The Wikipedia entry on One Time Pads (OTPs) states that if this cipher is used properly; ie, the keys are truly random and each part of the key is independent of every other part, it's uncrackable, and yields perfect secrecy, i.e., $H(M|C) = H(M)$.
It gives an example saying that a cryptanalysis on a plaintext "HELLO" will yield all plaintexts like "HELLO", "LATER", with equal probabilities.
Now, consider some OTP encrypted data that I know are English sentences. With infinite computational power, I generate all plaintexts. Now, becasuse each word of the sentence is related to nearby words, I can at the very least narrow down the list of possibilities (I don't know the combinatorics of cramming English sentences into M letters), which does not equal perfect secrecy (entropy $H(M)$ appears to have decreased!).
In short, OTP guarantees $H(M|C) = H(M)$, but my question is that $H(M)$ will be reduced by knowledge of the plaintext, so how is prefect secrecy being ensured?