A Frosty Deal?

Reading news articles on the will-they-won’t-they post-Brexit trade negotiations with the EU sees days of optimism jarred by days of gloom. Do negative news articles, when one wants a positive outcome, leave a deeper impression?

I wondered if I could get a more objective view from quantitative analysis of textual data. To do this, I’m going to look at hundreds of articles published in the Guardian newspaper over the course of the year to see how trade-talk sentiment has changed week-to-week.

library(tidyverse)
library(rebus)
library(wesanderson)
library(kableExtra)
library(lubridate)
library(GuardianR)
library(quanteda)
library(scales)
library(tictoc)
library(patchwork)
library(text2vec)
library(topicmodels)
library(zoo)
library(glue)

theme_set(theme_bw())

cols <- wes_palette(name = "Chevalier1")

The Withdrawal Agreement between the UK and the European Union was signed on the 24th of January 2020. I’ll import Brexit-related newspaper articles from that date.

The Guardian newspaper asks for requests to span no more than 1 month at a time. So I’ll first create a set of monthly date ranges.

dates_df <- tibble(start_date = seq(ymd("2020-01-24"), today(), by = "1 month")) %>%
  mutate(end_date = start_date + months(1) - 1)

dates_df %>%
  kable()

I’ll import the newspaper articles in monthly chunks. Note, access to the Guardian’s API requires a key which may be requested here.

tic()

article_df <-
  dates_df %>%
  pmap_dfr(., function(start_date, end_date) {
    Sys.sleep(1)
    get_guardian(
      "brexit",
      from.date = start_date,
      to.date = end_date,
      api.key = key
    )
  })

toc()

The data need a little cleaning, for example, to remove multi-topic articles, html tags and non-breaking spaces.

trade_df <-
  article_df %>%
  filter(!str_detect(id, "/live/"), sectionId %in% c("world", "politics", "business")) %>% 
  mutate(
    body = str_remove_all(body, "<.*?>") %>% str_to_lower(),
    body = str_remove_all(body, "[^a-z0-9 .-]"),
    body = str_remove_all(body, "nbsp")
  )

A corpus then gives me a collection of texts whereby each document is a newspaper article.

trade_corp <- trade_df %>% 
  corpus(docid_field = "shortUrl", text_field = "body")

Although I’ve only imported articles mentioning Brexit since the Withdrawal Agreement was signed, some of these articles will not be related to trade negotiations with the EU. For example, there are on-going negotiations with many countries around the world. So, I’m going to use word embeddings to help narrow the focus to the specific context of the UK-EU trade deal.

The chief negotiator for the EU is Michel Barnier, so I’ll quantitatively identify words in close proximity to “Barnier” in the context of these Brexit news articles.

window <- 5

trade_fcm <- 
  trade_corp %>% 
  fcm(context = "window", window = window, count = "weighted", weights = window:1)

glove <- GlobalVectors$new(rank = 60, x_max = 10)

set.seed(42)

wv_main <- glove$fit_transform(trade_fcm, n_iter = 10)

## INFO  [18:12:41.398] epoch 1, loss 0.3801 
## INFO  [18:12:43.295] epoch 2, loss 0.2511 
## INFO  [18:12:45.176] epoch 3, loss 0.2226 
## INFO  [18:12:47.151] epoch 4, loss 0.2020 
## INFO  [18:12:49.027] epoch 5, loss 0.1846 
## INFO  [18:12:50.908] epoch 6, loss 0.1710 
## INFO  [18:12:52.791] epoch 7, loss 0.1607 
## INFO  [18:12:54.655] epoch 8, loss 0.1527 
## INFO  [18:12:56.530] epoch 9, loss 0.1462 
## INFO  [18:12:58.403] epoch 10, loss 0.1407

wv_context <- glove$components
word_vectors <- wv_main + t(wv_context)

search_coord <- 
  word_vectors["barnier", , drop = FALSE]

word_vectors %>% 
  sim2(search_coord, method = "cosine") %>% 
  as_tibble(rownames = NA) %>% 
  rownames_to_column("term") %>% 
  rename(similarity = 2) %>% 
  arrange(desc(similarity)) %>% 
  slice(1:10) %>%
  kable()

Word embedding is a learned modelling technique placing words into a multi-dimensional vector space such that contextually-similar words may be found close by. Not surprisingly, one of the closest words contextually is “Michel”. And as he is the chief negotiator for the EU, we find “eu’s”, “chief”, and “negotiator” also in the top most contextually-similar words.

The word embeddings algorithm, through word co-occurrence, has identified the name of Michel Barnier’s UK counterpart David Frost. So filtering articles for “Barnier”, “Frost” and “UK-EU” should help narrow the focus.

context_df <- 
  trade_df %>% 
  filter(str_detect(body, "barnier|frost|uk-eu")) 

context_corp <- 
  context_df %>% 
  corpus(docid_field = "shortUrl", text_field = "body")

I can then use quanteda’s kwic function to review the key phrases in context to ensure I’m homing in on the texts I want. Short URLs are included below so I can click on any to read the actual article as presented by The Guardian.

