Airline Passenger Sentiment Models
Fine-tuned transformer checkpoints from the M.Sc. thesis "Multi-Source Natural Language Processing for Airline Passenger Sentiment and Satisfaction Analysis" (Enes Gerem, Hacettepe University).
Each checkpoint is a 3-class sentiment classifier (negative / neutral / positive) fine-tuned on one of seven source configurations drawn from Twitter, Skytrax, and TripAdvisor airline feedback.
Label mapping
| id | label |
|---|---|
| 0 | negative |
| 1 | neutral |
| 2 | positive |
Available checkpoints (subfolders)
| Family | Source config | Subfolder |
|---|---|---|
| BERT-base | twitter |
bert/twitter |
| BERT-base | skytrax |
bert/skytrax |
| BERT-base | tripadvisor |
bert/tripadvisor |
| BERT-base | twitter+skytrax |
bert/twitter+skytrax |
| BERT-base | twitter+tripadvisor |
bert/twitter+tripadvisor |
| BERT-base | skytrax+tripadvisor |
bert/skytrax+tripadvisor |
| BERT-base | twitter+skytrax+tripadvisor |
bert/twitter+skytrax+tripadvisor |
| RoBERTa-base | twitter |
roberta/twitter |
| RoBERTa-base | skytrax |
roberta/skytrax |
| RoBERTa-base | tripadvisor |
roberta/tripadvisor |
| RoBERTa-base | twitter+skytrax |
roberta/twitter+skytrax |
| RoBERTa-base | twitter+tripadvisor |
roberta/twitter+tripadvisor |
| RoBERTa-base | skytrax+tripadvisor |
roberta/skytrax+tripadvisor |
| RoBERTa-base | twitter+skytrax+tripadvisor |
roberta/twitter+skytrax+tripadvisor |
| DistilBERT-base | twitter |
distilbert/twitter |
| DistilBERT-base | skytrax |
distilbert/skytrax |
| DistilBERT-base | tripadvisor |
distilbert/tripadvisor |
| DistilBERT-base | twitter+skytrax |
distilbert/twitter+skytrax |
| DistilBERT-base | twitter+tripadvisor |
distilbert/twitter+tripadvisor |
| DistilBERT-base | skytrax+tripadvisor |
distilbert/skytrax+tripadvisor |
| DistilBERT-base | twitter+skytrax+tripadvisor |
distilbert/twitter+skytrax+tripadvisor |
Usage
from transformers import AutoTokenizer, AutoModelForSequenceClassification
repo = "enesgerem/airline-sentiment-models"
subfolder = "roberta/twitter" # pick any subfolder from the table above
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=subfolder)
model = AutoModelForSequenceClassification.from_pretrained(repo, subfolder=subfolder)
Citation / data
Datasets are used for non-commercial academic research. The Twitter component is the Crowdflower US Airline Sentiment benchmark (CC BY-NC-SA); the Skytrax and TripAdvisor components are point-in-time scrapes of publicly available review pages, used under each platform's terms of service.