Reflection

This lab shows how AI can help organize an R Markdown report and write code for cleaning text, counting words, scoring sentiment, and creating an interactive chart. During the process, my first attempt to knit the report produced a formatting error, which showed that AI-generated instructions still require careful copying and troubleshooting. The example uses simulated posts rather than actual BlueSky data, so its findings cannot be treated as evidence about real customers. I think the lab provides useful practice with AI and R, but a sample dataset and a checklist for verifying results would make the instructions easier to follow.

Human judgment remains necessary when choosing data, evaluating sentiment labels, and interpreting results. This report uses a small, manually defined sentiment dictionary, which makes the calculations easy to inspect but limits the words the analysis can recognize. A word-count method can miss sarcasm, negation, and context. For example, it would score “not good” positively because it recognizes “good” without understanding “not.” The report therefore includes data checks and hand-calculated examples to help evaluate the AI-written code. Any difference between the two simulated weeks reflects how these example posts were constructed, rather than a verified change in customer opinions.

Purpose and Data

This practice report analyzes 20 simulated BlueSky-style posts about a fictional coffee shop across two example weeks.

These posts are invented. They were not collected from BlueSky. This is a demonstration of the workflow. The assignment requests collected group data, so this example should only be submitted if simulated data is allowed.

The analysis asks:

1. Create the Example Dataset

posts <- tibble(
  post_id = 1:20,
  week = rep(c("Week 1", "Week 2"), each = 10),
  text = c(
    "I love this coffee shop. The latte was excellent.",
    "Friendly staff and delicious coffee.",
    "The service was slow and the coffee was cold.",
    "Great atmosphere and a wonderful pastry.",
    "My drink was bitter and disappointing.",
    "Fresh muffins and amazing service.",
    "The tables were dirty and the music was annoying.",
    "Good coffee and a comfortable seat.",
    "My order was wrong. A terrible experience.",
    "Stopped in for a coffee before work.",
    "Excellent coffee and friendly service today.",
    "I love the new menu. The cookies are delicious.",
    "Great service and a wonderful latte.",
    "The coffee was cold and the wait was frustrating.",
    "Amazing pastries and a beautiful space.",
    "Good drinks and helpful staff.",
    "My sandwich was disappointing and expensive.",
    "A pleasant visit with fresh coffee.",
    "Happy with my order. Everything was perfect.",
    "Ordered a latte and sat near the window."
  )
)

knitr::kable(posts)
post_id week text
1 Week 1 I love this coffee shop. The latte was excellent.
2 Week 1 Friendly staff and delicious coffee.
3 Week 1 The service was slow and the coffee was cold.
4 Week 1 Great atmosphere and a wonderful pastry.
5 Week 1 My drink was bitter and disappointing.
6 Week 1 Fresh muffins and amazing service.
7 Week 1 The tables were dirty and the music was annoying.
8 Week 1 Good coffee and a comfortable seat.
9 Week 1 My order was wrong. A terrible experience.
10 Week 1 Stopped in for a coffee before work.
11 Week 2 Excellent coffee and friendly service today.
12 Week 2 I love the new menu. The cookies are delicious.
13 Week 2 Great service and a wonderful latte.
14 Week 2 The coffee was cold and the wait was frustrating.
15 Week 2 Amazing pastries and a beautiful space.
16 Week 2 Good drinks and helpful staff.
17 Week 2 My sandwich was disappointing and expensive.
18 Week 2 A pleasant visit with fresh coffee.
19 Week 2 Happy with my order. Everything was perfect.
20 Week 2 Ordered a latte and sat near the window.

2. Check and Clean the Data

First, check the number of posts, missing values, blank text, duplicate text, and duplicate IDs.

data_checks <- posts %>%
  summarise(
    total_posts = n(),
    missing_text = sum(is.na(text)),
    blank_text = sum(str_trim(text) == "", na.rm = TRUE),
    duplicate_text = sum(duplicated(text)),
    duplicate_ids = sum(duplicated(post_id))
  )

knitr::kable(data_checks)
total_posts missing_text blank_text duplicate_text duplicate_ids
20 0 0 0 0
knitr::kable(count(posts, week))
week n
Week 1 10
Week 2 10

The expected result is 20 posts, 10 per week, with no missing text, blank text, duplicate text, or duplicate IDs.

The cleaning code removes missing or blank posts, standardizes spaces, and removes repeated text within each week.

clean_posts <- posts %>%
  filter(!is.na(text)) %>%
  mutate(text = str_squish(text)) %>%
  filter(text != "") %>%
  distinct(week, text, .keep_all = TRUE)

stopifnot(
  nrow(clean_posts) == 20L,
  !anyDuplicated(clean_posts$post_id),
  all(count(clean_posts, week)$n == 10L)
)

knitr::kable(
  tibble(
    posts_before_cleaning = nrow(posts),
    posts_after_cleaning = nrow(clean_posts)
  )
)
posts_before_cleaning posts_after_cleaning
20 20

No rows should be removed because this example dataset was constructed without those data problems.

