Texture Oracle: A Software Service for Predicting and Responding to COVID-19 Outbreaks in UK Regions Using Twitter Data

**雄心：**纹理甲骨文将使用推特数据(1)预测英国邮政编码中的新冠肺炎疫情，(2)评估旨在预防或应对疫情的公共卫生消息的有效性。学术研究表明，可以通过语言使用模式来预测疾病的爆发；我们将利用团队成员的原创研究来改进现有方法，并将其应用于新冠肺炎\。同样，人工智能和自然语言处理的进步使我们有可能在社交媒体上建立一条消息的“回声”：这将使我们能够对公共卫生消息的有效性进行评级。因此，甲骨文将为应对新冠肺炎的生物医学和公共卫生应对措施提供一层独立的基于语言数据的预测和评估能力。**预测新冠肺炎：**我们旨在通过两种方法预测某个地区爆发新冠肺炎的可能性增加：*_跟踪危险行为：当人们没有遵守社会距离和隔离措施时，新冠肺炎更有可能发生。通常，承担这种风险的人会体验到未来的回报不如现在的回报有价值；因此，避免感染被认为没有短期满足感有价值。未来回报贬值的趋势可以从一段语言提到未来的方式中推断出来：通常，它被表示为不那么确定。使用项目成员开发的人工智能方法，我们将在特定地区的推文中衡量这一趋势，并将其用于预测新冠肺炎爆发*_跟踪症状_：可以通过跟踪推文中提到的症状来预测流感爆发。项目成员的研究通过提供一种新的资源来跟踪症状表达，从而扩展了这种方法。兰开斯特感觉运动规范将40k个英语单词按照它们唤起六种知觉模式(触觉、听觉、嗅觉、味觉、视觉和体内感觉(内感))和五种动作效应器(嘴/喉咙、手/胳膊、脚/腿、头(不包括嘴/喉咙)和躯干)的程度进行分类。这在日常语言和身体状态之间建立了比症状术语更强的联系，使我们能够使用推文的感觉运动特征来预测疫情爆发。**评估公共卫生信息：**政府信息导致启动效应，这意味着有效的信息以人们的术语和语言结构重现。我们将使用人工智能方法提取推文中的这些“回声”，使我们能够识别公共卫生信息在地区基础上的渗透情况。这样做将为公共卫生机构提供所需的信息，以便将其传播目标(和重新目标)指向风险最高的地区。

**Ambition:** Texture Oracle will use Twitter data to (1) predict COVID-19 outbreaks in UK postcodes and (2) evaluate the effectiveness of public health messaging designed to prevent or tackle outbreaks. Academic research shows that disease outbreaks can be predicted by way of patterns in language use; we will use team members' original research to improve existing methods and apply them to COVID-19\. Similarly, advances in AI and natural language processing make it possible to establish a message's 'echo' in social media: this will allow us to grade the effectiveness of public health messaging. Oracle will therefore complement biomedical and public health responses to COVID-19 with an independent layer of predictive and evaluative capacity based on language data.**Predicting COVID-19:** There are two ways in which we aim to predict the increased likelihood of a COVID-19 outbreak in an area:* _Tracking risky behaviour_: COVID-19 is more likely to occur when people fail to observe social distancing and quarantine measures. Typically, people responsible for this kind of risk-taking experience future rewards as less valuable than present rewards; thus, avoiding infection is felt as less valuable than near-term gratification. The tendency to devalue future rewards can be inferred from how a piece of language refers to the future: typically, it is represented as being less certain. Using AI methods developed by project members, we will measure this tendency in region-specific tweets and use it to predict COVID-19 outbreaks* _Tracking symptoms_: Influenza outbreaks can be predicted by tracking symptom mentions in tweets. Research by project members extends this method by providing a new resource for tracking symptom expression. The Lancaster sensorimotor norms classify 40k English words for the extent to which they evoke six perceptual modalities (touch, hearing, smell, taste, vision, and feelings inside the body (interoception)) and five action effectors (mouth/throat, hand/arm, foot/leg, head (excluding mouth/throat), and torso). This creates a much stronger link between everyday language and bodily states than symptom terms alone, allowing us to use the sensorimotor profile of tweets to predict outbreaks.**Evaluating public-health messaging:** Government messaging causes priming effects, meaning that effective messages are reproduced in people's terminology and language structure. We will use AI methods to extract these 'echoes' in tweets, allowing us identify how well public health messages have penetrated on a regional basis. Doing this will give public health agencies the information needed to target (and re-target) their communications to areas most at risk.