Reflection draft: revise after running the code and completing the manual review.
My earlier RPubs report helped me understand how tokenization and stop-word removal turn short reviews into a word-frequency table. Extending that work with AI provides a way to explore sentiment and interactive charts, but I need to verify the code and its interpretations. In this report, I will check row counts, compare the word counts with the original sentences, and inspect which words the sentiment dictionary actually matches. A frequent word identifies a repeated topic; it does not establish whether people like that topic or why they mention it.
Human judgment contributes the context that a word dictionary cannot reliably capture. A review can praise product quality while criticizing shipping, so a single positive label can hide an important complaint. I will compare the automated results with my own reading of all five reviews and explain any disagreement. These sentences are classroom examples rather than a representative customer sample, so their percentages cannot establish broader customer attitudes. AI can help produce code and charts, but I remain responsible for interpreting the evidence, documenting limitations, and supporting any recommendation.
This report extends my published Text analysis–Week 13 report. The input consists of the same five example product-review sentences shown there. They are classroom examples rather than newly collected customer data.
The questions are: Which words recur, what sentiment words are detected, and where might a dictionary score fail to reflect the complete meaning of a review?
reviews_raw <- tibble(
review_id = seq_len(5),
text = c(
"I love the quality of this product. The price is reasonable, and delivery was fast.",
"The product quality is excellent, but the shipping took too long.",
"Great customer service! I would definitely recommend this product to my friends.",
"The price is too expensive. However, the product looks beautiful.",
"Fast delivery and excellent service. I am very satisfied with my purchase."
)
)
kable(reviews_raw, caption = "The original five classroom examples")
| review_id | text |
|---|---|
| 1 | I love the quality of this product. The price is reasonable, and delivery was fast. |
| 2 | The product quality is excellent, but the shipping took too long. |
| 3 | Great customer service! I would definitely recommend this product to my friends. |
| 4 | The price is too expensive. However, the product looks beautiful. |
| 5 | Fast delivery and excellent service. I am very satisfied with my purchase. |
Keep the original text. Remove blank records, trim repeated whitespace, and check for exact duplicate review text. In a real collection, identical text from different people may be meaningful; the appropriate duplicate rule depends on the source. For this small classroom exercise, exact repeated text would be retained once.
reviews_nonblank <- reviews_raw %>%
mutate(text_clean = str_squish(text)) %>%
filter(!is.na(text_clean), text_clean != "")
reviews <- reviews_nonblank %>% distinct(text_clean, .keep_all = TRUE)
cleaning_audit <- tibble(
check = c("Original rows", "Blank rows removed", "Exact duplicates removed", "Final rows"),
value = c(nrow(reviews_raw), nrow(reviews_raw) - nrow(reviews_nonblank),
nrow(reviews_nonblank) - nrow(reviews), nrow(reviews))
)
kable(cleaning_audit)
| check | value |
|---|---|
| Original rows | 5 |
| Blank rows removed | 0 |
| Exact duplicates removed | 0 |
| Final rows | 5 |
stopifnot(nrow(reviews) == 5L, !anyDuplicated(reviews$review_id),
all(nchar(reviews$text_clean) > 0))
Tokenize to one word per row. Use stop-word removal for the frequency analysis. Keep a separate token table for sentiment so that this frequency filter does not silently decide which words are available for sentiment matching.
tokens_all <- reviews %>%
select(review_id, text_clean) %>%
unnest_tokens(word, text_clean)
tokens_content <- tokens_all %>%
anti_join(tidytext::stop_words, by = "word") %>%
filter(str_detect(word, "[a-z]"))
word_counts <- tokens_content %>% count(word, sort = TRUE)
kable(tibble(measure = c("All word tokens", "Tokens after frequency filters", "Distinct retained words"),
value = c(nrow(tokens_all), nrow(tokens_content), nrow(word_counts))))
| measure | value |
|---|---|
| All word tokens | 60 |
| Tokens after frequency filters | 26 |
| Distinct retained words | 17 |
kable(word_counts, caption = "Word occurrences, not number of customers")
| word | n |
|---|---|
| product | 4 |
| delivery | 2 |
| excellent | 2 |
| fast | 2 |
| price | 2 |
| quality | 2 |
| service | 2 |
| beautiful | 1 |
| customer | 1 |
| expensive | 1 |
| friends | 1 |
| love | 1 |
| purchase | 1 |
| reasonable | 1 |
| recommend | 1 |
| satisfied | 1 |
| shipping | 1 |
word_chart_data <- word_counts %>%
slice_max(n, n = 15, with_ties = FALSE) %>%
arrange(n, word) %>%
mutate(word = factor(word, levels = word))
plot_ly(word_chart_data, x = ~n, y = ~word, type = "bar", orientation = "h",
marker = list(color = "#2563A6"),
hovertemplate = "%{y}<br>Occurrences: %{x}<extra></extra>") %>%
layout(title = "Most frequent retained words in five example reviews",
xaxis = list(title = "Word occurrences", dtick = 1),
yaxis = list(title = ""), margin = list(l = 110))
Hover over the bars to inspect counts. The chart summarizes this small set only.
