I wonder how the test was done, i.e. if the sample was 50% sarcasm, or the more realistic 1-5%. 80% accuracy on 50% frequency sample can mean a 10% false positive rate (i.e. 10 phrases, five sarcastic phrases were all true positives, four non-sarcastic phrases were true negatives, and one non-sarcastic phrase was a false positive).
If you use a 10% false positive rate test against a 1% frequency sample you get 90% false positive hits within your positive hits (out of a million phrases you'll get eleven thousand positives, out of which ten thousand will be false positives and only one thousand true positives).