A Clustering Approach for the Unsupervised Recognition of by rt3463df

VIEWS: 3 PAGES: 44

									A Clustering Approach for
 the Nearly Unsupervised
 Recognition of Nonliteral
                Language
           Julia Birke & Anoop Sarkar
       SIMON FRASER UNIVERSITY
                 Burnaby BC Canada
  Presented at EACL ‟06, April 7, 2006
 The Problem
                      She hit the ceiling.

               ACCIDENT?            or             OUTRAGE?
(as in “she banged her hand on the ceiling”)
                                          (as in “she got really angry”)

The Goal
            Nonliteral Language Recognition

                                                                           2
Outline
   Motivation
   Hypothesis
   Task
   Method
   Results
   TroFi Example Base
   Conclusion

                         3
Motivation
   “She broke her thumb while she was
    cheering for the Patriots and, in her
    excitement, she hit the ceiling.” [from Axonwave
                                       ACCIDENT
    Claim Recovery Management System]
   “Kerry hit Bush hard on his conduct on the
    war in Iraq.” → “Kerry shot Bush.” [from RTE-1
                                        FALSE
    challenge of 2005]
   Cannot just look up idioms/metaphors in a list
   Should be able to handle using same method
                                                       4
Motivation (cont)
   Literal/Nonliteral language recognition is a “natural
    task” as evidenced by high inter-annotator
    agreement
    (Cohen) and  (S&C) on a random sample of 200
    annotated examples annotated by two different
    annotators: 0.77
   As per ((Di Eugenio & Glass, 2004), cf. refs therein),
    standard assessment for  values is that tentative
    conclusions on agreement exists when .67    .8;
    definite conclusion on agreement exists when   .8.

                                                             5
Hypothesis
   It is possible to look at a sentence and
    classify it as literal or nonliteral
   One way to implement this is to use similarity
    of usage to recognize the literal/nonliteral
    distinction
   The classification can be done without
    building a dictionary


                                                     6
Hypothesis (cont)
   Problem: New task = No data
   Solution: Use nearly unsupervised algorithm
    to create feedback sets; use these to create
    usage clusters
   Output: An expandable database of
    literal/nonliteral usage examples for use by
    the nonliteral language research community


                                                   7
Task
   Cluster usages of arbitrary verbs into literal and nonliteral by
    attracting them to sets of similar sentences
                                Mixed




              Nonliteral                        Literal                8
Task (cont)
TroFi Example Base                    ABSORB
                                         EAT
***absorb***
*nonliteral cluster*            I sponge absorbs water.
                            Thiswant to eat chocolate.
                                  had to eat my curb the
wsj02:2251 U Another optionI will be to try to words. growth in education and
                            This company absorbs cash.
             other local assistance , which absorbs 66 % of the state 's budget ./.
wsj03:2839 N “ But in the short-term it will absorb a lot of top management 's
                   Literal                                Nonliteral
             energy and attention , '' says Philippe Haspeslagh , a business
             professor at the European management school , Insead , in Paris ./.
*literal cluster*
wsj11:1363 L An Energy Department spokesman says the sulfur dioxide
             might be simultaneously recoverable through the use of powdered
             limestone , which tends to absorb the sulfur ./.




                                                                                 9
Method
       TroFi
         uses a known word-sense disambiguation
          algorithm (Yael Karov & Shimon Edelman, 1998)
         adapts algorithm to task of nonliteral language
          recognition by regarding literal and nonliteral
          as two senses of a word and by adding various
          enhancements




                                                        10
Data Sources
   Wall Street Journal Corpus (WSJ)
   WordNet
   Database of known metaphors, idioms, and
    expressions (DoKMIE)
       Wayne Magnuson English Idioms Sayings &
        Slang
       Conceptual Metaphor WWW Server



                                                  11
Input Data                           Target Word
                                                    WSJ Sentences
                                                    → Feature Sets

