PowerTutor | Model based power estimator by application | Time Series Database library

All you can do is look for commonly occurring strings in your data, and populate the 32K with those strings. That's all there is to it. As noted in the reference you found, the compression will be better if the most common strings are towards the end of the 32K, and the least common towards the start.

The benefit of the dictionary has nothing to do with character frequencies, but rather with matching strings. If the compressor can find a string in the dictionary of three or more characters, it can replace that string with a reference to the dictionary, which is a distance and a length. (Shorter distances code in fewer bits, hence putting more common strings at the end of the dictionary, which is closer to the data being compressed.)

Character frequencies are already handled dynamically by the Huffman coder, which is independent of the dictionary.

I don't have your list of addresses, but you can look for common city, state combinations, possibly including the first few digits of the zip code. You can look for common address components, like " Ave.", spelled out ordinal numbers like "First", "Second", street names like "Main", "Hill", "Lake", "Elm", etc. Those could be combined, e.g. "First St.", "Main St.". Possibly quadrants " NE ", " SW ". Note the inclusion of spaces as part of the string to be matched.