Sentiment with rzeit2
In rzeit2: Client for the ZEIT ONLINE Content API

Load required packages

library(rzeit2)
library(robotstxt)
library(stringr)
library(dplyr)
library(tidytext)
library(ggplot2)
library(ggthemes)

Load meta data

We begin by downloading the meta data using get_content_all. We search for articles between 01st May and 31st May 2018. The dates are chosen randomly. We are looking for articles containing the keyword Merkel.

# load meta data
articles_merkel <- get_content_all("Merkel",
                                   begin_date = "20180501",
                                   end_date = "20180531")

# extract urls
urls <- articles_merkel$content$href

Scrape websites

First, let's check whether we are allowed to scrape ZEIT ONLINE. Every website usually contains a so-called robots.txt file which tells you, which automated processes are allowed. robotstxt is a very useful package which provides you all information right away for a given url.

robots <- robotstxt("https://zeit.de")
robots$permissions

The permissions table shows that we are allowed to scrape usual news articles. The Disallow indicates the subdirectories you should not index and scrape.

Second, we start downloading and parsing the articles from ZEIT ONLINE using get_article_text. The function works on charcters and character vectors. Please use a decend timeout to be gently. Since, we found 120 articles this may take some time. Some articles are premium content, those are indicated by the string "[PAYWALL] ZEIT PLUS CONTENT".

# get article content
article_texts <- get_article_text(urls, timeout = 2)

Calculate sentiment

We use the Sentiment Wortschatz, a sentiment dictionary for general German, to calculate the sentiment for each article. The dictionary contains 3,458 words and a respective positive or negative polarity. The dictionary is included in this package and can be loaded using load. We prepare a data frame including the url, date and scraped text of each article. Articles beyond the paywall are excluded, because we cannot access them using get_article_text. Afterwards, we calculate the score for each word and sum up all score for each article.

# prepare data frame
articles <- data.frame(url = urls,
                       text = article_texts,
                       date = as.Date(articles_merkel$content$release_date),
                       stringsAsFactors = F)

# exclude premium content
articles <- articles %>% 
  filter(!str_detect(text, "ZEIT PLUS CONTENT"))

# lazy loading german sentiment dictionary
data("senti_ws")

# calculate the sentiment for each day
sentiment_example <- articles %>%
  unnest_tokens(word, text) %>% 
  inner_join(senti_ws) %>% 
  group_by(url, date) %>%
  summarise(score = sum(score))

Due to copyright issues I am not allowed to include the article texts in this package. Hence, I saved the corresponding data frame from above in sentiment_example. The data set can also be loaded using load. We are grouping the data by date and sum up all score for each day. Subsequently, we divide the score by the number of articles per day to obtain a relative score.

data("sentiment_example")

# calculate sentiment by day
sentiment <- sentiment_example %>% 
  group_by(date) %>% 
  summarise(sentiment = sum(score) / n())

# plot the sentiment by day
ggplot(sentiment, aes(date, sentiment)) +
  geom_point(pch = 1, col = "#3a9b96") +
  geom_line(col = "#3a9b96", alpha = .6) +
  geom_hline(yintercept = mean(sentiment$sentiment), col = "#4e4e4e") +
  annotate("text", x = as.Date("2018-05-30"), y = -.18, 
           label = sprintf("mean = %s", round(mean(sentiment$sentiment), 2)), col = "#4e4e4e") +
  theme_fivethirtyeight() +
  theme(axis.title = element_text(), axis.title.x = element_blank()) + ylab('Articles Sentiment')

The graph shows that the sentiment differs strongly over time. Although, the mean equals nearly 0. A short look at the headings of the articles tells us that the articles are not about Merkel herself, but rather include articles about general German politics. Hence, the sentiment cannot be interpreted in advantage or disadvantage of Merkel. Using sentiment scores like this carries two risks. First, the articles probably do not only include content on the object of interest. Consequently, we probably measure paragraphs which are not of interest. Second, the dictionary does not match the content which may lead to incorrect scores. In this case it would probably be useful to use a dictionary concentrated on German politics. Besides these thoughts, sentiment scores probably provide a useful insights of the mood towards the object of interest.