1.In this environment,–it's pretty easy to get the ball rolling.
 roll, revolve, turn over –
    General turn over
     roll, revolve,                                Target Word
roll off the tongue to
      off the or cause
roll(to rotatetongue .. .. rotate;                        Examples
                                                          Def‟ns
                                                          Target Word
     Feature sets of stemmed                             → & Examples
                                                      Def‟nsFeature Sets
    "The child rolled down the hill"; nouns and verbs; → Seed Words
                                                        Synonyms
natural to say, targetto pronounce .. ..
         remove easy words, seed                        → Seed Sets
                                                      → FeatureWords
natural rolled the ball"; pronounce words, and frequent words
    "She to say, easy to
 "They rolled their eyes at his words";
    Target Set
    "turn over to your left side";
Podnzilowicz is a name that doesn't roll off the tongue.
Podnzilowicz is a name that doesn't roll off the tongue.
     WSJ sentences containing target word
    "Ballet dancers can rotate their legs outward")
2. wheel, roll –– Feedback Set
 Nonliteral
     wheel, roll
    (move along on or as if on wheels or a wheeled vehicle;
     WSJ sentences containing DoKMIE seeds
                      convoi rolled past the crowds")
    "The President's convoy rolled past the crowds")
       DoKMIE examples
   Literal Feedback Set
       WSJ sentences containing WordNet seeds
       WordNet examples
                                                                           12
Input Data (cont)
 Target
Word Similarity MatrixSimilarity Matrix
Original Sentence
               environ child convoi
             ball ballet ball suv hillcrowd dancer environ ey hill leg name paper podnzilowicz presid rotat side suv tongu vehicl wheel word
environ ball
ball            1      0    0
                            0      0   0 0      0   Target Set 0
                                                      0 00 0 0        0                               0    0   0   0      0      0   0     0
suv hill                    0          0
ballet                 1    0      0     0     0
                                                    environ 0ball
                                                  0 0 0 0       0     0                               0    0   0   0      0      0   0     0

child                       1      0     0     0  0 0 0 0
                                                    suv hill 0 0      0
                                             The SUV rolled 0down the hill.
                                                                                                      0    0   0   0      0      0   0     0

convoi                             1     0     0  0 0 0 0       0     0                               0    0   0   0      0      0   0     0

crowd                                    1      0        0 0   0   0     0      0            0        0    0   0   0      0      0   0     0

dancer                                          1
                    Literal Sentence Similarity Matrix 0 0     0   0     0      0            0        0    0   0   0      0      0   0     0

environ                              1 0 0                         0     0      0            0 ballet
                                                                                                    0      0   0   0    presid
                                                                                                                          0      0   0     0

ey
          Literal Feedback Set          1 0                        0      0    0
                                                                                    Nonliteral Feedback Set 0
                                                                                           dancer     wheel     convoi
                                                                                  I can‟t0 pronounce0 that word.
                                                                                         0
                                                                                                 leg 0
                                                                                           rotat 0       0        0
                                                                                                      vehicl 0 crowd 0                     0

hill
          paper the rotat child
       She turned paper paper over. hill ball
               environ                     1                       0
                                                                         ey word side
                                                                          0    0
                                                                                    word         0    0  0        0    0 0                 0

leg       rotatball      0      0      0
                                                                   1
                                                                       0
                                                                          0    0
                                                                                 0
                                                                                    podnzilowicz name0 tongu 0
                                                                                        0
                                                                                         0       0
                                                                                                    0
                                                                                                      0  0 0
                                                                                                              0        0
                                                                                                                       0                   0

name      child hill
               suv hill  0      0      0                               0 1      00           00       0   00   0   00     0      0
                                                                                                                                 0   0     0

paper
          ball                                                                  1            0        0    0   0   0      0      0   0     0

podnzilowicz                                                                                 1        0    0   0   0      0      0   0     0

presid
          ey word                                                                                     1    0   0   0      0      0   0     0

rotat
          side                                                                                             1   0   0      0      0   0     0

side      ballet dancer rotat leg                                                                 Nonliteral Sentence Similarity Matrix
                                                                                                               1 0      0     0    0
                                                                                                                          podnzilowicz
                                                                                                                                           0

suv       wheel vehicl                                                                                        word1           0    0
                                                                                                                        0 name tongu       0
                                                                                                  environ
tongu
          presid convoi crowd                                                                     ball
                                                                                                                        1
                                                                                                                        0
                                                                                                                              0    0
                                                                                                                                          0
                                                                                                                                           0

vehicl                                                                                                                        1    0      0
                                                                                                  suv hill              0                 0
wheel                                                                                                                                1     0
                                                                                                                                               13
word                                                                                                                                       1
Similarity-based Clustering
   Principles
       Sentences containing similar words are similar; words
        contained in similar sentences are similar
       Similarity is transitive: if A is similar to B and B is similar to
        C, then A is similar to C
       Mutually iterative updating between matrices; stop when
                            changes in similarity values < threshold
           Target SSM




