linguistics

这篇论文通过在Tagalog上的受控实验（MorphAustro-Tok与BPE及打乱形态素的对照组比较），发现形态感知分词确实能提升机器翻译效果，但形态正确性本身并不贡献收益——增益仅来自分词粒度与一致性，而非语言学上正确的切分。

Georgetown linguistics researchers propose PropAlign, a multiple-summary-based method for graded proposition salience evaluation, and release PropAlignBank, a dataset with 138K alignments of 35K+ propositions. They show that summary-derived salience metrics are more consistent than direct model judgments, and that model ensembles with linguistic features approach human annotation quality.

该论文利用 Pythia 和 OLMo 等开源模型的中间检查点，系统研究了大语言模型在训练过程中形态结构、嵌入漂移、重复主导与共现效应的演化规律，发现模型在语义嵌入稳定之前就已习得形态结构，且训练数据规模有助于发散识别，而共现和重复则有负面影响。

The paper proposes using the caron and circumflex diacritics to mark Yorùbá contour tones (rising and falling) in a single syllable, addressing orthographic errors that can invert the meaning of names, and describes implementations in the WriteYoruba keyboard and TTSYoruba speech synthesizer.

The paper diagnoses systematic errors in neural morphological generation for Japanese past-tense verb inflection, focusing on a rare irregular subtype that causes disproportionate errors. Ablation experiments demonstrate that removing this subtype improves accuracy more than removing all irregular verbs, emphasizing the importance of fine-grained subclass analysis.

The study examines how large language models generate and detect fake news under different scenarios, revealing variations in performance and that refined prompts do not always improve detection.

This paper presents evidence supporting the Pastiche Hypothesis for the Voynich Manuscript by applying large language models to analyze its structure and generative grammar.

This blog post analyzes the rule for using 'a' versus 'an' in English based on phonetic sounds, with data showing only 129 out of 32,455 words need exceptions, and discusses its application in procedural text generation code.

A new preprint shows that many neural language models exhibit abstract representation separation for grammaticality, addressing debates about confounds in measuring model grammatical knowledge.

This arXiv preprint analyzes the historical development of spoken political language in the United States, likely employing computational or linguistic methods.

This study uses word embeddings to analyze reduplicative constructions in Mandarin Chinese, revealing semantic and pragmatic differentiation and validating distributional semantics for linguistic investigation.

This paper introduces a data-free universal prior over syntactic structures from a cognitive model of language production, demonstrating it assigns higher probabilities to attested trees and correlates with language-specific data across 138 languages.

This study examines charismatic prosody in bilingual political speeches across Luxembourgish and French, finding that speaker identity accounts for most variance while language shows systematic differences in acoustic features.

This study investigates temporal organization in bilingual political speech, revealing that consonant-based timing metrics are speaker-specific while vowel-based metrics are language-dependent, with French exhibiting longer and more variable vowels.

The paper introduces NegCue, a large-scale negation dataset, and shows that pre-training with affixal negation yields better improvements for negation understanding in language models compared to single-word negation.

This paper proposes using semantic uncertainty derived from large language models to anticipate transition relevance places in spoken turn-taking, showing improved performance over baselines in dialogue systems.

This tweet reflects on a rewarding conversation about linguistics versus NLP, the Bitter Lesson, and the evolution of the Transformer architecture, linked to a podcast episode on AI tokenomics.

This paper examines language models as linguistic artifacts, using Italian models to explore their training and questioning their representation of language. It argues for a distinction between technical products and research tools in NLP.

This paper discusses how aligning large language models with brain functions underdetermines neural computation, highlighting limitations in modeling cognition.

This paper disentangles statistical preemption from entrenchment in language models' avoidance of overgeneralizations through controlled experiments, finding that LMs exhibit abstract preemption rather than verb-specific preemption, with implications for human language learning.