The Bing dictionary assigns words positive or negative labels. For each review, score = positive word occurrences minus negative word occurrences. A positive score is labeled Positive and a negative score Negative. A zero score with matches is called Balanced; a review with no matches is called No matches. Neither label proves that a human would regard a review as neutral. This method does not model negation, sarcasm, or the target of an opinion.
bing <- tidytext::get_sentiments("bing")
sentiment_matches <- tokens_all %>% inner_join(bing, by = "word")
match_counts <- sentiment_matches %>%
group_by(review_id) %>%
summarise(positive_words = sum(sentiment == "positive"),
negative_words = sum(sentiment == "negative"), .groups = "drop")
scored <- reviews %>%
left_join(match_counts, by = "review_id") %>%
mutate(across(c(positive_words, negative_words), ~replace_na(.x, 0L)),
matched_words = positive_words + negative_words,
score = positive_words - negative_words,
automated_label = case_when(
matched_words == 0 ~ "No matches",
score > 0 ~ "Positive",
score < 0 ~ "Negative",
TRUE ~ "Balanced"
))
sentiment_summary <- scored %>% count(automated_label, name = "reviews") %>%
mutate(percent = 100 * reviews / sum(reviews))
kable(scored %>% select(review_id, positive_words, negative_words, score, automated_label))
| review_id | positive_words | negative_words | score | automated_label |
|---|---|---|---|---|
| 1 | 3 | 0 | 3 | Positive |
| 2 | 1 | 0 | 1 | Positive |
| 3 | 2 | 0 | 2 | Positive |
| 4 | 1 | 1 | 0 | Balanced |
| 5 | 3 | 0 | 3 | Positive |
kable(sentiment_summary, digits = 1, caption = "Shares of the five examples under the stated scoring rule")
| automated_label | reviews | percent |
|---|---|---|
| Balanced | 1 | 20 |
| Positive | 4 | 80 |
kable(sentiment_matches %>% count(review_id, word, sentiment),
caption = "Inspect the actual word matches before interpreting the scores")
| review_id | word | sentiment | n |
|---|---|---|---|
| 1 | fast | positive | 1 |
| 1 | love | positive | 1 |
| 1 | reasonable | positive | 1 |
| 2 | excellent | positive | 1 |
| 3 | great | positive | 1 |
| 3 | recommend | positive | 1 |
| 4 | beautiful | positive | 1 |
| 4 | expensive | negative | 1 |
| 5 | excellent | positive | 1 |
| 5 | fast | positive | 1 |
| 5 | satisfied | positive | 1 |
plot_ly(sentiment_summary, x = ~automated_label, y = ~reviews, type = "bar",
text = ~paste0(round(percent, 1), "%"), textposition = "auto",
marker = list(color = "#24887C"),
hovertemplate = "%{x}<br>Reviews: %{y}<br>Share: %{text}<extra></extra>") %>%
layout(title = "Dictionary labels for five example reviews",
xaxis = list(title = "Automated label"),
yaxis = list(title = "Review count", dtick = 1))
These checks compare independently defined quantities. They do not prove that sentiment labels match human interpretation.
checks <- tibble(
check = c("Cleaning accounting balances", "Five reviews retained",
"One sentiment result per review", "Sentiment counts add to five",
"Percentages add to 100", "Dictionary matches accounted for"),
passed = c(
nrow(reviews_raw) == cleaning_audit$value[2] + cleaning_audit$value[3] + nrow(reviews),
nrow(reviews) == 5,
nrow(scored) == nrow(reviews) && !anyDuplicated(scored$review_id),
sum(sentiment_summary$reviews) == nrow(reviews),
abs(sum(sentiment_summary$percent) - 100) < 1e-8,
sum(scored$matched_words) == nrow(sentiment_matches)
)
)
kable(checks)
| check | passed |
|---|---|
| Cleaning accounting balances | TRUE |
| Five reviews retained | TRUE |
| One sentiment result per review | TRUE |
| Sentiment counts add to five | TRUE |
| Percentages add to 100 | TRUE |
| Dictionary matches accounted for | TRUE |
stopifnot(all(checks$passed))
Reading the original sentences shows that the word product appears once in each of reviews 1, 2, 3, and 4, and does not appear in review 5. The hand count is therefore 4. Compare that with the code result below.
product_count <- sum(tokens_all$word == "product")
kable(tibble(word = "product", hand_count = 4L, code_count = product_count,
agrees = product_count == 4L))
| word | hand_count | code_count | agrees |
|---|---|---|---|
| product | 4 | 4 | TRUE |
stopifnot(product_count == 4L)
Read each original sentence yourself. Replace the five empty labels below with Positive, Negative, Mixed, or Neutral; enter a short reason for each. These labels are intentionally blank so this report does not claim a manual review you have not performed. Mixed is useful when a review contains both praise and criticism.