              WSM                 Literal SSM         Nonliteral SSM

                                                                             14
 1


0.9


0.8
                                                                                                                                                                                                                  director
                                                                                                                                                                                                                  essenti
0.7
                                                                                                                                                                                                                  fipp
                                                                                                                                                                                                                  financ
0.6
                                                                                                                                                                                                                  hand
                                                                                                                                                                                                                  idea
0.5
                                                                                                                                                                                                                  institut
                                                                                                                                                                                                                  kaisertech
 0.4
                                                                                                                                                                                                                  mother
                                                                                                                                                                                                                  philosophi
 0.3
                                                                                                                                                                                                                  president
                                                                                                                                                                                                                  principl
 0.2                                                                                                                                                                             trouble
                                                                                                                                                                               quandari                           quandari
                                                                                                                                                                             president                            strait
 0.1
                                                                                                                                                                         mother                                   trouble
                                                                                                                                                                       institut
      0
                                                                                                                                                                     hand
          director
                     essenti
                               fipp
                                      financ




                                                                                                                                                                  fipp            1
                                               hand
                                                      idea
                                                             institut
                                                                        kaisertech
                                                                                     mother
                                                                                              philosophi




                                                                                                                                                                director
                                                                                                           president

                                                                                                                       principl

                                                                                                                                  quandari




                                                                                                                                                                                 0.9
                                                                                                                                             strait

                                                                                                                                                      trouble




                                                                                                                                                                                 0.8


                                                                                                                                                                                 0.7


                                                                                                                                                                                 0.6


                                                                                                                                                                                  0.5


 1 She grasped her mother's hand.                                                                                                                                                 0.4

 2 He thinks he has grasped the essentials of the                                                                                                                                 0.3

    institute's finance philosophies.                                                                                                                                             0.2
 3 The president failed to grasp KaiserTech's                                                                                                                                                                                                                            3
                                                                                                                                                                                                                                                                     president kaisertech financ quandari
                                                                                                                                                                                                                                                                               kaisertech financ quandari

    finance quandary.                                                                                                                                                             0.1

                                                                                                                                                                                        0
                                                                                                                                                                                                                                                                 2
                                                                                                                                                                                                                                                                essenti institut financ philosophi
                                                                                                                                                                                                                                                                        institut financ philosophi



                                                                                                                                                                                               1
                                                                                                                                                                                            mother hand                                                     1
                                                                                                                                                                                                                                                           mother hand                     15
                                                                                                                                                                                                                   2
                                                                                                                                                                                                          essenti institut financ
                                                                                                                                                                                                               philosophi
                                                                                                                                                                                                                                            3
                                                                                                                                                                                                                                    president kaisertech
                                                                                                                                                                                                                                      financ quandari
                                                                                                                                                                                       1


 1                                                                                                                                                                                   0.9


                                                                                                                                                                                     0.8
0.9
                                                                                                                                                                                      0.7


0.8                                                                                                                                                                                   0.6


                                                                                                                                                                                      0.5
0.7
                                                                                                                                                                                      0.4


 0.6                                                                                                                                                                                  0.3                                                   president k. financ quandari
                                                                                                                                                                                                                                                president kaisertech financ quandari
                                                                                                                                                                                                                                                          kaisertech financ quandari



                                                                                                                                                                                      0.2
                                                                                                                                                                                                                                         essenti institut financ philosophi
                                                                                                                                                                                                                                       essenti institut financ philosophi
                                                                                                                                                                                                                                               institut financ philosophi
 0.5
                                                                                                                                                                                       0.1


 0.4                                                                                                                                                                                        0                               mother hand
                                                                                                                                                                                                                                 hand
                                                                                                                                                                                                                          mother hand




 0.3
                                                                                                                                                                                                 L1
                                                                                                                                                                                                mother hand



