Part-of-speech tagging

Part-of-speech tagging is the process of marking up the words in a text with their corresponding parts of speech. People commonly learn a simplified form of this in their early years of school, identifying nouns, verbs, and so on. However, the term is generally used to refer to computer algorithms to do much the same thing.

Principle

Part-of-speech tagging is harder than just having a list of words and their parts of speech, because some words can represent more than one part of speech at different times. This is not rare -- in many (all?) languages, a huge percentage of word-forms are ambiguous. For example, even "dogs" which is usually thought of as a just a plural noun, can also be a verb:

The sailor dogs the hatch.

"Dogged", on the other hand, can be either an adjective or a past-tense verb. Just which parts of speech a word can represent varies greatly.

Schools commonly teach that there are 8 parts of speech in English: noun, verb, adjective, preposition, pronoun, adverb, conjunction, and interjection. However, there are clearly many more categories and sub-categories. For example, adjectives divide into sub-classes for color, size, number, and other types of properties. And this is not just at the semantic level, because when these sub-types come together, they can only go in certain syntactic orders:

The three big red dogs

is grammatical, but

The red three big dogs

is not. For nouns, plural, possessive, and singular forms can be distinguished. In many languages words are also marked for their "case" (role as subject, object, etc.), grammatical gender, and so on; while verbs are marked for tense, aspect, and other things. In part-of-speech tagging by computer, it is typical to distinguish from 50 to 150 separate parts of speech for English. Work on stochastic methods for tagging Koine Greek has used over 1,000 parts of speech, and found that about as many words were ambiguous there as in English.

History

Research on part-of-speech tagging has been closely tied to corpus linguistics. The first major corpus of English for computer analysis was the Brown Corpus developed at Brown University by Henry Kucera and Nelson Francis, in the mid-1960s. It consists of about 1,000,000 words of running English prose text, made up of 500 samples from randomly chosen publications. Each sample is 2,000 or more words (ending at the first sentence-end after 2,000 words, so that the corpus contains only complete sentences).

The Brown Corpus was painstakingly "tagged" with part-of-speech markers over many years. A first approximation was done with a program by Greene and Rubin, which consisted of a huge handmade list of what categories could co-occur at all. For example, article then noun can occur, but article verb (arguably) cannot. The program got about 70% correct. Its results were repeatedly reviewed and corrected by hand, and later users sent in errata, so that by the late 70s the tagging was nearly perfect (allowing for some cases even human speakers might not agree on).

This corpus has been used for innumerable studies of word-frequency and of part-of-speech, and inspired the development of similar "tagged" corpora in many other languages. Statistics derived by analyzing it formed the basis for most later part-of-speech tagging systems, such as CLAWS and VOLSUNGA. However, by this time (2005) it has been superseded by larger corpora such as the 100 million word British National Corpus.

For some time, part-of-speech tagging was considered an inseparable part of natural language processing, because there are certain cases where the correct part of speech cannot be decided without understanding the semantics or even the pragmatics of the context. This is extremely expensive, especially because analyzing the higher levels is much harder when multiple part-of-speech possibilities must be considered for each word.

In the mid 1980s, researchers in Europe began to use Hidden Markov Models (HMMs) to disambiguate parts of speech, when working to tag the Lancaster-Oslo-Bergen Corpus of British English. HMMs involve counting cases (such as from the Brown Corpus), and making a table of the probabilities of certain sequences. For example, once you've seen an article, perhaps the next word is a noun 40% of the time, an adjective 40%, and a number 20%. Knowing this, a program can decide that "can" in "the can" is far more likely to be a noun than a verb or a modal. The same method can of course be used to benefit from knowledge about following words.

More advanced ("higher order") HMMs learn the probabilities not only of pairs, but triples or even larger sequences. So, for example, if you've just seen an article and a verb, the next item may be very likely a preposition, article, or noun, but even less likely another verb.

When several ambiguous words occur together, the possibilities multiply. However, it is easy to enumerate every combination and to assign a relative probability to each one, by multiplying together the probabilities of each choice in turn. The combination with highest probability is then chosen. The European group developed CLAWS, a tagging program that did exactly this, and achieved accuracy in the 93-95% range.

It is worth remembering, as Eugene Charniak points out in Statistical techniques for natural language parsing [1] (http://www.cs.brown.edu/people/ec/home.html), that merely assigning the most common tag to each known word and the tag "proper noun" to all unknowns, will approach 90% accurate because many words are unambiguous.

CLAWS pioneered the field of HMM-based part of speech tagging, but was quite expensive since it enumerated all possibilities. It sometimes had to resort to backup methods when there were simply too many (the Brown Corpus contains a case with 17 ambiguous words in a row, and there are words such as "still" that can represent as many as 7 distinct parts of speech).

In 1987, Steve DeRose and Ken Church independently developed dynamic programming algorithms to solve the same problem in vastly less time. Their methods were similar to the Viterbi algorithm known for some time in other fields. DeRose used a table of pairs, while Church used a table of triples and an ingenious method of estimating the values for triples that were rare or nonexistent in the Brown Corpus (actual measurement of triple probabilities would require a much larger corpus). Both methods achieved accuracy over 95%. DeRose's 1990 dissertation at Brown University included analyses of the specific error types, probabilities, and other related data, and replicated his work for Greek, where it proved similarly effective.

These findings were surprisingly disruptive to the field of Natural Language Processing. The accuracy reported was higher than the typical accuracy of very sophisticated algorithms that integrated part of speech choice with many higher levels of linguistic analysis: syntax, morphology, semantics, and so on. CLAWS, DeRose's and Church's methods did fail for some of the known cases where semantics is required, but those proved negligibly rare. This convinced many in the field that part-of-speech tagging could usefully be separated out from the other levels of processing; this in turn simplified the theory and practice of computerized language analysis, and encouraged researchers to find ways to separate out other pieces as well. Markov Models are now the standard method for part-of-speech assignment.

The methods already discussed involve working from a pre-existing corpus to learn tag probabilities. It is also possible to bootstrap, using "unsupervised" tagging. Unsupervised tagging techniques use an untagged corpus for their training data and produce the tagset by induction. That is, they observe patterns in word use, and derive part-of-speech categories themselves. For example, statistics readily reveal that "the", "a", and "an" occur in similar contexts, while "eat" occurs in very different ones. With sufficient iteration, similarity classes of words emerge that are remarkably similar to those human linguists would expect; and the differences themselves sometimes suggest valuable new insights.

These two categories can be further subdivided into rule-based, stochastic, and neural approaches. Some current major algorithms for part-of-speech tagging include the Viterbi algorithm, Brill Tagger, and the Baum-Welch algorithm (also known as the forward-backward algorithm). Hidden Markov model and visible Markov model taggers can both be implemented using the Viterbi algorithm.

Expanded article. Help Wikipedia by expanding stub articles.