Need to find positive reviews in Portuguese, but have no labeled examples for training a new classifier? Zero-shot learning lets a pretrained model classify text using candidate labels you supply at inference time. It draws on knowledge from prior training rather than examples labeled for this particular review task. section 6.2.8
A practical workflow starts by asking an AI assistant how to extract positive Portuguese reviews with a zero-shot model. Install `transformers`, then inspect and adapt the suggested Python code to your DataFrame. The example uses a Hugging Face classification pipeline with `positive` and `negative` as candidate labels; it keeps each review’s top label and filters for positive results. section 6.2.8
Listing 6.5 Sentiment analysis with a zero-shot learning model
from transformers import pipeline
import pandas as pd
# Load the zero-shot classification pipeline
classifier = pipeline("zero-shot-classification",
model="facebook/bart-large-mnli")
# Specify the candidate labels
candidate_labels = ["positive", "negative"]
# Define a function to classify a single review
def classify_review(review):
result = classifier(review, candidate_labels=candidate_labels,
hypothesis_template="This review is {}.", multi_label=False)
return result['labels'][0]
# Apply the classification to each review
df['sentiment'] = df['review_comment_message'].apply(classify_review)
# Filter the DataFrame to only include positive reviews
positive_reviews_df = df[df['sentiment'] == 'positive']
The listing shows the model, candidate labels, hypothesis template, and DataFrame filter needed to reproduce the workflow. section 6.2.8
Do not treat a plausible label as proof of accuracy. On 500 non-empty reviews, evaluation against `review_score` gave the zero-shot model 0.87 sensitivity and 0.74 specificity. A multilingual sentiment model scored 0.79 and 0.93 respectively: the zero-shot approach found more positive reviews but was less effective at excluding non-positive ones. Test candidate approaches on your data before choosing one. section 6.2.8 section 6.2.9
Table 6.3 Quality of sentiment analysis with different pretrained models
|
Sentiment analysis (
n=500)
|
Number of reviews classified as positive
|
Sensitivity
|
Specificity
|
|---|---|---|---|
|
Reference review score
|
320 (63%)
true positives
|
1
|
1
|
|
Basic keyword search
|
163 (33%)
|
0.45
|
0.9
|
|
ChatGPT 4 API
|
250 (50%)
|
0.74
|
0.93
|
|
Suboptimal model
|
193 (39%)
|
0.56
|
0.93
|
|
Translated inputs
|
295 (59%)
|
0.86
|
0.89
|
|
Multilingual model
|
265 (53%)
|
0.79
|
0.93
|
|
Zero-shot learning model
|
325 (65%)
|
0.87
|
0.74
|
The comparison table places the zero-shot results alongside multilingual classification, translated inputs, keyword search, and an API-based approach. section 6.2.9
Flexibility is the attraction, not a guarantee: a single zero-shot model can tackle tasks without task-specific labeled training data, but a model fine-tuned for a nuanced domain may perform better. The named model is an example, not necessarily the best choice for every dataset. section 6.2.4 section 6.2.8