3. Text Analysis

Tokenization splits each post into lowercase words. For word-frequency analysis, common stop words such as “the” and “and” are removed.

tokens <- clean_posts %>%
  unnest_tokens(word, text)

frequency_tokens <- tokens %>%
  anti_join(tidytext::stop_words, by = "word") %>%
  filter(str_detect(word, "[a-z]"))

top_words <- frequency_tokens %>%
  count(word, sort = TRUE) %>%
  arrange(desc(n), word) %>%
  slice_head(n = 10)

knitr::kable(top_words)
word n
coffee 8
service 4
latte 3
amazing 2
cold 2
delicious 2
disappointing 2
excellent 2
fresh 2
friendly 2
ggplot(
  top_words,
  aes(x = reorder(word, n), y = n)
) +
  geom_col(fill = "#8B6551") +
  coord_flip() +
  labs(
    title = "Most Frequent Words in the Simulated Posts",
    x = NULL,
    y = "Word occurrences"
  ) +
  theme_minimal()

These counts show repeated topics in the example. A frequent word does not necessarily indicate positive or negative sentiment. Related words such as “pastry” and “pastries” are counted separately.

4. Define a Transparent Sentiment Dictionary

This demonstration uses a small, manually defined dictionary rather than a standard validated sentiment lexicon. It requires no additional dictionary download.

Each matched positive word contributes +1 and each matched negative word contributes -1. Words outside the dictionary contribute zero.

positive_words <- c(
  "love", "excellent", "friendly", "delicious",
  "great", "wonderful", "fresh", "amazing",
  "good", "comfortable", "beautiful", "helpful",
  "pleasant", "happy", "perfect"
)

negative_words <- c(
  "slow", "cold", "bitter", "disappointing",
  "dirty", "annoying", "wrong", "terrible",
  "frustrating", "expensive"
)

sentiment_dictionary <- tibble(
  word = c(positive_words, negative_words),
  sentiment = c(
    rep("positive", length(positive_words)),
    rep("negative", length(negative_words))
  ),
  value = c(
    rep(1, length(positive_words)),
    rep(-1, length(negative_words))
  )
)

stopifnot(!anyDuplicated(sentiment_dictionary$word))

knitr::kable(sentiment_dictionary)
word sentiment value
love positive 1
excellent positive 1
friendly positive 1
delicious positive 1
great positive 1
wonderful positive 1
fresh positive 1
amazing positive 1
good positive 1
comfortable positive 1
beautiful positive 1
helpful positive 1
pleasant positive 1
happy positive 1
perfect positive 1
slow negative -1
cold negative -1
bitter negative -1
disappointing negative -1
dirty negative -1
annoying negative -1
wrong negative -1
terrible negative -1
frustrating negative -1
expensive negative -1

The labels reflect choices that a human analyst should review. For example, “cold” is negative here because it describes an unwanted cold coffee, but it could be appropriate in a post about iced coffee.

5. Score Each Post

Sentiment scoring uses all word tokens, before stop-word removal. Keeping all tokens makes it easier to inspect context later, although this simple scoring rule still does not interpret that context.

Posts with no dictionary matches receive a score of zero.

matched_words <- tokens %>%
  inner_join(sentiment_dictionary, by = "word")

scores <- matched_words %>%
  group_by(post_id) %>%
  summarise(
    positive_matches = sum(value == 1),
    negative_matches = sum(value == -1),
    net_sentiment = sum(value),
    .groups = "drop"
  )

post_scores <- clean_posts %>%
  left_join(scores, by = "post_id") %>%
  mutate(
    across(
      c(positive_matches, negative_matches, net_sentiment),
      ~ replace_na(.x, 0)
    ),
    classification = case_when(
      net_sentiment > 0 ~ "Positive score",
      net_sentiment < 0 ~ "Negative score",
      TRUE ~ "Zero score"
    )
  )

knitr::kable(post_scores)
post_id week text positive_matches negative_matches net_sentiment classification
1 Week 1 I love this coffee shop. The latte was excellent. 2 0 2 Positive score
2 Week 1 Friendly staff and delicious coffee. 2 0 2 Positive score
3 Week 1 The service was slow and the coffee was cold. 0 2 -2 Negative score
4 Week 1 Great atmosphere and a wonderful pastry. 2 0 2 Positive score
5 Week 1 My drink was bitter and disappointing. 0 2 -2 Negative score
6 Week 1 Fresh muffins and amazing service. 2 0 2 Positive score
7 Week 1 The tables were dirty and the music was annoying. 0 2 -2 Negative score
8 Week 1 Good coffee and a comfortable seat. 2 0 2 Positive score
9 Week 1 My order was wrong. A terrible experience. 0 2 -2 Negative score
10 Week 1 Stopped in for a coffee before work. 0 0 0 Zero score
11 Week 2 Excellent coffee and friendly service today. 2 0 2 Positive score
12 Week 2 I love the new menu. The cookies are delicious. 2 0 2 Positive score
13 Week 2 Great service and a wonderful latte. 2 0 2 Positive score
14 Week 2 The coffee was cold and the wait was frustrating. 0 2 -2 Negative score
15 Week 2 Amazing pastries and a beautiful space. 2 0 2 Positive score
16 Week 2 Good drinks and helpful staff. 2 0 2 Positive score
17 Week 2 My sandwich was disappointing and expensive. 0 2 -2 Negative score
18 Week 2 A pleasant visit with fresh coffee. 2 0 2 Positive score
19 Week 2 Happy with my order. Everything was perfect. 2 0 2 Positive score
20 Week 2 Ordered a latte and sat near the window. 0 0 0 Zero score

