主旨演讲

The so-called "verb-doubling" (or "verb-copying") construction (VDC) has recurrently come into focus in studies on the syntax and semantics of Mandarin Chinese, ever since it was brought to the fore by Huang (1982). Its basic schema involves two identical spellouts of a verb V, the first followed by its internal argument (object), and the second by some complementative element. Two major lines of analysis can be distinguished: one assuming the two occurrences of V are movement-related, and another holding they are base-generated independently. The talk reassesses earlier analyses in light of recent contributions, exploring whether "VDC" is in fact a cover term for at least two different constructions, where one type is characterized by strict form-identity of V1 and V2 (potentially subject to movement analysis), while in the other, perfect form-identity is not strictly required (requiring a base-generation account).

在上古汉语中，动词的地位十分凸显。上古汉语中的动词不仅包含动词性信息，还往往包含名词性信息。把动词作为方法，用动词的视角重新审视上古汉语，就会发现上古汉语句法至少有以下三个特点：（1）一些按现代汉语语感理解为名词的词，在上古汉语中其实是动词；（2）一些按照现代汉语语感会理解为介词或连词或副词的词，在上古汉语其实是动词；（3）一些所谓的连动或动宾结构，在上古汉语中很可能是动词并联结构。从动词出发来把握名词，最早是Yaska在Nirukta中提出来的研究方法。本文认为Nirukta的这一观点对今天的语言学分析仍然具有很大的启发意义。

自结构主义语言学开始，对于并列结构的句法分析一直是句法界的难题。原因之一是有一类语料长期被主流所忽视：基于绑定条件A、C的测试，并列项之间不存在成分统御关系（c-command），跟偏正结构中的修饰与被修饰成分形成鲜明对照。设语义关系向结构关系的投射必须同构（isomorphic），即两者之间存在象似性，则可以通过最简方案中的省力原则和标签机制确保在并列项之间不存在成分统御。本文涉及普遍语法和功能语法是否需要合作、怎样实现合作的问题，而合作的关键在于理论互补：普遍语法长于精确表述种种非对称语法关系，而象似性则作为功能语法中的核心元素，二者合作可以完整解释偏正结构和并列结构的句法差异。

One defining characteristic of the Chinese language is its lack of inflectional morphology (tense, number, case, etc.), which sets it free to build a variety of quantificational and predicational dependencies through Merge instead of Move. The talk examines eight types of structural inner-outer dichotomy in Mandarin Chinese — including wh-adverbials, reflexive adverbials, causative light verbs, foci, affectives, A-not-A questions, deontic modals, and bouletic modals — arguing that Chinese displays a most fully-developed spectrum of analyticity effects. This approach provides a straightforward way to link synchronic with diachronic syntax, as almost every outer expression can be shown to derive from grammaticalization of its inner counterpart, introducing new types of speaker-oriented adverbial expressions in the "treetops" of the clausal spine. It also provides a new way to map different parts of the clausal spine to their respective semantic/pragmatic functions in the history of the Chinese language.

This talk investigates English and Korean speakers' oral production of L2 Chinese elliptical sentences — sentences with Chinese verb-phrase ellipsis licensed by modal verbs, and sentences with ΣP-ellipsis licensed by the Chinese auxiliary shi 'BE'. An elicited imitation task was administered to learners at beginner, intermediate and advanced levels. Results show that instead of elliptical sentences, learners at early stages produced significantly more complete sentences with no ellipsis, challenging syntactic complexity analyses based on sentence length and dependency distance. No L1 influence was observed at beginner levels, disconfirming an influential hypothesis by Schwartz & Sprouse (1994, 1996); however, L1 influence appeared at intermediate and advanced levels. The absence of L1 influence at beginner levels is attributed to a breakdown at the syntax-stylistics interface and to beginners' difficulty in implementing checking and deletion operations in L2 oral production.

Large language models (LLMs) are among the most consequential recent developments in natural language processing (NLP) and computational linguistics. This talk provides a technically grounded overview of major contemporary LLM architectures, their strengths and limitations, clarifying key distinctions: quantity vs. quality, creation vs. innovation, rule deduction vs. probabilistic imitation, NLP vs. AI, and AI vs. AGI. On empirical grounds, current LLMs are best viewed as advanced NLP systems rather than genuine artificial intelligence, predictably lacking robust causal reasoning. Explaining and improving "black-box" model behavior necessarily requires linguistic analysis, and core issues in AI explainability — such as "unintentional hallucinations" — demand linguistic theory for adequate diagnosis. LLMs track statistical co-occurrence patterns over symbols but do not genuinely encode rich semantic relations; addressing this semantic deficit lies at the heart of linguistic inquiry.