The biggest difficulty in Bayesian modeling is usually choosing the prior i.e. the P(W) term. There are a few maximum entropy techniques, but basically the rule is: "what works best".
In this case P(W) is a 1-gram probability which is clearly wrong, because it amounts to saying that every word has a probability of appearing that is independent of its surroundings. For example, while 'the' is very likely to appear you don't want it to appear next to another 'the'.
Even when using more sophisticated priors that take n-grams into account, one never has a real model of the English language, because that will depend on the context, the period in which it was written, the linguistic domain…
That is why tweaks are actually justified theoretically.
Another interesting thought that popped out in the comments to the original post is that if one uses a statistical mechanics interpretation of probability, where P(W) = exp(-H(W)/T) where H is some energy function and T is temperature, One can interpret the 1.5 factor as lowering the temperature and making the probability less fuzzy.
In this case P(W) is a 1-gram probability which is clearly wrong, because it amounts to saying that every word has a probability of appearing that is independent of its surroundings. For example, while 'the' is very likely to appear you don't want it to appear next to another 'the'.
Even when using more sophisticated priors that take n-grams into account, one never has a real model of the English language, because that will depend on the context, the period in which it was written, the linguistic domain…
That is why tweaks are actually justified theoretically.
Another interesting thought that popped out in the comments to the original post is that if one uses a statistical mechanics interpretation of probability, where P(W) = exp(-H(W)/T) where H is some energy function and T is temperature, One can interpret the 1.5 factor as lowering the temperature and making the probability less fuzzy.