Positive Pointwise Mutual Information §

= Pointwise Mutual Information, in which negative PMI values get replaced by 0

Formula §

$\mathrm{PPMI}\left({\text{ word }}_1,{\text{ word }}_2\right)=\max \left({\log }_2\frac{P\left({\text{ word }}_1,{\text{ word }}_2\right)}{P\left({\text{ word }}_1\right)P\left({\text{ word }}_2\right)},0\right)$

Problems §

• biased towards infrequent events
  • → smoothing or give rare words higher probabilities

🔗 Links

🧮 tf-idf