🔄 Self-Supervised Learning

💬 ACL2026 · 2 paper notes

[b] = [d] − [t] + [p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic

This paper systematically demonstrates that linear phonological feature vectors exist in the representation space of self-supervised speech models (S3M), satisfying word2vec-like vector arithmetic relations, with their scaling factors continuously correlating with acoustic measurements.

ConlangCrafter: Constructing Languages with a Multi-Hop LLM Pipeline

This paper proposes ConlangCrafter, a multi-hop LLM pipeline that decomposes constructed language (conlang) design into three modular stages — phonology, grammar, and lexicon — ensuring typological diversity through randomness injection and internal consistency through self-refinement loops, along with an automatic evaluation framework incorporating typological diversity analysis and translation consistency assessment.