                                                                                                                                                                                                His aging mother gripped his hands tightly.
 0.2                                                                                                                                                                              trouble
                                                                                                                                                                                quandari
  0.1                                                                                                                                                                         president
                                                                                                                                                                          mother
                                                                                                                                                                        institut     0.45
       0
                                                                                                                                                                      hand
           director
                      essenti
                                fipp




                                                                                                                                                                                       0.4
                                       financ




                                                                                                                                                                   fipp
                                                hand
                                                       idea
                                                              institut
                                                                         kaisertech
                                                                                      mother
                                                                                               philosophi




                                                                                                                                                                 director
                                                                                                            president

                                                                                                                        principl

                                                                                                                                   quandari




                                                                                                                                                                                      0.35
                                                                                                                                              strait

                                                                                                                                                       trouble




                                                                                                                                                                                       0.3


                                                                                                                                                                                      0.25


N1 After much thought, he finally grasped      0.2

   the idea.                                  0.15

N2 This idea is risky, but it looks like the    0.1
   director of the institute has finally                                                                                                                                                                                                                                                      3
                                                                                                                                                                                                                                                                                            president kaisertech financ quandari
                                                                                                                                                                                                                                                                                            presidentkaisertech financ quandari

   comprehended the basic principles behind it.
                                              0.05


N3 Mrs. Fipps is having trouble comprehending 0                                                                                                                                                                                                                                         2
                                                                                                                                                                                                                                                                                       essenti institut financ philosophi


                                                                                                                                                                                                                                                                                                               16
   the legal straits.                                                                                                                                                                           N1idea                                                                        1
                                                                                                                                                                                                                                                                            mother hand


                                                                                                                                                                                                                   N2
                                                                                                                                                                                                              idea director institut
                                                                                                                                                                                                                    principl
                                                                                                                                                                                                                                                 N3
                                                                                                                                                                                                                                             fipp trouble strait
                                                                                  1 The girl and her brother grasped their mother's
0.45                                                                                 hand.
                                                                                  2 He thinks he has grasped the essentials of the
                                                                                     institute's finance philosophies.
 0.4


0.35                                                                              3 The president failed to grasp KaiserTech's
                                                                                     finance quandary.
 0.3


 0.25                                                                             L1 The man's aging mother gripped her husband's
  0.2
                                                                                             shoulders tightly.
                                                                                  L2 The child gripped her sister's hand to cross
 0.15                                                                                        the road.
   0.1
                                                                                  L3 The president just doesn't get the picture,
                                                                              3              does he?
                                                                             president kaisertech financ quandari
  0.05


        0
                                                                         2
                                                                       essenti institut financ philosophi
                                                                                   0.45

            man mother husband                                    1
                                                                 girl brother mother hand
                 L1
                 shoulder        child sister hand
                                                                                      0.4
                                      L2
                                    cross road
                                                       L3
                                                     president
                                                                                    0.35


            N1 After much thought, he finally                                         0.3

               grasped the idea.                                                     0.25
            N2 This idea is risky, but it looks like
               the director of the institute has                                      0.2


               finally comprehended                                                  0.15

               the basic principles behind it.                                         0.1
            N3 Mrs. Fipps is having trouble
               comprehending the legal straits
                                                                                      0.05                                                                                             3
                                                                                                                                                                                     president ka


               of the institute.                                                            0                                                                                    2
                                                                                                                                                                                essenti institut fin

            N4 She had a hand in his finally fully                                                  N1
                                                                                                     idea

               comprehending their quandary.
                                                                                                            idea director institut
                                                                                                                  N2                                                        1
                                                                                                                                                                           girl brother mother han
                                                                                                                                                                                17
                                                                                                                  principl
                                                                                                                                         N3
                                                                                                                                     fipp trouble strait
                                                                                                                                           institut           N4
                                                                                                                                                           hand quandari
High Similarity vs. Sum of
Similarities
                  0.3
                 0.45



                  0.4

                 0.25

                 0.35



                  0.2
                  0.3



                 0.25
    Similarity




                                                                                                                                       Sum of Sim
                                                                                                                               Literal Highest Sim
                 0.15
                                                                                                                                          Sum of Sim
                                                                                                                               Nonliteral Highest Sim
                  0.2



