sentiment_breakdown: Breakdown the sentiment into topical components

Description

Break down the sentiment series obtained with sentiment_series() into topical components. Sentiment is broken down at the document level using estimated topic proportions, then processed to create a time series and its components.

Usage

sentiment_breakdown(
  x,
  period = c("year", "quarter", "month", "day", "identity"),
  rolling_window = 1,
  scale = TRUE,
  scaling_period = c("1900-01-01", "2099-12-31"),
  plot = c(FALSE, TRUE, "silent"),
  as.xts = TRUE,
  ...
)
plot_sentiment_breakdown(
  x,
  period = c("year", "quarter", "month", "day"),
  rolling_window = 1,
  scale = TRUE,
  scaling_period = c("1900-01-01", "2099-12-31"),
  ...
)

Value

A time series of sentiment, stored as an xts::xts object or as a data.frame.

Arguments

x: a LDA() or rJST() model populated with internal dates and/or internal sentiment.
period: the sampling period within which the sentiment of documents will be averaged. period = "identity" is a special case that will return document-level variables before the aggregation happens. Useful to rapidly compute topical sentiment at the document level.
rolling_window: if greater than 1, determines the rolling window to compute a moving average of sentiment. The rolling window is based on the period unit and rely on actual dates (i.e, is not affected by unequally spaced data points).
scale: if TRUE, the resulting time series will be scaled to a mean of zero and a standard deviation of 1. This argument also has the side effect of attaching scaled sentiment values as docvars to the input object with the _scaled suffix.
scaling_period: the date range over which the scaling should be applied. Particularly useful to normalize only the beginning of the time series.
plot: if TRUE, prints a plot of the time series and attaches it as an attribute to the returned object. If 'silent', do not print the plot but still attaches it as an attribute.
as.xts: if TRUE, returns an xts::xts object. Otherwise, returns a data.frame.
...: other arguments passed on to zoo::rollapply() or mean() and sd().

Details

The sentiment is broken down at the sentiment level assuming the following composition: $$s = \sum^K_{i=1} s_i \times \theta_i$$, where $s_i$ is the sentiment of topic i and $theta_i$ the proportion of topic i in a given document. For an LDA model, the sentiment of each topic is considered equal to the document sentiment (i.e. $s_i = s \forall i \in K$). The topical sentiment attention, defined by $s*_i = s_i \times \theta_i$ represent the effective sentiment conveyed by a topic in a document. The topical sentiment attention of all documents in a period are averaged to compute the breakdown of the sentiment time series.

Examples

Run this code

lda <- LDA(ECB_press_conferences_tokens)
lda <- fit(lda, 100)
scores <- compute_PicaultRenault_scores(ECB_press_conferences)
sentopics_sentiment(lda) <- scores[names(lda$tokens), "EC"]
sentiment_breakdown(lda)

# plot shortcut
plot_sentiment_breakdown(lda)

# also available for rJST models (with topic-level sentiment)
rjst <- rJST(ECB_press_conferences_tokens, lexicon = LoughranMcDonald)
rjst <- fit(rjst, 100)
sentopics_sentiment(rjst) # compute sentiment
plot_sentiment_breakdown(rjst)