Figure 8

Normalized sentiment time series on Twitter using \(\pmb{\Delta_{h}}\) of 1.0 for all dictionaries. To normalize the sentiment score, we subtract the mean and divide by the absolute range. The resolution is 1 day, and draws on the 10% gardenhose sample of public Tweets provided by Twitter. All of the dictionaries exhibit wide variation for very early Tweets, and from 2012 onward generally track together strongly with the exception of MPQA and LIWC. The LIWC and MPQA dictionaries show opposite trends: a rise until 2012 with a decline after 2012 from LIWC, and a decline before 2012 with a rise afterwards from MPQA. To analyze the trends we look at the words driving the movement across years using word shift Figures in S7 Appendix. An interactive version of this Figure using the labMT dictionary be found at http://hedonometer.org.