每天早上一封邮件，把昨天的 AI 梳理好订阅邮件

METAL LAB

SynFlow: A Multidimensional Diachronic Semantic Analysis Toolkit

arXiv:2608.194722026-08-21

一款把单词意义变化拆解到语法细节的开源分析工具

SynFlow是一款开源工具包，它不像传统的向量空间方法那样只给出一个笼统的变化分数，而是从句法、词形、结构搭配等多个维度追踪词义的历时变化。研究团队以德语形容词viral为例，展示了它从“病毒性的”演变为“在网上走红”这一意义变化，如何同时体现在多个语言维度上。在SemEval-2020 Task 1基准测试中，该工具的表示方法排名超过了许多基于神经网络的系统，同时保持了可解释性。

他们做了什么

  1. 传统基于向量空间的词汇语义变化(LSC)方法只能指出一个词发生了变化，却难以说明具体是哪方面的用法在变
  2. SynFlow把依存句法共现、词形特征、结构搭配组合，以及Frame Semantics等外部表示统一转换成按时间段划分的分布，用同一套工作流程处理
  3. 支持余弦距离、詹森-香农散度(JSD)、总变异距离(TVD)等多种距离度量，并提供置换检验（一种统计显著性检验方法）和数值级别的变化分解
  4. 还提供基于word2vec词嵌入的增量聚类功能，用来区分单纯的词汇替换和更广泛的主题性变化
  5. 案例研究显示，德语形容词viral在2011至2015年前后从形容词用法转向副词用法（如viral gehen，意为“走红”），这一变化在句法关系、词汇搭配、结构组合和词形特征等多个维度上都保持一致
Figure 2: Change of normalised frequencies of all slot types of viral.
Figure 2: Change of normalised frequencies of all slot types of viral.
Table 1: Dimensions and example values natively extracted by SynFlow.
TypeDimensionExample valuesInterpretation
Slot typeSlotTypechi_nsubj, chi_objgrammatical profile
Slot fillerFILLER[chi_obj]cakelexical preferences
ConstructionConstruction[chi_nsubj + chi_obj]multi-slot configurations
Feature typeFeatureTypeNumber, Tensemorphological profile
FeatureFEAT[Number]Sing, Plurvariation within feature
Figure 3: Contributions of individual slot types to TVD of viral. Green bars indicate an increase in relative frequency; red bars indicate a decrease.
Figure 3: Contributions of individual slot types to TVD of viral. Green bars indicate an increase in relative frequency; red bars indicate a decrease.
Table 2: Ranking of previously evaluated SynFlow representations against systems reported for the English SemEval-2020 benchmark. Full results are reported in 32; 31.
RepresentationSubtask 1Subtask 2
Slot filler4th /2812th /28
Frame Semantics9th /289th /28
Figure 4: Weighted TVD trajectories for slot-filler distributions across dependency slots.
Figure 4: Weighted TVD trajectories for slot-filler distributions across dependency slots.

为什么重要

以往语言学研究者需要用不同工具分别做句法画像、词形分析和基于向量的语义变化检测，SynFlow把这些整合进一个统一工作流程，让分析更快、结果更易解释。研究还证明该工具在普通笔记本电脑上就能实际处理大规模历时语料库（20个时期、超过11.6万个已解析文件），说明它已经具备实际科研可用性。

Figure 5: Incremental clustering results with social-media-related clusters growing rapidly from 2011.
Figure 5: Incremental clustering results with social-media-related clusters growing rapidly from 2011.

本文术语

  • 词汇语义变化(LSC) · 研究词语含义随时间演变的语言学领域
  • 总变异距离(TVD) · 衡量两个概率分布之间差异程度的统计方法
  • 詹森-香农散度(JSD) · 一种对称的统计方法，用于衡量两个概率分布之间的差异
  • 置换检验 · 通过随机打乱数据来检验观测到的变化是否可能只是偶然结果的统计方法
  • Frame Semantics · 一种语言学理论，通过特定情境中的角色结构来描述词义
Figure 6: Frequencies of different constructions around viral
Figure 6: Frequencies of different constructions around viral

论文原文摘要（英文）

Lexical semantic change (LSC) is commonly modelled through vector-space representations, but these approaches often provide limited insight into which aspects of usage are changing. Diachronic corpus research instead examines interpretable dimensions such as syntactic behaviour, morphology, and constructional patterns, but typically through separate analytical workflows. We present SynFlow, an open-source toolkit for multidimensional diachronic analysis of linguistic usage. SynFlow converts linguistic observations into period-specific distributions and applies a shared workflow across dependency-based co-occurrences, morphological features, constructional configurations, and externally derived representations such as Frame Semantics. It supports different distance measures, together with value-level decomposition, statistical testing, and incremental clustering of lexical fillers. We demonstrate SynFlow through a qualitative case study of the German adjective viral, showing how a single semantic development is reflected across syntactic, lexical, constructional, and morphological dimensions. We further report previously published results on SemEval-2020 Task 1 to situate the performance of these representations relative to existing lexical semantic change detection systems.

作者 · Bach Phan-Tat, Kris Heylen, Dirk Geeraerts, Stefano De Pascale, Dirk Speelman

在 arXiv 阅读

最新论文

全部论文 →

METAL LAB 最新报道

图片来源： Bach Phan-Tat et al., arXiv:2608.19472, CC BY 4.0