---
license: mit
language:
- en
base_model:
- FacebookAI/roberta-large
pipeline_tag: text-classification
---

# Classification of sentiments towards AI perceptions from snippets of news-related texts. 
This model classifies sentiments towards AI perceptions "Positive" (2), "Negative" (0), or "Neutral/Mixed" (1)

## Training Dataset

This was fine-tuned on the [Long-Term Trends of Public Perception of Artificial Intelligence (AI)](https://ojs.aaai.org/index.php/aaai/article/view/10635) with [`FacebookAI/roberta-large`](https://huggingface.co/FacebookAI/roberta-large) serving as the base model.  

The Long-Term Trends of Public Perception of Artificial Intelligence (AI) is a dataset that captures nearly 30 years of public perceptions regarding AI. Annotators labeled perceptions based on 5,685 paragraphs extracted from New York Times (NYT) articles related to AI, spanning 1986 to 2016. Perceptions were condensed to "Positive", "Negative", or "Neutral/Mixed".  

2,000 data points were sampled through continous-time series clustering and were used to finetune the data in order to mimic the constraints of annotation and finetuning in policy-related studies. (To understand why we do that, refer to our paper at the reference below) 

## How to use model

Here is some source code to get you started on using the model to classify the sentiments. 

```python
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
from tqdm import tqdm
import pandas as pd


def classify_tweets(df, text_col, model, tokenizer):
    df[text_col]=df[text_col].astype(str)
    device = 0 if torch.cuda.is_available() else -1  # Use GPU if available
    classifier = pipeline(
        "text-classification",
        model=model,
        tokenizer=tokenizer,
        device=device,
        truncation=True,  # Ensures inputs don't exceed max length
        max_length=512,   # Manually set to avoid exceeding model's limit
        padding="max_length"  # Ensures all inputs have the same length
    )
    
    outcomes, probs, pred_labels = [], [], []
    for text in tqdm(df[text_col]):  # Fixed tqdm syntax
        preds = classifier(text, return_all_scores=True)
        outcomes.append(preds)
        
        # Extract probabilities and predicted label
        label_scores = {entry['label']: entry['score'] for entry in preds[0]}
        probs.append(list(label_scores.values()))
        pred_labels.append(max(label_scores, key=label_scores.get))
    df["predicted_label"] = pred_labels
    return df

model_name="cja5553/AI-perception-sentiments-roberta-large"
label2id={"negative":0,"neutral/mixed":1,"positive":2}
id2label = {v: k for k, v in label2id.items()}
model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=len(label2id), 
                                                           label2id=label2id,id2label=id2label,
                                                           use_auth_token=False).to("cuda")
tokenizer = AutoTokenizer.from_pretrained(model_name,use_auth_token=False)
text_col="text" # change text column accordingly
df_with_classification=classify_tweets(df, text_col, model, tokenizer)
```

## Citation
If you find this model useful, please cite the following paper:
```
@inproceedings{
author={Charles Alba, Benjamin C Warner, Akshar Saxena, Jiaxin Huang, Ruopeng An},
title={Towards Robust Sentiment Analysis of Temporally-Sensitive Policy-Related Online Text},
year={2025},
booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), Volume 4: Student Research Workshop.},
url={https://aclanthology.org/2025.acl-srw.70/}}
```



## Code 

Code used to train these models are available on GitHub at [github.com/cja5553/ctscams](https://github.com/cja5553/ctscams)

## Questions?
contact me at alba@wustl.edu