This investigation draws from research on negative polarity item (NPI) illusions in order to explore a new and interesting instance of misalignment observed for grammatical sentences containing two negative markers. Previous research has shown that unlicensed NPIs can be perceived as acceptable when occurring soon after a structurally inaccessible negation (e.g., ever in *The bills that no senators voted for have ever become law). Here we examine the opposite configuration: grammatical sentences created by substituting the NPI ever with the negative adverb never (e.g., The bills that no senators voted for have never become law). The processing and acceptability of these sentences were studied using three tasks: a speeded acceptability judgment (Experiment 1), a self-paced reading task (Experiment 2), and an offline acceptability rating (Experiment 3). The results are consistent across measures in showing that the integration of the adverb never is disrupted by the linearly preceding but structurally inaccessible negative quantifier no in the relative clause. In our view, this pattern of results is in line with Parker and Phillips’ (2016) proposal that NPI illusions arise when the context containing the inaccessible negation has not been fully encoded by the time the NPI ever is encountered, making the embedded negative quantifier transparently available as a licensor. In a similar vein, the disruption effects observed for grammatical sentences containing two negative elements could arise if the negative quantifier is still being integrated when never is encountered, forcing the parser to deal with two negative elements simultaneously. This interpretation suggests that the same incomplete encodings that could be ameliorating the online perception of unlicensed NPIs could also be responsible for deteriorating the perception of the sentences under investigation here. This would represent an illusion of ungrammaticality. Furthermore, these results provide evidence against the speculation that NPI illusions are the consequence of misrepresenting ever as its near neighbor never, given that continuations with never are judged as unacceptable in spite of their grammaticality. Together, these findings inform the landscape of hypotheses on NPI illusions and offer valuable insights into the complexity of multiple negations and the relation between processing difficulty and acceptability.

No abstract

Resumo: No passado, os debates eram realizados em ambientes organizados em um espaço onde a participação das pessoas necessitava da sua presença física. Com a Internet, e, principalmente, a Web 2.0, um novo contexto de comunicação surgiu e os debates, antes momentâneos, agora podem ser estendidos por um período mais longo, e seu ambiente pode ser ampliado além dos limites espaciais que antes o limitavam. Com as atuais tecnologias, é possível assumir a posição de interlocutor nas discussões por meio de projetos independentes, ou via coletivos ou associações que se servem das atuais ferramentas online disponíveis. É nesse contexto que o presente trabalho, realizado a partir de uma pesquisa documental, objetiva estabelecer uma discussão sobre as particularidades e o papel de ferramentas tecnológicas no incremento de debates democráticos. Neste artigo são discutidos os conceitos de debate, a sua contribuição para a democracia, sua evolução para os debates online e algumas das ferramentas disponíveis para isso. Palavras-chave: Debates online. Web 2.0. Democracia. eDemocracia. Governo Participativo.

This paper explores the ability of Transformer models to capture subjectverb and noun-adjective agreement dependencies in Galician. We conduct a series of word prediction experiments in which we manipulate dependency length together with the presence of an attractor noun that acts as a lure. First, we evaluate the overall performance of the existing monolingual and multilingual models for Galician. Secondly, to observe the effects of the training process, we compare the different degrees of achievement of two monolingual BERT models at different training points. We also release their checkpoints and propose an alternative evaluation metric. Our results confirm previous findings by similar works that use the agreement prediction task and provide interesting insights into the number of training steps required by a Transformer model to solve long-distance dependencies.