A zero score can mean no dictionary matches or an equal number of positive and negative matches. It does not prove that a post is neutral.

6. Verify the Results

Hand calculation

Post 1 contains “love” and “excellent,” so its expected score is +2.

Post 3 contains “slow” and “cold,” so its expected score is -2.

Post 10 contains no dictionary words, so its expected score is 0.

manual_checks <- tibble(
  post_id = c(1L, 3L, 10L),
  expected_score = c(2, -2, 0)
) %>%
  left_join(
    post_scores %>% select(post_id, net_sentiment),
    by = "post_id"
  ) %>%
  mutate(passed = expected_score == net_sentiment)

knitr::kable(manual_checks)
post_id expected_score net_sentiment passed
1 2 2 TRUE
3 -2 -2 TRUE
10 0 0 TRUE
stopifnot(all(manual_checks$passed))

Inspect the matched words

matched_words %>%
  filter(post_id %in% c(1, 3)) %>%
  select(post_id, word, sentiment, value) %>%
  knitr::kable()
post_id word sentiment value
1 love positive 1
1 excellent positive 1
3 slow negative -1
3 cold negative -1

Check totals

stopifnot(
  nrow(post_scores) == nrow(clean_posts),
  sum(post_scores$positive_matches) ==
    sum(matched_words$value == 1),
  sum(post_scores$negative_matches) ==
    sum(matched_words$value == -1),
  sum(post_scores$net_sentiment) ==
    sum(matched_words$value)
)

cat("All programmed verification checks passed.")
## All programmed verification checks passed.

These checks verify selected calculations and accounting totals. They do not establish that the dictionary correctly understands every post.

7. Compare the Example Weeks

weekly_summary <- post_scores %>%
  group_by(week) %>%
  summarise(
    posts = n(),
    positive_matches = sum(positive_matches),
    negative_matches = sum(negative_matches),
    net_sentiment = sum(net_sentiment),
    average_score = mean(net_sentiment),
    .groups = "drop"
  )

knitr::kable(weekly_summary, digits = 2)
week posts positive_matches negative_matches net_sentiment average_score
Week 1 10 10 8 2 2
Week 2 10 14 4 10 10

Using this dictionary, Week 1 has a net score of 2 and Week 2 has a net score of 10.

Each week contains 10 posts. The average score per post allows a clearer comparison if the number of posts changes in a future dataset. It still depends on post length and dictionary coverage.

8. Interactive Plotly Chart

Hover over a bar to see the average score, number of posts, and matched-word counts.

chart_data <- weekly_summary %>%
  mutate(
    hover_text = paste0(
      week,
      "<br>Posts: ", posts,
      "<br>Average score: ", round(average_score, 2),
      "<br>Positive matches: ", positive_matches,
      "<br>Negative matches: ", negative_matches
    )
  )

plot_ly(
  data = chart_data,
  x = ~week,
  y = ~average_score,
  type = "bar",
  text = ~hover_text,
  hoverinfo = "text",
  marker = list(color = c("#A88770", "#58786B"))
) %>%
  layout(
    title = "Sentiment in Simulated Coffee Shop Posts",
    xaxis = list(title = "Example week"),
    yaxis = list(
      title = "Average net sentiment score per post",
      zeroline = TRUE
    )
  )

9. Critique the AI-Assisted Analysis

The code provides a reproducible demonstration, but the interpretation requires human review.

For real BlueSky data, I would review a sample of posts manually, compare those labels with the automated scores, investigate disagreements, and document how the posts were collected. I would also compare this custom dictionary with an established sentiment method.

10. AI Prompt Record

The following requests were used during this conversation.

Initial request:

i dont have anything. come up wiht an example and walk me thru step by step

AI response summary: The assistant proposed clearly labeled, simulated coffee-shop posts and began guiding the creation of an HTML R Markdown report.

Follow-up request:

just give me something i can press cntrl a and delete it all and copy everything else into the top lef t

AI response summary: The assistant supplied a complete replacement R Markdown document, including data checks, text analysis, a custom sentiment dictionary, manual verification, and an interactive plotly chart.

These requests helped obtain an example and a complete document, but they did not precisely specify packages or verification requirements. A stronger follow-up would explicitly request the packages, input structure, expected output, and checks. Screenshots of the actual conversation should accompany the discussion submission.