SiEBERT – English-Language Sentiment Classification

Overview

This model (“SiEBERT”, prefix for “Sentiment in English”) is a fine-tuned checkpoint of RoBERTa-large (Liu et al. 2019). It enables reliable binary sentiment analysis for various types of English-language text. For each instance, it predicts either positive (1) or negative (0) sentiment. The model was fine-tuned and evaluated on 15 data sets from diverse text sources to enhance generalization across different types of texts (reviews, tweets, etc.). Consequently, it outperforms models trained on only one type of text (e.g., movie reviews from the popular SST-2 benchmark) when used on new data as shown below.

Predictions on a data set

If you want to predict sentiment for your own data, we provide an example script via Google Colab. You can load your data to a Google Drive and run the script for free on a Colab GPU. Set-up only takes a few minutes. We suggest that you manually label a subset of your data to evaluate performance for your use case. For performance benchmark values across various sentiment analysis contexts, please refer to our paper (Hartmann et al. 2022).

Use in a Hugging Face pipeline

The easiest way to use the model for single predictions is Hugging Face’s sentiment analysis pipeline, which only needs a couple lines of code as shown in the following example:
from transformers import pipeline
sentiment_analysis = pipeline("sentiment-analysis",model="siebert/sentiment-roberta-large-english")
print(sentiment_analysis("I love this!"))


Use for further fine-tuning

The model can also be used as a starting point for further fine-tuning of RoBERTa on your specific data. Please refer to Hugging Face’s documentation for further details and example code.

Performance

To evaluate the performance of our general-purpose sentiment analysis model, we set aside an evaluation set from each data set, which was not used for training. On average, our model outperforms a DistilBERT-based model (which is solely fine-tuned on the popular SST-2 data set) by more than 15 percentage points (78.1 vs. 93.2 percent, see table below). As a robustness check, we evaluate the model in a leave-one-out manner (training on 14 data sets, evaluating on the one left out), which decreases model performance by only about 3 percentage points on average and underscores its generalizability. Model performance is given as evaluation set accuracy in percent.