Learning from Social Media Texts

近年来，在自然语言处理(NLP)和机器学习(ML)领域的应用变得非常流行。这是由于数据和评估基准的可获得性、算法的进步，以及在我们的日常生活和商业产品开发中对这些应用程序的需求增加所致。我建议将NLP和ML技术应用于社交媒体中的用户建模。它有三个具体的目标：(1)从社交媒体文本中了解用户特征。这些特征可能包括：年龄、性别、个性类型、地点、种族、健康问题、兴趣和生活事件。(2)了解各种社交媒体(如Twitter、论坛、Facebook)中这些特征的人口分布情况。(3)将提取的信息作为概念证明，用于市场研究和健康监测等应用。我提出的科学方法是基于机器学习、深度学习、自动文本分类和信息提取技术。我们将重点关注社交媒体文本，由于拼写不规范、缺乏编辑、缩写、行话和噪音，这些文本比常规文本更具挑战性，技术需要适应这类文本。调整它们的各种方式包括再培训、消息的部分标准化，以及添加每种类型的社交媒体特有的功能。此外，我还建议整合利用社交网络结构的技术。以前关于相关主题的大部分工作要么使用消息文本，要么使用网络结构。我认为，将它们结合起来可能会提高提取信息的精度。我们还将特别注意保护社交媒体用户的隐私。拟议工作的新颖性包括对现有技术进行全面调查，提高其复杂性，为拟议任务开发新技术(特别是在人口一级，这是一个研究较少的领域)，以及开发几个概念验证应用程序，这些应用程序需要有关用户或社交媒体用户群体的信息。我们预计，研究结果将有助于更好地了解社交媒体中的人类交流，这将有助于在挖掘社交媒体用于市场研究和其他决策应用方面取得进展，并将加强加拿大作为信息技术领域主要参与者的地位。

Applications in the field of Natural Language Processing (NLP) and Machine Learning (ML) have become popular in recent years. This is due to the availability of data and evaluation benchmarks, to progress in the algorithms, and finally to an increased need for these applications in our daily life and in commercial product development. I propose to apply NLP and ML techniques for user modelling in social media. There are three specific objectives: (1) Learn user characteristics from social media texts. These characteristics may include: age, gender, personality type, location, ethnicity, health issues, interests, and life events. (2) Learn population distributions in various social media (e.g., Twitter, forums, Facebook) for each of these characteristics. (3) Use the extracted information, as a proof of concept, in applications such as marketing research and health monitoring. The scientific approach that I propose is based on machine learning, deep learning, automatic text classification, and information extraction techniques. We will focus our attention on social media texts, which are more challenging than regular texts due to non-standard spelling, lack of editing, abbreviations, jargon, and noise, The techniques need to be adapted to this kind of text. The various ways of adapting them include retraining, partial normalization of the messages, and adding features specific to each type of social media. In addition, I propose to integrate techniques that exploit the structure of the social network. Most of the previous work on related topics uses either the texts of the messages or the network structure. I believe that combining them may lead to an increase in the precision of the extracted information. We will also pay special attention to protecting the privacy of social media users. The novelty of the proposed work consists of a comprehensive investigation of the existing techniques, in increasing their sophistication and in developing new techniques for the proposed tasks (particularly at population level, a less-studied area), as well as in the development of several proof-of-concept applications that require information about users or about populations of users in social media. We anticipate that the outcomes of the research will contribute to better understanding of human communication in social media, which will allow advances in mining the social media for market research purposes and other decision-making applications, and will strengthen Canada's position as a major player in the field of information technology.