set.seed(123)

context_corp %>%
  tokens(
    remove_punct = TRUE,
    remove_symbols = TRUE,
    remove_numbers = TRUE
  ) %>%
  kwic(pattern = phrase(c("trade negotiation", "trade deal", "trade talks")), 
       valuetype = "regex", window = 7) %>%
  as_tibble() %>%
  left_join(article_df, by = c("docname" = "shortUrl")) %>% 
  slice_sample(n = 10) %>% 
  select(docname, pre, keyword, post, headline) %>%
  kable()

Quanteda provides a sentiment dictionary which, in addition to identifying positive and negative words, also finds negative-negatives and negative-positives such as, for example, “not effective”. For each week’s worth of articles, I’ll calculate the proportion of positive sentiments.

tic()

sent_df <- 
  context_corp %>% 
  dfm(dictionary = data_dictionary_LSD2015) %>% 
  as_tibble() %>%
  left_join(context_df, by = c("doc_id" = "shortUrl")) %>% 
  mutate(
    pos = positive + neg_negative,
    neg = negative + neg_positive,
    date = ceiling_date(as_date(webPublicationDate), "week"),
    pct_pos = pos / (pos + neg)
  )

sent_df %>% 
  select(doc_id, starts_with("pos"), starts_with("neg")) %>% 
  slice(1:10) %>% 
  kable()

summary_df <- sent_df %>% 
  group_by(date) %>% 
  summarise(pct_pos = mean(pct_pos), n = n())

toc()

## 0.686 sec elapsed

Plotting the changing proportion of positive sentiment over time did surprise me a little. The outcome was more balanced than I expected which perhaps confirms the deeper impression left on me by negative articles.

The upper violin plot shows the average weight of the sentiment across multiple articles for each week. Individually the articles range from 20% to 80% positive, with discernible periods of relatively negative and relatively positive sentiment.

The lower plot shows the volume of articles. As we draw closer to the crunch-point the volume appears to be picking up.

p1 <- sent_df %>% 
  ggplot(aes(date, pct_pos)) +
  geom_violin(aes(group = date), alpha = 0.5, fill = cols[1]) +
  geom_line(data = summary_df, aes(date, pct_pos), colour = cols[1], linetype = "dashed") +
  geom_hline(yintercept = 0.5, linetype = "dashed", colour = cols[4], size = 1) +
  scale_y_continuous(labels = percent_format(), limits = c(0.2, 0.8)) +
  labs(title = "Changing Sentiment Towards a UK-EU Trade Deal",
       subtitle = "Week-to-week Since the Withdrawal Agreement",
       x = NULL, y = "Positive Sentiment")

p2 <- summary_df %>% 
  ggplot(aes(date, n)) +
  geom_line(colour = cols[1]) +
  labs(x = "Weeks", y = "Article Count",
       caption = "Source: Guardian Newspaper")

p1 / p2 + 
  plot_layout(heights = c(2, 1))

Adopting titaniumtroop’s Disqus approach we could visualise a rolling 7-day mean with a narrowing ribbon representing a narrowing variation in sentiment.

width <- 7

sent_df2 <- sent_df %>%
  mutate(web_date = as_date(webPublicationDate)) %>% 
  group_by(web_date) %>%
  summarise(pct_pos = sum(pos) / sum(neg + pos)) %>% 
  mutate(
    roll_mean = rollapply(pct_pos, width, mean, partial = TRUE, align = "right"),
    roll_lq = rollapply(pct_pos, width, quantile, probs = 0.25, partial = TRUE, align = "right"),
    roll_uq = rollapply(pct_pos, width, quantile, probs = 0.75, partial = TRUE, align = "right")
  )

sent_df2 %>%
  ggplot(aes(web_date)) +
  geom_line(aes(y = roll_mean), colour = cols[1]) +
  geom_ribbon(aes(ymin = roll_lq, ymax = roll_uq), alpha = 0.33, fill = cols[1]) +
  geom_hline(yintercept = 0.5, linetype = "dashed", colour = cols[4], size = 1) +
  scale_y_continuous(labels = percent_format(accuracy = 1)) +
  labs(
    title = "Changing Sentiment Towards a UK-EU Trade Deal",
    subtitle = glue("Rolling {width} days Since the Withdrawal Agreement"),
    x = NULL, y = "Positive Sentiment"
  )

Some writers exhibit more sentiment variation than others.

byline_df <- 
  sent_df %>% 
  mutate(byline = word(byline, 1, 2) %>% str_remove_all(PUNCT)) %>% 
  group_by(byline, date) %>% 
  summarise(pct_pos = mean(pct_pos), n = n())

top_3 <- byline_df %>% 
  count(byline, sort = TRUE) %>% 
  ungroup() %>% 
  filter(!is.na(byline)) %>% 
  slice(c(3, 2)) %>% 
  pull(byline)

byline_df %>% 
  filter(byline %in% top_3) %>% 
  ggplot(aes(date, pct_pos, colour = byline)) +
  geom_line() +
  geom_hline(yintercept = 0.5, linetype = "dotted", colour = cols[2]) +
  scale_y_continuous(labels = percent_format(), limits = c(0.2, 0.8)) +
  scale_colour_manual(values = cols[c(1, 4)]) +
  labs(title = "Changing Sentiment Towards a UK-EU Trade Deal",
       subtitle = "Week-to-week Since the Withdrawal Agreement",
       x = "Weeks", y = "Positive Sentiment", colour = "Byline", 
       caption = "Source: Guardian Newspaper")