A Sentiment-labelled Corpus of Hansard Parliamentary Debate Speeches

Hansard transcripts provide access to Members of Parliament’s opinions on many important issues, but are difficult for people to process. Existing corpora for sentiment analysis in Hansard debates rely on speakers’ votes as sentiment labels, but these votes are known to be constrained by speakers’ party affiliations. We develop an annotation scheme and create a novel corpus designed for use in the evaluation of sentiment analysis systems using automatically and manually applied speech labels. Observing the effects on speech sentiment of differing sentiment polarities in debate motions (proposals), we also apply sentiment labels to these motions. We find that humans are able to reach high agreement in identifying sentiment polarity in these debates, and that manually applied and automatically retrieved class labels differ somewhat, suggesting that speech content does not always reflect the voting behaviour of Members of Parliament.