Polite Bert

Model Details

Polite Bert is, as the name implies, a BERT model trained to classify a given sentence on a scale of politeness:

Not Polite (aka Rude or Impolite)
Neutral
Somewhat Polite
Polite

Training

Polite Bert was trained by fine-tuning a BERT model on annotated politeness-level data. The model was trained using SFT for 4 epochs, with a batch size 16, and a max sequence length of 128 tokens.

Data Details

The training data consisted of 2000 annotated sentences. This training data was composed of the following:

Manually annotated data:

250 sentences sampled from EUROPARL dataset. Specifically from the English version of PT-EN data.
250 sentences sampled from SIMMC2.0 dataset. From any domain (Fashion or Furniture) and speaker (System or User).
250 sentences sampled from the Philosophy and Politics data of the StackExchange dataset.
250 sentences sampled from a collection of hotel review replies from Trip Advisor.

Automatically annotated data:

1000 sentences from 4Chan Pol dataset. Specifically, we only considered sentences annotated with TOXICITY > 0.85, SEVERE_TOXICITY > 0.85, and INSULT > 0.5.

While we manually labelled the first 1000 sentences, the 1000 sentences from 4ChanPol were automatically set to Not Polite.

These source datasets were chosen due to their likelihood of containing distinct, but pronounced, politeness levels (hate speech from 4chan, formal and polite speech from hotel staff and parliament members, etc)

License

Apache 2.0 license.

Made by Diogo Glória-Silva. PhD Student at NOVA FCT and Affiliated PhD Student at CMU