                  0.1
                 0.15



                  0.1

                 0.05

                 0.05



                   0
                        girl brother mother hand   essenti institut financ philosophi   president kaisertech financ quandari
                                                        Original Sentences
                                                                                                                                                        18
                               C
                       LEARNER A
                               D
                               B

                                     overlap
                       INDICATOR : phrasal/expression words AND overlap
                                     n/a
Scrubbing & Learners          feature
                       TYPE : synset set
                              n/a
                                 remove
                       ACTION : move
                                 n/a

   Scrubbing          child sister hand cross firmly)
                   1. grasp, grip, hold on -- (hold road
                   2. get the the feedback savvy,
     cleaning noise out ofpicture, comprehend,sets dig, grasp,
                                                       hand quandari
                   compass, apprehend -- (get the meaning of something;
     Scrubbing Profile comprehend the meaning of this letter?")
                   "Do you
      INDICATOR
        the linguistic phenomenon that triggers the scrubbing
               phrasal/expression verbs, overlap
      TYPE
        the kind of item to be scrubbed
               word, synset, feature set
      ACTION
        the action to be taken with the scrubbed item
               move, remove                                   19
  Voting
              0.3




             0.25




              0.2
Similarity




                                                                                                                                              Literal
             0.15
                                                                                                                                              Nonliteral




              0.1




             0.05




               0
                    Learner   Learner   Learner   Learner   Learner   Learner    Learner    Learner   Learner   Learner   Learner   Learner
                      A         B         C         D         A         B          C          D         A         B         C         D

                          girl brother mother hand             essenti institut financ philosophi      president kaisertech financ quandari     20
                                                                 Sentences and Learners
SuperTags and Context
   SuperTags
    A/B_Dnx person/A_NXN needs/B_nx0Vs1 discipline/A_NXN
       to/B_Vvx kick/B_nx0Vpls1 a/B_Dnx habit/A_NXN like/B_nxPnx
       drinking/A_Gnx0Vnx1 ./B_sPU
    →
    disciplin habit drink kick/B_nx0Vpls1_habit/A_NXN

   Context
    foot drag/A_Gnx0Vnx1_foot/A_NXN
    →
    foot everyon mcdonnel dougla commod anyon paul nisbet aerospac
       analyst prudentialbach secur mcdonnel propfan model spring
       count order delta drag/A_Gnx0Vnx1_foot/A_NXN
                                                                   21
Results – Evaluation Criteria 1
    25 target words
                                                    absorb      assault    die         drag    drown
                                    Lit Target       4             3      24           12        4
                                     Nonlit Target    62            0      11           41        1

   Target sets:                     Target
                                     Lit FB
                                     Nonlit FB
                                                      66
                                                     286
                                                      1
                                                                    3
                                                                  119
                                                                    0
                                                                           35
                                                                          315
                                                                            7
                                                                                        53
                                                                                       118
                                                                                       241
                                                                                                  5
                                                                                                 25
                                                                                                 21
    1 to 115 sentences each          Lit Target
                                                    escape
                                                      24
                                                               examine
                                                                   49
                                                                           fill
                                                                           47
                                                                                        fix
                                                                                        39
                                                                                                flow
                                                                                                 10

   Feedback sets:                   Nonlit Target
                                     Target
                                     Lit FB
                                                      39
                                                      63
                                                     124
                                                                   37
                                                                   86
                                                                  371
                                                                           40
                                                                           87
                                                                          244
                                                                                        16
                                                                                        55
                                                                                       953
                                                                                                 31
                                                                                                 41
                                                                                                 74
    1 to ~1500 sentences each        Nonlit FB        2
                                                     grab
                                                                    2
                                                                 grasp
                                                                           66
                                                                          kick
                                                                                       279
                                                                                      knock
                                                                                                  2
                                                                                                lend

   Total target sentences:          Lit Target
                                     Nonlit Target
                                     Target
                                                      5
                                                      13
                                                      18
                                                                    1
                                                                    4
                                                                    5
                                                                           10
                                                                           26
                                                                           36
                                                                                        11
                                                                                        29
                                                                                        40
                                                                                                 77
                                                                                                 15
                                                                                                 92
    1298                             Lit FB
                                     Nonlit FB
                                                      76
                                                      58
                                                                   36
                                                                    2
                                                                           19
                                                                          172
                                                                                        60
                                                                                       720
                                                                                                641
                                                                                                  1


