Part of speech assignment provides a pos to a word. In many pos systems this can occasionally produce errors due multi-word expressions of one form or another. When 'we' look at the text, we may see it as a compound, common sequence, set phrase or idiomatic expression etc. And we tend to see them as singular pieces. For example, frying pan, computer screen, kicked off. These are not really individual words, but singular concepts represented with spaces in them.

So the thought in my mind is whether it would make sense to assign only a single pos to some/all mwe's? Has anyone done research into such an approach? Are there any obvious grammatical flaws people can see?

  • 2
    Noun Phrase covers all the varieties of words, clauses, and phrases that can function as subject or object in a clause, for instance. That includes pronouns, nouns, modified nouns, extremely modified nouns, complement clauses, headless relatives, prepositional phrases and locatives, and participials of several varieties. Mostly these are governed by the matrix predicate.
    – jlawler
    Oct 15 '14 at 18:02
  • 1
    There's two meanings for "part of speech". The older and traditional one is the simplistic set of word classes imposed on languages by classicists studying Latin and Greek. This is the sense still used by dictionaries but doesn't see much use in linguistics. Now in computational linguistics and NLP there is a more modern sense based on practical work on parsing the syntax of languages by computer. This is usually the sense in "POS tagger" and there can be many many times more POSes than dictionaries use, and many differing sets of POSes devised by different groups, even for the same language. Oct 15 '14 at 21:35
  • Having said that, I personally believe there is a problem lurking here in assigning POSes to orthographic words in written language that becomes much more obvious when trying to parse languages which don't use breaks between words (Burmese, Chinese, Japanese, Khmer, Lao, Thai) or which use breaks between all syllables including word-internally (Dzongkha, Tibetan, Vietnamese). I wonder what state of the art speech recognition software does. I would not regard systems that did speech-to-text as an early step followed by POS-tagging to be state of the art. Oct 15 '14 at 21:40
  • 1
    "These are not really individual words" who says? They look like two word phrases to me. A good parser would tag the individual words as well as identifying them as constituencies.
    – curiousdannii
    Oct 16 '14 at 4:51

I find the question and much of the discussion so far contaminated by confusions between language and writing and between word and phrase. "Frying pan" is a noun; it is a compound, made up of two words, "frying" and "pan", which are both nouns. It is not a noun phrase, though you might be able to find a noun phrase (e.g., the subject of a sentence) which just this noun occurring alone, but at the moment, no such example occurs to me.

The fact that "frying pan" would customarily be spelled with a space between the two nouns that make up the compound, rather than "frying-pan" or "fryingpan", has really nothing to do with its grammatical category. Why should it? The space there is just a spelling convention, not a linguistic hypothesis.

It may be part of your understanding of the term "phrase" that a phrase has more than one word. You ought to try not to carry over this understanding to grammar, because the terms "noun" and "noun phrase", as used in grammatical analysis, have nothing to do with whether a constituent has more than one word. One of the first things that has to be done in teaching modern grammar to a student who is trained in traditional grammar is to convince the student that a noun can never be the subject of sentence -- only a noun phrase can.

Now, can a multi-word constituent ever be a word, as I have claimed above for "frying pan"? Yes, and the case of word compounds is one such case. There are others. Here is a pretty clear one. A general principle of coordination is that the constituent created by combining two constituents has the same grammatical category as the constituents that were combined. Combining the verb phrases "run" and "eat sardines" with "and" gives the verb phrase "run and eat sardines", for instance. Note that I've categorized "run" as a verb phrase, even though it has just one word in it.

Now consider the perfectly ordinary sentence "The man and woman left". What category is "man and woman"? It has more than one word -- does that make it a phrase? No, it's just a noun, formed by the general rule of coordination mentioned above. Start with two nouns, connect them with "and", and you get something of that same category -- a noun. So, if we can agree that a noun is a word, clearly here is a case of a multi-word word.


Sorry, but for some reason I cannot comment - I've had to create an account!

Indeed, NounPhrases cover many things, but it is also far larger in scope than what I referred to (no articles or determiners unless they are part of the "entity"). I think it also misses phrases such as "kicked off". I'm not disputing that NP is correct, it just seems slightly further along the chain than where I'm looking at at present.

I'm aware of the differences and volume of PoS tagsets. That doesn't negate the question though. I do agree that some of the approaches are somewhat askew. Look at how long people went before fathoming the issues regarding multi-word units.

Maybe a rephrase would be better?

If you had the choice, when processing text for PoS tagging, would you treat set phrases and obvious multi-word entities as singular "words" (regardless of the presence of hyphens/spaces)?
Would you tag "frying pan" as [Adjective Noun], or just [Noun], or an alternative noun representation ([Comp.Noun], [Noun-Comp], [Noun Comp-Adj/Nn etc.)? Does anyone know of a tagging system or approach that does or has tried this?

Thank you both for commenting. I hope I've made it a little clearer.

Yes. The examples given were two-word phrases. And yes, many taggers will accurately tag them most of the time, in most cases. And yes, higher level parsers would also identify them as being adjoined.
But that is Not the question being asked. The question asked was whether it should be the case that they are tagged as a collective, rather than as individually. The question asked was whether there were any known systems that did this, or theories that covered treating MWE's as singular pieces rather than multi-part pieces. Think of it this way, they are not individual words when you read them, so why should they be tagged as such? In certain structures/context, they convey different semantics when together than individually. When that happens, they stop being separate words, and instead produce a singular entity. "Big pan" is not the same as "frying pan". "falling pan" is not the same either. Nor is "full-", "deep-", "shiny-", "old-", "ageing-" etc. Yet many of those will fit one of the patterns that "frying pan" does, depending on what PoS you assign to the first word.


And Moderators/Admin - sorry for the answer-abuse ... but unsure how to respond considering the situation.

  • Are you trying to say that "loudbob" and "Robert Noise" are the same person but you had to create the second account due to some difficulty with the system?? Oct 17 '14 at 6:45
  • What do you mean by they are not individual words when you read them? Do you mean "frying pan" is not an individual word or "frying" and "pan" are not each individual words? Oct 17 '14 at 6:50
  • 1
    Yes hippietrail. Something went wrong with my tablet, and I couldn't continue the session from my laptop. I don't think it was a site issue. I should have created an account in the first place really. As for the individual words - "frying" and "pan" are individual words, but when they appear in immediate succession, in a specific order within a given pattern/sequence, those two words convey a singular meaning. They stop being 2 distinct entities, and instead become a single entity - thus why I was wondering if it would make sense to treat them as such whilst at the PoS stage? Oct 17 '14 at 11:59
  • I think you could also fairly say that they are clearly individual orthographic words but a single phonological word. Remember that language is primarily about sounds and that writing is just a recently imposed technology with a good but far from perfect mapping. Thinking only about written words can be distracting to literate people. I personally do think it would make sense to treat them as single words at the PoS stage. But I'm just an interested amatuer and I don't even believe a discrete PoS stage is necessarily such a good thing. Oct 17 '14 at 12:10
  • I agree regarding the sounding, it's all representation of concept - be it audible or visible. Why do you think it makes sense? Oct 17 '14 at 12:28

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.