Tagging parts of speech when proper noun is a composite

Question

By means of parts of speech tagging, words of a given sentence can be assumed to be noun/verb etc, but if the sentence is for instance:

"My favourite book is harry potter and the prizoner of azkaban"

note that the inputs I receive would be from a chat interface so having a fixed format for the data can't be expected. Is there a way to identify "harry potter and the prizoner of azkaban" as a proper noun from such messages?

Currently this query tags as:

My|PRP$ 
favourite|JJ 
book|NN 
is|VBZ 
harry|JJ 
potter|NN 
and|CC 
the|DT 
prizoner|NN 
of|IN 
azkaban|NN

I would like to know if this can be handled some way, or if there is another algorithm that can handle this?

Answer 1

I guess your problem is a form of NER (Named-Entity Recognition) tagging. NER tags consist of PER(person), LOC(location), ORG(organization) and MISC(miscellaneous) and O(other). with the help of a NER-tagger algorithm, you probably would have:

My(O) favorite(O) book(O) is(O) harry(B-MISC) potter(I-MISC) and(I-MISC) the(I-MISC) prisoner(I-MISC) of(I-MISC) azkaban(I-MISC).

now you can tokenize your text with identifying different NER tags, and join "B-" and "I-" prefixes with the same NER tag. e.g. you have "harry potter and ..." as a single token, which is a MISC and starts from harry (since harry is B-MISC) and ends to azkaban (since azkaban has the last consecutive I-MISC tag). Now you can let your POS tagger act the "harry potter ..." as a single token, and it must tag it with "NN"

Another solution for this problem is called "chunking". It works based on a set of rules and detect Noun Phrases (NPs). Here, you define a rule and create the regexp statement for it. e.g. you define all consecutive NNs as a single NP (harry potter): <NN>+ or a DT followed by a string of NNs (the brown fox): <DT>?<NN>+. Now you tokenize your text based on a regexp matching method. but in your case, chunking such a long NP is almost impossible. because a rule that finds this NP as a NP candidate, would by mistake find a lot of other string of words as a NP candidate too, while they're not a NP at all. (see nltk: information extraction)