What do customers complain about most often? First identify the negative reviews; then condense them into something a decision-maker can read quickly. Text summarization turns longer text into a shorter, coherent account that retains essential information. It can also make reports, support messages, and research easier to review. section 6.3 section 6.3.1
Choose an output that fits the source:
- An extractive summary selects important words or sentences from the original. Rule-based scoring can use sentence position, length, or frequent terms. section 6.3.4
- An abstractive summary conveys the main ideas in new wording. Generative AI can produce one, but may drift toward rewriting even when asked to extract. section 6.3.2
- For a short customer review, keywords or topics—possibly paired with sentiment analysis—may be more useful than a miniature prose summary. section 6.3.4
Start with a review you can check
Select a negative review using a review score, then prompt ChatGPT to summarize it. One example uses a Portuguese complaint about a product not arriving and asks for a Portuguese summary plus English translations of both the review and summary. Compare those outputs with the original: a shorter answer is useful only if it preserves the complaint. section 6.3.3
Listing 6.6 Summarizing the longest review—a frequency-based approach
import pandas as pd
import nltk
import string
from collections import Counter
# Download the required NLTK resources
nltk.download('punkt')
nltk.download('stopwords')
# Sample data (manually adapted to remove empty records)
df = pd.read_csv('olist_order_reviews_dataset.csv')
df = df.dropna(subset = ['review_comment_message'])
# Function to tokenize and remove stopwords
def preprocess(text):
stopwords = nltk.corpus.stopwords.words('portuguese')
tokens = nltk.word_tokenize(text.lower())
tokens = [token for token in tokens if token not in string.punctuation
and token not in stopwords]
return tokens
# Function to create word frequency distribution
def word_frequency(tokens):
frequency = Counter(tokens)
return frequency
# Function to summarize short reviews
def summarize_reviews(text, num_keywords=3):
tokens = preprocess(text)
frequency = word_frequency(tokens)
important_words = [word for word, count in
frequency.most_common(num_keywords)]
summary = ' '.join(important_words)
return summary
# Apply the function to the DataFrame
df['summary'] = df['review_comment_message'].apply(summarize_reviews)
# Display the results (manually adapted to print the summary of the longest
message)
print("Longest review:", df.loc[1316]["review_comment_message"])
print("Summary:", df.loc[1316]["summary"])
The frequency-based Python workflow offers a simpler comparison. It reads a review CSV with pandas, removes records without review text, tokenizes Portuguese with NLTK, removes stopwords and punctuation, and counts remaining words. Inspect what the selected words capture—and what context they leave behind. section 6.3.4
Table 6.5 Summary efforts on a sample of five customer reviews
|
Translated review
|
Three-word summary by frequency-based methods
|
Three-word summary by ChatGPT 4
|
ChatGPT’s haiku
|
|---|---|---|---|
|
I received it well before the stipulated deadline.
|
received
well
before
|
Early delivery
|
Arrived well before
Stipulated deadline nears
Satisfaction found
|
|
Congratulations Lannister stores, I loved shopping online, safe and practical. Congratulations to everyone and happy Easter.
|
congratulations
lannister
stores
|
Lannister stores praised
|
Lannister stores cheer
Safe and practical shopping
Happy Easter joy
|
|
Efficient device. On the website, the device brand is printed as 3desinfector, and upon arrival, it has another name... Update with the correct brand since it’s the same device.
|
efficient
brand
device
|
Efficient, mislabeled device
|
Efficient device
Misnamed but still performs well
Update with true brand
|
|
But a little, lagging... for the price, it’s good.
|
a bit
lagging
...
|
Lagging, decent value
|
Slight lag appears,
Yet value stands its ground,
Acceptable deal.
|
|
Reliable seller, product ok, and delivery before the deadline
|
reliable
seller
product
|
Reliable seller, prompt
|
Trustworthy vendor,
Product meets expectations,
Early delivery.
|
Five translated reviews place three-word frequency outputs beside ChatGPT summaries and haiku. Use the original review as the check, rather than treating the most fluent output as automatically the most faithful. section 6.3.4
Match the method to the job. Longer narratives may benefit from abstractive overviews; technical documents may call for a mix of extracted details and synthesis. At scale, also weigh generative AI’s speed and cost against the control lost when a company depends on third-party models or APIs. section 6.3.4 section 6.3.2 section 6.4
