Lecture 14: Linguistic Challenges

Final project details:

• Groups, just like normal projects.
  
• Scope: something that with an extra month of work could turn into a paper at a major conference (ACL, EMNLP, NAACL, etc....)
  
• Four styles of projects:
• Reimplement some cool, interesting system from the past 2 or 3 years in an area of NLP/CL that you think is interesting.  (+1 month would be to figure out what it does wrong, propose some new idea, and then get a paper out of it.)
  
• Take some problem you care about from your own research and look at how NLP can help it.  (You're a better judge than I am about turning this into a paper, but at least you care about the topic!)
  
• Pick some interesting, and under-studied linguistic phenomena (like the things we're talking about today) and do a study on this phenomena.  Things not in English are totally allowed, but it would help if you knew the language you were working with.  If you want to do data collection using mechanical turk, I have a budget of $200 for this: we can see how many people want to do this and then try to divide fairly (note: fairly is not the same as evenly).
  
• Something else you can convince me is interesting...
  
• You will be graded entirely on your writeup, which should look more or less like a 4-6 page ACL workshop paper (you're encouraged to use the ACL style files).

Here are some examples of linguistic phenomena that I really don't have much idea how to handle.  After are some additional project ideas....

1. Metaphor.  This can be narrowly construed ("John devoured his steak like a lion") or broadly construed ("John drove Mary crazy").  The usual theory is that there are concrete, physical verbs/nouns; and then when we want to talk about something abstract, we ground it in one of these physical verbs.  So a metaphor will usually consist of one abstract thing and one physical thing.  There are interesting things to be done both on interpreting metaphors as well as generating metaphors.  I have a small amount of data.
   
2. Quantifier degree.  I can say "Most men are jerks" and I can say "Most jerks drink beer" but does this mean that most men drink beer (or rather, that I think that most men drink beer?).  What about other quantifiers?  How consistent is this across different statements ("Most men are jerks" versus "Most animals are invertebrates" -- is the meaning of "most" the same)?  I don't have any data, but you could collect it.
   
3. Metonymy.  This is when I say "The White House" and really mean "Obama" (or at least I mean "someone from the White House").  This combines with ideas of selectional preferences, where we might be able to figure out that the agent (or subject) of "said" should be animate, and might know that the White House is not animate, and so it must be metonymic.  Can we identify metonymic usages?  Can we resolve them?  I have a reasonable amount of data.
   
4. Quantifier scope.  This is a classic semantics problem.  "All the women on the island have a husband."  Does this mean that each woman has her own husband, or that there's one man who is married to all of the women simultaneously.  Similarly, "Each man entered the bar with a dog" -- how many dogs are there?  I have no data.
   
5. Negation scope.  This is perhaps particularly interesting in the case of sentiment analysis.  "I did not like the plot but liked the acting."  -- Does "not" scope over just "like the plot" or over the full "like the plot but liked the acting"?  There is also an interaction with quantifiers ("No man should marry a leprechaun"), but this is both rare and hard.  I have no data, but I think we could find some.
   
6. Language change.  We have English documents going back 500 years, but English back then looks a lot different than English now.  Moreover, our part of speech taggers and parsers that work well now won't work well on Shakespeare (I would guess :P).  Can we track changes back in time, and use this information to get today's models to do a good job way back when?  I have a lot of data.
7. Traces and Ellipsis.  "John went to the store but Jerry didn't ______" -- what was it that Jerry didn't do.  This is a form of a trace.  The Penn Treebank comes with some but not all traces spelled out.  Try to figure out where they should be and what they refer to.
   
8. Exaggeration.  No one has ever looked at this problem.  Okay that's not true, but only because it's an overstatement.  Can we automatically identify overstatements in text?  Can we tone them down or at least identify the degree of overstatement?  I have no data, but it shouldn't be too hard to find some.
   
9. Emotion from text.  Can we identifying character's emotions by looking at what they say, for instance in novels.  I have data of quoted speech, paired with emotional strength -- can we predict emotional strength given the quoted speech and perhaps the context in which it was spoken?  I have lots of data.
   
10. Language and X.  Often we can get language paired with objects in another medium, for instance images and captions.  Can we identify correspondences?  Or perhaps generate captions from images?  We have (or can get) similar data for comments and code, or formal specifications (variant of first order logic) and descriptions.  Or database queries and text.  I have some data, depending on what you want "X" to be.
    
11. Shannon game (word/character prediction) with domain knowledge.... The Shannon game is a way of measuring entropy of a language.  Perhaps make an online game (like hangman?) where we can get people to play this game, and then can also tell them things like "this is a news story" or "this is a children's book" or "this is a medical text" or something and see how it affects their ability to guess the next work.
    
12. "Translating" from tweets to language that your parents could understand.  We have lots of SMS/twitter data, and lots of normal text.  A lot of this is abbreviation expansion, but some is not.  Can we figure out which terms in tweets should be expanded, and then try to find their expansions in normal text?  There are at least two types ("something" -> "sth" or "laugh out loud" -> "lol"), but maybe more.  We can get lots of data easily.
13. Identifying character relationships in literature.  We can track character interactions in literature, but the relationship between these characters is not known.  And perhaps it changes over time.  Can we figure out how two characters in a novel feel about each other on the basis of what they do/say to each other? I have some data and we can get a lot more easily.
14. Many other languages have phenomena that English does not.  For example, subject dropping happens in lots of languages (eg., Japanese) when it's clear from context.  Also stuff in the Semantic Analysis chapter of J+M is pretty much up for grabs.

I'm sure there are lots of things I'm not thinking of that would be fun/interesting to do.  This is your opportunity to be creative and choose something that just plain sounds cool.