Assuming there is no inherent bias in terms of sentiment and vocabulary, one approach would be to repeatedly randomly sample 300 negative words from the corpus

Assuming there is no inherent bias in terms of sentiment and vocabulary, one approach would be to repeatedly randomly sample 300 negative words from the corpus and generate a vector of sentiments. You could then average the elements of the vector to get an average sentiment, or use another metric from basic stats. That could decrease the bias.