Investigating the diffusion of morphosyntactic innovations using social media

我们建议研究语言语法的变化（“创新”）是如何从少数使用者传播到更大一部分人口（“扩散”）的。用传统方法收集的数据类型并不特别适合研究扩散过程。传统上，方言学家通过调查不同地区的说话者来收集资料。然而，对每个地区的大量说话者进行调查通常是不切实际的，这意味着当生活在同一地区的说话者之间存在差异时，这种调查效果很差（在持续变化期间通常是这种情况）。社会语言学家通常在一个地方采访大量的说话者；然后通过比较这些数据集来研究扩散过程是可能的，但这只能提供非常有限的地理范围。因此，我们建议使用来自社交媒体平台（如Twitter）的数据来调查传播过程。虽然是一种书面语言，但社交媒体数据往往是非正式的，通常提供的数据与口头数据很接近。通过使用来自社交媒体的数据，我们将能够在大范围内从许多地方收集大量的本地化数据。我们将探索使用社交媒体数据来调查语言变异和变化的新方法，并展示使用这些数据来调查地理空间扩散和语法模式变化的有效性。我们将使用Twitter收集三个数据集（“语料库”），其中包括英国的英语和威尔士语以及挪威的挪威语，时间跨度为9个月。这三种语言的选择使我们能够比较不同的人口和地理情景对传播模式的影响：威尔士语作为少数/较少使用的语言与英语和挪威语作为多数语言；挪威人口密度低，而英格兰大部分地区人口密度高；等。我们将从挪威语和威尔士语特定的社交媒体上通过数据抓取获得额外的数据来补充这些Twitter语料库。然后，我们将确定目前在这些人群中扩散的语言变化，并调查它们在这些语料库中的分布。我们期望调查的变化包括威尔士语中新的第二人称代词“chdi”（“你”）的传播，挪威语中带有“komme til <s:1>”和“bli <s:1>”的将来式的传播，以及英语中带有“需要”的结构的句法变化（例如“你需要洗头”和“你需要洗头”）。“你需要洗头了”)。通过识别语料库中创新选项和旧选项的所有实例，我们将能够映射出每个不同选项通常使用的位置。通过将这些发现与早期研究中已知的地理模式或在其他数据集中确定的地理模式进行比较，我们将能够绘制出新形式随着时间的推移而传播的地图，从而确定扩散过程的特性。通过这种方式，我们将能够回答以下问题：这些语言社区的变化是在陆地上持续传播（“传染性扩散”），还是在到达农村地区之前从一个城市跳到另一个城市（“分层扩散”）？这种扩散模式是否受到所讨论的变化类型、该地区的人口结构或其他因素的影响？我们将使用语料库中用户之间的交互（转发、@direct消息、相互关注）来构建这些用户的社交网络模型。然后，我们将能够比较该网络模型作为扩散路径预测器与地理模型的有效性。我们的研究结果将通过网络应用程序进行演示，这些应用程序通过用户对语言使用问题的回答来预测用户的母语，此外，这些结果还将通过一个在线地图册式的网站向公众公布。

We propose to research the way in which changes in the grammar of languages ("innovations") spread out from a small number of speakers to a larger section of the population ("diffusion"). The types of data gathered by traditional means are not especially well-suited to studying diffusion processes. Dialectologists have traditionally gathered data by surveying speakers at a large range of different localities. However, it is usually impractical to survey a large number of speakers at each locality, meaning that such surveys do poorly when there is variation between speakers living in the same places (as is usually the case during ongoing change). Sociolinguists typically interview a large number of speakers at a single locality; it is possible to then investigate diffusion processes by comparing such datasets, but this only ever offers a very limited geographical range. Accordingly, we propose to use data from social-media platforms such as Twitter to investigate diffusion processes. Although a form of written language, social-media data tend to be highly informal and often provide a good approximation to spoken data. By using data from social media, we will be able to gather a very large quantity of localised data from many places across large areas. We will explore new ways of using social media data to investigate language variation and change and demonstrate the effectiveness of using such data to investigate geospatial diffusion and change in grammatical patterns.We will use Twitter to collect three datasets ("corpora") of tweets in English and Welsh in Britain, and in Norwegian in Norway covering a period of nine months. The selection of these three languages enables us to compare the effect of different demographic and geographic scenarios on patterns of diffusion: Welsh as a minority/lesser-used language vs. English and Norwegian as majority languages; low population density in Norway vs. higher population density in much of England; etc. We will supplement these Twitter corpora with additional data obtained by data-scraping from Norwegian- and Welsh-language-specific social media.We will then identify language changes currently diffusing in these populations and investigate their distribution in these corpora. Changes we expect to investigate include the spread of a new second-person pronoun 'chdi' ('you') in Welsh, the spread of future constructions with 'komme til å' and 'bli å' in Norwegian and changes in the syntax of constructions with 'need' in English (such as 'you need your hair washing' vs. 'you need your hair washed'). By identifying all instances of both the innovative and the older option(s) in our corpora we will be able to map where each different option is typically used. By comparing these findings with geographical patterns known from earlier studies or identified in other datasets, we will be able to map the spread of new forms over time and so identify the properties of processes of diffusion. In this way, we will be able to answer questions such as: do changes in these speech communities spread continuously over land ("contagious diffusion") or jump from city to city before reaching rural regions ("hierarchical diffusion")? Is this mode of diffusion affected by the type of change in question, by the demographics of the region or by some other factors?We will use interactions between users in our corpora (retweets, @direct messages, mutual following) to construct a model of these users' social network. We will then be able to compare the effectiveness of this network model as a predictor of the pathway of diffusion to the geographical model. Our results will be demonstrated in action through web-apps that predict users' origins using their responses to questions about their language use and they will furthermore be made available to the public via an online atlas-style website.