An Empirical and Computational Investigation of Generalisation in Nonword Reading

将文本转换为语音是认知科学、人工智能和技术研究人员感兴趣的问题。对这一过程的调查已经广泛使用了计算模型和经验方法来揭示阅读背后的机制（colheart, 2005）。由于人类读者经常会遇到新单词，对阅读模型的一个关键测试是它是否能从它的训练集中泛化——也就是说，它是否能像人类一样，对以前从未遇到过的单词产生同样的发音。由于泛化的重要性，阅读非词已成为评估阅读模式的关键方法。以往的绝大多数工作都集中在单音节单词和非单词的建模阅读上。然而，排除多音节单词并不能提供阅读英语的代表性说明。为了纠正这种情况，Mousikou等人（2017）将当前模型的表现与人类在双音节非单词发音中的表现进行了比较。在此过程中，作者为英式英语创建了第一个规范的非词语料库。对双音节阅读中的模型和人类反应的比较使作者得出结论，这些模型并不能提供双音节阅读过程的代表性说明。模型仍然能发出人类参与者无法发出的发音。因此，我们离能够真正像人类一样阅读非单词的模型还很远。建模非单词发音的一个挑战是人类非单词阅读的高度可变性，即使在熟练的读者中也是如此。这种差异的来源似乎部分在于个别读者，部分在于个别项目（Coltheart & Ulicheva, 2018）。遵循从拼写到发音的规则被认为是应用于非单词阅读的主要机制，但依赖词汇知识也被证明发挥了作用，至少对于一些非单词（Andrews & Scarratt, 1998）。因此，非单词阅读的可变性可能部分基于个体对词汇知识和亚词汇知识的依赖差异，或者拼写到声音规则和词汇知识的差异。目前博士项目的第一个目标是开发一个大声朗读双音节单词的模型。为此，我们将利用最近开发的英国英语双音节非词语料库（Mousikou et al. 2017）来识别现有模型中的问题。根据这些调查，将开发一个具有现有模型的潜在组合特征的新模型。该项目的第二个目标是收集经验非词阅读数据，以评估模型的泛化能力，并探索人类非词阅读的可变性。后者可以通过比较不同读者群在非单词发音中的表现来探索。更具体地说，数据收集将包括来自参与者的两种类型的反应：非词的发音和由新开发的模型给出的非词发音评级。在比较人类和模型的表现时，利用定性发音数据的丰富性和定量评分数据的客观性可以最大限度地提高这些调查的效率。对这种细粒度数据的分析还可以促进对非词阅读模式的检测，这些模式以特定的读者亚群为特征。更好地理解大声朗读有可能有助于更好地筛查和支持有阅读困难的人群。此外，更准确的大声朗读模型可以为辅助技术的发展提供信息，比如为视障人士设计的语音合成器。推动这些应用程序发展的前景使得理解大声朗读成为一项非常有价值的努力。

Converting text to speech is of interest to researchers in cognitive science, artificial intelligence and technology. Investigations of this process have widely utilized computational modelling and empirical approaches to uncover the mechanisms underlying reading (Coltheart, 2005). As human readers routinely encounter new words, a critical test for a model of reading is whether it can generalize from its training set - i.e. whether it can produce the same kind of pronunciations as humans do for items it has never encountered before. Because of the importance of generalisation, reading nonwords has become a key method in evaluating models of reading.The vast majority of previous work has focused on modelling reading of monosyllabic words and nonwords. However, excluding multisyllabic words does not provide a representative account of reading English. In an attempt to remedy this situation, Mousikou et al. (2017) compared the performance of current models against human performance in disyllabic nonword pronunciation. In doing so, the authors created the first normative nonword corpus for British English. Comparisons of model and human responses in disyllabic reading led the authors to conclude that these models do not provide a representative account of processes involved in disyllabic reading. Models still produced pronunciations that none of the human participants produced. As such, we are still far from a model capable of truly human-like nonword reading.One challenge in modelling nonword pronunciation is the high variability in human nonword reading, even among skilled readers. The source of this variability seems to be partly in individual readers and partly in individual items (Coltheart & Ulicheva, 2018). Following spelling-to-sound rules is believed to be the main mechanism applied in nonword reading, but relying on lexical knowledge has also been shown to play a role, at least for some nonwords (Andrews & Scarratt, 1998). The variability in nonword reading may thus be partly based on individual differences in reliance on lexical vs sublexical knowledge, or differences in spelling-to-sound rules and lexical knowledge.The first aim of the current PhD project is developing a model of reading aloud disyllabic words. To this end, the recently developed corpus of disyllabic nonwords in British English (Mousikou et al. 2017) will be utilized to identify issues in the existing models. Informed by these investigations, a new model with potentially combined features from the existing models will be developed. The second aim of the project is to collect empirical nonword reading data against which to assess the model's generalisation ability, and to explore the variability in human nonword reading. The latter can be explored by comparing the performance of different subgroups of readers in nonword pronunciation. More specifically, the data collection would include two types of responses from the participants: pronunciation of nonwords and ratings of nonword pronunciations given by the newly developed model. Utilizing the richness of the qualitative pronunciation data and the objectivity of the quantitative ratings data in comparisons of human and model performance would maximise the efficiency of these investigations. Analysis of such fine-grained data could also facilitate detection of patterns in nonword reading characterised by particular subgroups of readers.Better understanding of reading aloud has the potential to contribute to the development of better screening and support for populations with reading difficulties. Additionally, more accurate models of reading aloud can inform the development of assistive technologies, such as speech synthesisers for the visually impaired. The prospect of advancing the development of these applications makes understanding reading aloud a highly valuable endeavour.