Whether large language models (LLMs) possess linguistic intuition comparable to that of humans remains a matter of debate. Drawing on example sentences from The Syntax of Chinese, this study systematically compares the acceptability judgments of GPT-5.5 and DeepSeek-V4-Pro with those of native Mandarin speakers, The 7-point Likert-scale ratings are analyzed using a Bayesian cumulative probit mixed-effects model. The results show that GPT-5.5 yields higher latent acceptability estimates overall for both grammatical and ungrammatical sentences, whereas DeepSeek-V4-Pro yields lower estimates overall for ungrammatical sentences. Both models distinguish grammatical from ungrammatical sentences, exhibiting greater overall sensitivity to grammaticality than native Mandarin speakers. However, both calibration deviations and differences in grammatical sensitivity vary across linguistic phenomena. We argue that LLMs' linguistic judgments differ systematically from those of native Mandarin speakers. These differences may reflect statistical learning from massive amounts of data rather than the acquisition of abstract grammatical rules of the same kind as those underlying human linguistic intuition.

Signing avatars have proliferated rapidly in mainland China in public settings such as news broadcasting, legal and political affairs, and government services. However, a significant gap remains between the quality of their output and the community sign language used by the Deaf community. Systematic linguistic analysis can reveal the specific features of and underlying reasons for the avatars’ deviations from sign language grammar, thereby providing a clear theoretical basis for technological revision and consequent quality improvement. This paper analyzes randomly sampled videos of sign language avatars with respect to features such as classifiers, word order, fingerspelling, and pauses. The findings indicate that classifier use in these avatars is mostly limited to lexicalized vocabulary and fails to generate dynamic classifier predicates based on three-dimensional spatial syntax; word order is influenced by linear Chinese structure, and sign-language-specific features such as verb postpositioning are not fully realized; moreover, phenomena such as homophone and polyphone confusion, inappropriate fingerspelling, and abnormal pausing are common. The paper argues that these issues stem from the system treating sign language merely as a static list of approved signs, devoid of morphology and syntax, alongside deficiencies in Chinese text processing. Accordingly, we proposed that emphasis should be placed on multidimensional descriptions grounded in sign language linguistics, with linguistic research providing the underpinning for future development of sign language avatars.

Leveraging linguistic knowledge to support scalable and reproducible data synthesis remains an important challenge in language resource development in the era of artificial intelligence. Focusing on Chinese constructions, this study addresses three issues in evaluatiing large language models’ understanding of Chinese constructions: limited data scale, insufficient evaluation dimensions, and the difficulty of reusing expert experience in item design. To address these issues, we propose a data synthesis method based on a semi-structured construction knowledge repository (SCKR). The method organizes constructional forms, meanings, pragmatic constraints, relations among related constructions, positive and negative example corpora, and item templates into callable knowledge units. It then generates evaluation items through a three-stage process: item planning, item generation, and quality validation. Using this method, we developed the CCL-CUE V2 dataset for evaluating Chinese construction understanding. The results show that the proposed method improves the efficiency of evaluation data synthesis while maintaining linguistic quality, thereby providing a scalable data foundation for the fine-grained evaluation of Chinese construction understanding.

Large language models (LLMs) often diverge significantly from human linguistic intuition in their inference about factivity in Chinese, resulting in erroneous truth-value judgments. Grounded in a linguistic framework, this study evaluates the reasoning performance of multiple LLMs, focusing on how their judgments respond to shifts in factivity induced by different syntactic and contextual conditions. The error analysis reveals that current models struggle to detect changes in factivity triggered by syntactic downgrading, modal and aspectual markers, and variations in speaker authority and firsthand knowledge. More fundamentally, factivity inference in Chinese poses a challenge because it requires the integration of multiple layers of linguistic information. Current LLMs, however, remain heavily reliant on shallow lexical-semantic cues and exhibit systematic biases in weighting contextual information and processing conceptual structures. Through the analysis of representative cases, this study develops a linguistically grounded explanatory framework for understanding factivity inference and the reasoning errors  exhibited by LLMs.

Generative phonology (abbreviated as GP) is a theory of phonological derivations or mappings from underlying phonological forms to surface phonetic forms by a set of phonological rewrite rules, and it has long been widely accepted as the mainstream phonological theory for over 70 years, Prior to it, structuralist phonology was centered mainly on the search and establishment of phonemes. After that, the concept of phoneme has been excluded outside of GP. Many scholars and experts in China mistakenly believe that GP naturally bases its theory on structuralist phonology, and still takes “phonemes” as an its indispensable basic unit. Due to this reason, the present paper will start with the fundamental hypotheses of GP, systematically explain its theoretical principles and analytical methods, and explain in detail the theoretical reasons why there are no phonemes as well as the theoretical change from segmental to nonsegmental phonological representation in GP. This will help us comprehensively and systematically get to know retrospect and prospect of GP, accurately and deeply recognize its theoretical essence, and effectively promote its application and development in Chinese language.

This paper systematically investigates the synonymous expressions in which the negative marker occurs before or after the causative marker, focusing on the syntactic and semantic mechanisms underlying their synonymous interpretations. The analysis shows that pure causative semantics is a prerequisite for the synonymous readings of the two constructions. Strong volitional or purposive modality, together with the psychologization of caused-result predicates, construes the predicate denoting the caused event in terms of a binary opposition, thereby providing a pragmatic inferential basis for the extension of negative scope across structural layers. By contrast, factors such as lexicalization, adjacency constraints on modal verbs, and imperative constructions block this synonymous alternation. This phenomenon was already well established in pre-modern Chinese, where post-causative negation constituted the unmarked and dominant word order. An analysis based on the Force Dynamics model reveals that the causative verb + negative pattern corresponds to dynamic achievement events and tends to serve an objective narrative function, whereas the negation + causative verb + affirmative clause pattern corresponds to static non-achievement events and inherently carries counter-expectational and argumentative pragmatic implications. This phenomenon represents a distinctive manifestation of the core logic of Negative Raising within the causative domain. The two structures can thus be understood as alternative constructional choices speakers make in accordance with their pragmatic communicative purposes.

Classifying disyllabic compound words of the “parallel compound initial meaning” type (i.e. words whose original compositional meaning arises from the combination of their constituent morphemes) facilitates in-depth understanding and research of them. The following are not regarded as “parallel compound initial meaning” disyllabic compound words: words with identical written forms but different pronunciations; words whose forms contain interchangeable or differentiated characters; words with identical written forms but different morphemes; words in which one element is an abbreviation, a transliterated word, or a dialect word; and homonyms. According to the number of meanings of the two morphemes, four basic types can be distinguished: “single + single”, “single + multiple”, “multiple + single”, and “multiple + multiple”. On this basis, and according to the five relationships between the “parallel compound initial meaning” and the meanings of the constituent morphemes (the “additive meaning”, “predominantly one meaning”, “greater meaning”, “lesser meaning”, and “fused meaning”), 20 static types can be derived. Once the two generative elements of lexicalization and online word formation are incorporated, these can be expanded to 40 dynamic types. Both the static and dynamic types merit in-depth research.