages to predefined dictionary senses, making changes immediately verifiable by human linguists.
- vs. Direct Multi-Dimensional Change Analysis (Baes et al., 2024): Previous dimensional methods trace word-level shifts (e.g., sentiment drops) directly from context words without knowing which sense changed. LSC-Pattern explains why the dimension shifted by attributing the trajectory to the redistribution of specific dictionary senses.
Rating¶
- Novelty: ⭐⭐⭐⭐☆ Bridges LLM-based dictionary-grounded WSD with multi-dimensional diachronic trajectories, establishing an inspectable paradigm for LSCD.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ 150-year diachronic span across three major Japanese corpora (47.7M words, 213K usages), 976 expert annotations, 5 baselines, and rigorous case studies.
- Writing Quality: ⭐⭐⭐⭐⭐ Methodological design, error analysis of Sense 0, and limitations are articulated with exemplary clarity and academic rigor.
- Value: ⭐⭐⭐⭐⭐ Delivers an actionable, linguistically sound analytical framework for historical semantics, computational linguistics, and lexicography.