I am looking to know about the distribution of word senses, of a particular word in a corpus. How often a word sense will occur out of all the occurrences of that word with any meaning.

For example: if Zipf's law applied we would have a rule relating how often each word sense occurs, that says something like:

The most frequent sense of some given word (generally) occur roughly twice as often as the second most frequent word

I'm not in particular looking for Zipf Distribution, though, any distribution will do.

My general impression of sense counts in Semcor/WordNet suggest that suggests that the most common sense hugely dominates. I have not run and proper regressions to check.

I am looking for a paper I can cite to back up statements about incredibly rare, the rare word senses are compared to the common ones.

  • 1
    The problem is that word senses, unlike word forms, can not so easily be counted, because meaning is always very vague - the biggest problem being that meaning is not discrete and linear, but structured to some degree hierarchically and with a lot of mutual relations and overlapping. For example, with the notion of hyperonym and hyponym, when the sense of "cat" occurs, does that mean that the sense of "animal" occurs? And what about presuppositions or entailment relations; when the sense "succeed in" occurs, does that mean that the sense "try to" occurs? Word senses are not discrete entities.
    – lemontree
    Oct 4 '16 at 9:51
  • 1
    You are right in the that word senses are not clearly defined, but Lexicographers have spent a lot of time making resources like WordNet and BabelNet that do define a finite list of discrete entities (as a approximation to the truth), and have annotated corpora with them like Semcor; I'm happy for my question to treat these deiscrete-ized senses as word senses. And I am not looking for a count of how often a Meaning occurs, just how often a particular sense occurs for a given Lemma. So I'm not interested in counting occurances of cat as a occurance of animal. Oct 4 '16 at 10:02
  • There is certainly published work relating to this, using this meaning of word-sense. For example Zipf 1945 (The meaning-frequency relationship of words, The Journal of general psychology, 1945, 33, 251-256), which related how many senses a word has (in a dictionary) to how frequent the word is in a corpus. Oct 4 '16 at 10:03
  • It's much easier technically to count word string frequencies, you just count the instances of the string. A word sense, a distinct entry under the same string, needs a lot more context to determine. I don't know if anyone has done a corpus analysis distinguishing polysemes. A lot of technical work would have to go into determining for every string which entry is meant. One can easily write a program that determines the number of distinct meanings for a string (given a dictionary). From this I expect that: 1) shorter words have more polysemy 2) more common strings have more polysemy .
    – Mitch
    Oct 4 '16 at 16:09
  • Mathematical inituition suggests that even if the domain is distinct meanings (rather than Zipf's distinct strings), it'll still be a power-law (negative exponential). Of course Zipf's law isn't exact itself either (not exactly a power-law)
    – Mitch
    Oct 4 '16 at 16:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.