Between Conjecture and Memento: Shaping A Collective Emotional Perception of the Future

Large scale surveys of public mood are costly and often impractical to perform. However, the web is awash with material indicative of public mood such as blogs, emails, and web queries. Inexpensive content analysis on such extensive corpora can be used to assess public mood fluctuations. The work presented here is concerned with the analysis of the public mood towards the future. Using an extension of the Profile of Mood States questionnaire, we have extracted mood indicators from 10,741 emails submitted in 2006 to futureme.org, a web service that allows its users to send themselves emails to be delivered at a later date. Our results indicate longterm optimism toward the future, but medium-term apprehension and confusion. Introduction and Scope The ongoing popularization and growing availability of the Internet has led, in recent years, to a sharp increase in the amount of personal, social and emotional information that is publicly and openly shared. Manifestations of mood can be found in personal blogs, online diaries, social networking websites, message boards and chat rooms. Many of these online services offer ad hoc tools to let their users annotate their writings or webpages with a description of their current mood. For example, popular services such as Livejournal (http://www.livejournal.com/), an online blogging software and community, and myspace (http://www.myspace.com/), a social networking website, allow users to add a mood indicator to their blog entries and their personal profiles, respectively. Besides user-contributed mood descriptions, online textual content of various kinds can be aggregated and analyzed to infer public mood levels. Natural language processing, text mining and machine learning can be employed to assess the mood and opinions of groups of users in an automated fashion. Such analytic methods, often referred to as sentiment analysis (Nasukawa & Yi 2003), bypass the costs of repeated, large-scale public surveys and may yield accurate results when used on extensive corpora. There exists a considerable amount of experimentation in Copyright c © 2008, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. this field. Researchers at the Information and Language Processing Systems group at the Informatics Institute of the University of Amsterdam have used blog data to deduct trends and seasonality (Balog & De Rijke 2006), forecast mood (Mishne & De Rijke 2006), and predict movie sales (Mishne & Glance 2006). Investigating a similar online corpus, Mihalcea & Liu (2006) have performed a “linguistic ethnography” to find out where happiness lies in everyday lives. Some similar analytical tools function entirely on the web: We Feel Fine (http://www.wefeelfine.org/) constantly harvests blog posts for occurrences of the phrases ”I feel” and ”I am feeling” and matches it to a pre-identified database of feelings; similarly, Moodviews (http://moodviews.com) constantly tracks a stream of Livejournal weblogs that are user-annotated with predefined moods. The results generated via the analysis of such collective mood aggregators are compelling and indicate that accurate public mood indicators can be extracted from online materials. However, such indicators are limited to near-present observations due to their source material, i.e. mood indicators extracted from presently available material pertain to past and present moods. They need to be statistically extrapolated to assess future public mood or sentiment. In this paper, we investigate the content of emails submitted by anonymous users to the futureme web site (http://www.futureme.org). The futureme service allows users to compose email messages to themselves for future delivery; emails are held on the server for a certain time (from 1 month up to 30 years in the future) and then sent to the specified recipients at the specified dates. The content of futureme emails is thus explicitly directed towards a future date. In fact, the most common use of this service is for users to address their future selves in the form of a personal memento or conjecture about their future, e.g. ”I am sure your AAAI paper was accepted for publication and you are now happily married.” An analysis of the mood of these emails will thus yield a direct, rather than extrapolated, assessment of public mood toward the future. Futureme currently hosts nearly 500,000 emails to the future. Although a user might send more than one email, it is reasonable to estimate that the total population of the futureme community is in the range of few hundred thousands. Its policy of limitless access and the breadth of attention that ar X iv :0 80 1. 38 64 v1 [ cs .C L ] 2 5 Ja n 20 08 futureme has recently received from the press (Moehringer 2007; Associated Press 2005) has contributed to highly diversify its user community. Nevertheless, the findings presented in this paper are based on a sample of the population that is limited to the community of futureme users, i.e. those that submitted their emails to the futureme website. Although the size and characteristics of the futureme community are unknown due its focus on user anonymity, it clearly abides by some of the key criteria required by the ”wisdom of crowd” phenomena as observed in finance and prediction markets (Surowiecki 2004). In particular, the futureme crowd manifests: a) independence (users are not influenced by other users while composing their emails), b) decentralization (futureme is an online, thus distributed, community), and finally c) diversity of opinion (users write privately to themselves). Instead of relying on user-provided mood annotations and ad hoc models of mood states, we used a popular psychometric instrument to extract mood indicators from the body of futureme emails submitted in 2006, namely an expanded version of the Profile of Mood States questionnaire (POMS) (McNair, Loor, & Droppleman 1971). Mood indicators were thus extracted from 10,741 emails submitted to futureme in 2006 and aggregated to query the “wise” futureme crowd to compute collective sentiment towards forthcoming years. The remainder of the paper is organized as follows. In the next section we discuss the data collection process, the content of the data and some statistical facts about the data employed in the study. In the following section, we introduce the methodology, in particular with respect to the adaptation of the POMS test to extensive email corpora. In the last two sections, we present and discuss our results. Data collection and processing. The data used in the study was obtained from the archives of futureme, a web service that allows its users to send themselves emails to be delivered at a later date, up to 30 years in the future. Upon submission of the email, users can decide whether to make the content of their email publicly available. Subsequently, futureme publishes the following information for public emails without revealing the user’s identity: a) date of email dispatch (when the message was written), b) textual content of email and c) intended date of delivery (when the message is due to arrive). Futureme has experienced significant growth in the 2004 to 2007 period. Whereas in the early years of the service (2002), no more than 500 to 1000 public messages were sent at a yearly basis, this number peaked at about 70,000 in 2006. At the time this analysis was conducted, complete 2007 data was not yet available although extrapolation indicates a continued yet less hyperexponential growth of the service. For this reason, only messages submitted in 2006 were selected for further analysis. All public emails written between January 1 and December 31, 2006 were harvested from the futureme archive and stored in a format suitable for textual analysis. A total of 10,741 emails were collected for the year 2006, with delivery dates ranging from 2006 to 2036. The distributions of emails per destination year as well as the average lag between origin and destination year are shown in Figure 1. 2005 201