tfs_distilbert

This model is a fine-tuned version of distilber-base-cased designed to classify parliamentary speech into past, present or future orientation.

The training data for the model consists of roughly 3,600 sentences from the UK House of Commons Hansard which have been hand-coded by two coders working independently and reconciling their differences.

Sentences have been pre-processed using Duckling to turn absolute temporal references into relative temporal references. Thus, the sentence

"In the year 2050, global temperatures are forecast to rise by 1 degree. "

said in the year 2023 becomes

"27 years from now, global temperatures are forecast to rise by 1 degree. "

Note that sentences which dealt with conditional claims or were in the irrealis were classified as present-oriented. Sentences or sentence fragments which could not be classified in any other way were classified as present-oriented. The distribution of temporal foci in the training data was as follows:

• Present: 2421 instances
• Past: 825 instances
• Future: 363 instances