Temporally Informed Distillation of Embedding Semantics: Beyond Continued Pretraining for Modeling Gender Ideology in Dated Texts
Open Access
Abstract
Modeling historically situated gender ideology remains challenging for language models, as contemporary embeddings struggle to reflect temporally specific semantic structures beyond surface lexical patterns. Although large language models exhibit extensive general-purpose performance, their direct use with history-specific semantic analysis is limited by the distributional mismatch between contemporary training data and historical linguistic patterns. These constraints encourage the distillation of temporally based semantic knowledge into small student architectures. To solve this issue, we propose Temporally Informed Distillation of Embedding Semantics (TIDES), which integrates continued pretraining on temporally specific corpora with feature-level distillation from large embedding teachers. We evaluate TIDES across teacher architectures with distinct pretraining objectives. While continued pretraining provides lexical and syntactic adaptation, our results show that improvements in ideological modeling cannot be attributed to additional training exposure alone. Rather, teacher–student structural alignment is also critical to transfer effectiveness. Contrastive, encoder-aligned teachers yield substantially more stable preservation of fine-grained, historically situated semantic distinctions. These findings suggest that temporal ideology transfer is representation-dependent: ideological meaning can be shaped by the geometry and training objectives of embedding spaces. By introducing TIDES and providing evidence that architectural compatibility can influence ideological inheritance, this study advances a representation-centered account of modeling ideology in temporally grounded semantic research.Keywords
Funding Information
- Research Grants Council of Hong Kong SAR, China (15609725)
- Hong Kong Polytechnic University (#P0051089)
- Hong Kong Polytechnic University (#P0056456)
This publication has 63 references indexed in Scilit:
- Knowledge distillation and dataset distillation of large language models: emerging trends, challenges, and future directionsArtificial Intelligence Review, 2025
- Word Embedding Bias in Large Language Models2025
- A survey on knowledge distillation: Recent advancementsMachine Learning with Applications, 2024
- RoFormer: Enhanced transformer with Rotary Position EmbeddingNeurocomputing, 2023
- Extensive study on the underlying gender bias in contextualized word embeddingsNeural Computing and Applications, 2020
- Word embeddings quantify 100 years of gender and ethnic stereotypesProceedings of the National Academy of Sciences of the United States of America, 2018
- Language and Gender2013
- The Corpus of Contemporary American English as the first reliable monitor corpus of EnglishLiterary and Linguistic Computing, 2010
- Gender and Politeness2003
- A NEW MEASURE OF RANK CORRELATIONBiometrika, 1938