Table 3.
Memorisation assessment for 1K samples per n-gram group in the top+meta train-gen-mhr model. high denotes n-grams from the upper frequency quartile; low—n-grams from the lower frequency quartile; %,in denotes percentage of target n-grams in the input key phrases; and %,out—in the respective generated output. Highest PPL values are highlighted in bold.
2-gram | 3-gram | 5-gram | |||||
---|---|---|---|---|---|---|---|
High | Low | High | Low | High | Low | ||
%, In | 16 | 40 | 4 | 12 | 0.3 | 0.8 | |
%, Out | 48 | 48 | 43 | 34 | 41 | 29 | |
PPL, K | 18 | 25 | 17 | 24 | 21 | 24 |