Hi Peter, This sounds like a great project, and you have done a good job identifying what is likely to work. Part of speech tagging is important simply because it reduces the number of possible senses for a word in a coarse way - for example "train" as a verb has very different senses than "train" as a noun (e.g., train like a boxer versus train engine). So that's a very important first step.
Of all the methods we have looked at, WordNet::SenseRelate::AllWords is going to be the one most likely to be useful. Supervised learning will require a lot of manually annotated data, and typically only deals with one word at a time. So AllWords is probably the best starting point, because it will assign a sense to every content word in a text You could limit it to only dealing with nouns and adjectives, although you might still want to consider every content word as input to the algorithm and then only look at the results of the nouns and adjectives. The reason for considering all the words in the text is that the AllWords methods finds the sense of a word that is most related to its neighbors - so the more neighbors you use the better the results. You can find AllWords here, available for download or via a Web interface. http://www.d.umn.edu/~tpederse/senserelate.html You are also correct, word sense disambiguation as performed by AllWords is not highly accurate in some cases, and so you have to be a little careful in using the results. However, part of speech tagging will help, and in fact we are working on setting some benchmarks that compare part of speech tagged text versus untagged text - our early indications are that it helps quite a bit. You can see a fairly extensive discussion of AllWords results without using part of speech tags in the following MS thesis: Semantic Relatedness Applied to All Words Sense Disambiguation (Michelizzi) - Master of Science Thesis, Department of Computer Science, University of Minnesota, Duluth, July, 2005. http://www.d.umn.edu/~tpederse/Pubs/jason-thesis.pdf What you see there is that AllWords does not beat the "most frequent sense" or "WordNet sense 1" baseline. These are the results that you get by assigning each word its most frequent sense (according to WordNet in this case). In fact it is very hard to beat this baseline, but we are cautiously optimistic that we can if we use more part of speech information. Finally, as to extracting relations, that is not something that is supported in any of our tools, but I think this might be where knowing the degree to which verbs and nouns are semantically similar could be a reasonable approximation. For example... The doctor treated the patient The verb "treat" can be considered the "relation" between doctor and patient, and if "doctor" and "treat" are highly related, and if "treat" and "patient" are also highly related, we might assume that the relation "treat" holds between them. Note that in this case you'd need to use a measure of semantic relatedness (like lesk or vector, in WordNet::Similarity) that would allow you to make measurements between nouns and verbs. The measures of similarity (most of the rest of the measures) only allow for measuring similarity between nouns and nouns or verbs and verbs. That's a very quick approximation of how you might try and deal with relations, there are quite a few other options too. In any case, you raise many interesting issues and I think your instincts on this are dead on. It sounds like a very interesting problem, please feel free to raise any additional questions or concerns you might encounter! Cordially, Ted On Thu, Dec 4, 2008 at 1:14 PM, Peter Muhl <[EMAIL PROTECTED]> wrote: > I'm a social scientist who is hoping to apply sense tagging / WSD as a > first step in a statistical analysis of political speeches (e.g. > presidential candidates 2008) and discussion. I've had some success > applying POS tagging as the first step and am now investigating > whether I can do better w/ sense tagging / WSD. There are, however, > quite a few attractive looking algorithms on Ted Pedersen's web page, > and I'm hoping someone could give me a bit of advice on which one or > ones might be best for my purposes and where I could learn more about > what they can do. > > I'm trying to extract speaker's political concepts and the > relationships between them. So, I'm interested in processing all the > nouns and adjectives in a body of text. What I could ideally use is a > sense tagger that would run through text and with some accuracy > classify a given word into a broad cluster of meaning. For example, > it should distinguish between "bank" in the sense of river bank or > financial institution. SenseRelate looks like it could do this, but > I'm not sure how accurate it is. I'm under the impression, perhaps > incorrect, that supervised WSD would take too long if I end up trying > to disambiguate a thousand or more unique words. > > -- Ted Pedersen http://www.d.umn.edu/~tpederse