manual_review <- tibble(
review_id = 1:5,
human_label = c("", "", "", "", ""),
human_reason = c("", "", "", "", "")
)
valid_labels <- c("", "Positive", "Negative", "Mixed", "Neutral")
stopifnot(all(manual_review$human_label %in% valid_labels))
review_comparison <- scored %>%
select(review_id, text, automated_label, score) %>%
left_join(manual_review, by = "review_id") %>%
mutate(review_status = if_else(human_label == "", "Needs human review", "Human label entered"))
kable(review_comparison)
| review_id | text | automated_label | score | human_label | human_reason | review_status |
|---|---|---|---|---|---|---|
| 1 | I love the quality of this product. The price is reasonable, and delivery was fast. | Positive | 3 | Needs human review | ||
| 2 | The product quality is excellent, but the shipping took too long. | Positive | 1 | Needs human review | ||
| 3 | Great customer service! I would definitely recommend this product to my friends. | Positive | 2 | Needs human review | ||
| 4 | The price is too expensive. However, the product looks beautiful. | Balanced | 0 | Needs human review | ||
| 5 | Fast delivery and excellent service. I am very satisfied with my purchase. | Positive | 3 | Needs human review |
cat("Completed human labels:", sum(manual_review$human_label != ""), "of", nrow(manual_review))
## Completed human labels: 0 of 5
Complete: Identify at least one disagreement or limitation from this table and explain what the dictionary missed. Human Mixed and Neutral labels need not map directly to the automated Balanced or No matches categories.
My final interpretation: [Replace this sentence with 3–5 sentences describing your computed results, your manual check, and one cautious recommendation.]
The following are suggested prompts to send, not a record of completed exchanges. Use at least two AI turns and screenshot the actual prompts and responses in your chat. Include your code or results in the improved prompt.
Prompt 1 — specific initial request
Extend my existing RPubs exercise at https://rpubs.com/michelenorrego/1477019. Use only its five example product-review sentences. Write an HTML R Markdown report using dplyr, tidytext, ggplot2, and plotly. Include a cleaning audit, word frequencies, Bing sentiment scores by review, an interactive bar chart, and verification checks. Keep the original text. Do not invent customer data, execution results, or completed human verification. Explain the limits of the five-sentence dataset.
Prompt 2 — improved follow-up after inspecting the first response
Review and improve your previous R Markdown code. Keep stop-word removal separate from sentiment tokenization; retain reviews without dictionary matches. Distinguish positive, negative, balanced, and no-match scores. Show the exact matched words for each review and add a hand-check of the word product, which appears four times. Provide a blank table for my human labels and explain how mixed reviews and negation can mislead the method. Check the code I paste below and explain any corrections without claiming you ran it. [Paste the first response and any actual Console errors or output here.]
Evidence checklist: Save screenshots of each original prompt, the response, your improved prompt, and its response. Use the exchanges you actually complete. Post the screenshots to the discussion forum; do not submit example prompts as though they were already used.
sessionInfo()
## R version 4.6.2 (2026-10-29)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 24.04.5 LTS
##
## Matrix products: default
## BLAS: /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3
## LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so; LAPACK version 3.12.0
##
## locale:
## [1] LC_CTYPE=C.UTF-8 LC_NUMERIC=C LC_TIME=C.UTF-8
## [4] LC_COLLATE=C.UTF-8 LC_MONETARY=C.UTF-8 LC_MESSAGES=C.UTF-8
## [7] LC_PAPER=C.UTF-8 LC_NAME=C LC_ADDRESS=C
## [10] LC_TELEPHONE=C LC_MEASUREMENT=C.UTF-8 LC_IDENTIFICATION=C
##
## time zone: UTC
## tzcode source: system (glibc)
##
## attached base packages:
## [1] stats graphics grDevices utils datasets methods base
##
## other attached packages:
## [1] knitr_1.52 plotly_4.12.1 ggplot2_4.0.4 tidytext_0.4.3 tibble_3.3.1
## [6] stringr_1.6.0 tidyr_1.3.2 dplyr_1.2.1
##
## loaded via a namespace (and not attached):
## [1] Matrix_1.7-6 gtable_0.3.6 jsonlite_2.0.0
## [4] janeaustenr_1.0.0 compiler_4.6.2 tidyselect_1.2.1
## [7] Rcpp_1.1.2 jquerylib_0.1.4 scales_1.4.0
## [10] yaml_2.3.12 fastmap_1.2.0 lattice_0.23-1
## [13] R6_2.6.1 SnowballC_0.7.1 generics_0.1.4
## [16] htmlwidgets_1.6.4 RColorBrewer_1.1-3 bslib_0.12.0
## [19] pillar_1.11.1 tokenizers_0.3.0 rlang_1.3.0
## [22] cachem_1.1.0 stringi_1.8.9 xfun_0.62
## [25] S7_0.2.2 sass_0.4.10 otel_0.3.0
## [28] viridisLite_0.4.3 cli_3.6.6 withr_3.0.3
## [31] magrittr_2.0.5 crosstalk_1.2.2 digest_0.6.39
## [34] grid_4.6.2 lifecycle_1.0.5 vctrs_0.7.3
## [37] data.table_1.18.6.1 evaluate_1.0.5 glue_1.8.1
## [40] farver_2.1.2 httr_1.4.9 rmarkdown_2.32
## [43] purrr_1.2.2 tools_4.6.2 pkgconfig_2.0.3
## [46] htmltools_0.5.9