   Total literal FB sentences:      Lit Target
                                     Nonlit Target
                                                     miss
                                                      58
                                                      40
                                                                 pass
                                                                    0
                                                                    1
                                                                          rest
                                                                            8
                                                                           20
                                                                                       ride
                                                                                        22
                                                                                        26
                                                                                                 roll
                                                                                                 25
                                                                                                 46
    7297                             Target
                                     Lit FB
                                                      98
                                                     236
                                                                    1
                                                                 1443
                                                                           28
                                                                           42
                                                                                        48
                                                                                       221
                                                                                                 71
                                                                                                132


   Total nonliteral FB sentences:   Nonlit FB

                                     Lit Target
                                                      13
                                                    smooth
                                                      0
                                                                  156
                                                                  step
                                                                   12
                                                                            6
                                                                          stick
                                                                            8
                                                                                         8
                                                                                      strike
                                                                                        51
                                                                                                 74
                                                                                               touch
                                                                                                 13
    3726                             Nonlit Target
                                     Target
                                                      11
                                                      11
                                                                   94
                                                                  106
                                                                           73
                                                                           81
                                                                                        64
                                                                                       115
                                                                                                 41
                                                                                                 54
                                     Lit FB           28            5     132          693      904
                                     Nonlit FB        75          517     546          351      406
                                     Totals: Target=1298; Lit Feedback=7297; Nonlit Feedback=3726


                                                                                                  22
    Results – Evaluation Criteria 2
   Target set sentences hand-annotated for testing
   Unknowns sent to cluster opposite to manual label
   Literal Recall = correct literals in literal cluster / total correct literals
   Literal Precision = correct literals in literal cluster
                          / size of literal cluster
   If no literals: Literal Recall = 100%;
                    Literal Precision = 100% if no nonliterals in literal cluster,
                                           else 0%
   f-score = (2*precision*recall) / (precision+recall)
   Nonliteral scores calculated in same way
   Overall Performance = f-score of averaged literal/nonliteral
    precision scores and averaged literal/nonliteral recall scores
                                                                              23
Baseline
   Baseline – Simple Attraction
       Target sentence attracted to feedback set
        containing sentence with which it has the most
        words in common
       Unknowns sent to cluster opposite to manual
        label
       Attempts to distinguish between literal and
        nonliteral
       Uses all data used by TroFi

                                                         24
                                                                                                                                                                                        f-score
                                                                                                                                                                                        f-score
                                                                                                                                                                                        f-score
                                                                                                                 aab
                                                                                                                  ab




                                                                                                                                       0
                                                                                                                                                   10
                                                                                                                                                                20
                                                                                                                                                                        30
                                                                                                                                                                                 40
                                                                                                                                                                                              50
                                                                                                                                                                                                           60
                                                                                                                                                                                                                70
                                                                                                                                                                                                                     80
                                                                                                                                                                                                                          90
                                                                                                                                                                                                                               100




                                                                                                                                       0
                                                                                                                                                   10
                                                                                                                                                                20
                                                                                                                                                                        30
                                                                                                                                                                                 40
                                                                                                                                                                                              50
                                                                                                                                                                                                           60
                                                                                                                                                                                                                70
                                                                                                                                                                                                                     80
                                                                                                                                                                                                                          90
                                                                                                                                                                                                                               100
                                                                                                                   bssso
                                                                                                                      o
                                                                                                                  as rrb
                                                                                                                aa s b
                                                                                                                  ssssa
                                                                                                                      aau
                                                                                                                        ull
                                                                                                                          tt
                                                                                                                        d
                                                                                                                       diie
                                                                                                                          e
                                                                                                                     d
                                                                                                                    ddrra
                                                                                                                      raa
                                                                                                                          g
                                                                                                                   d gg
                                                                                                                  ddrro
                                                                                                                    roow
                                                                                                                       ww




        Baseline
                                                                                                                  es n
                                                                                                                eesc nn




         Baseline
                                                                                                                  scca




           Baseline
                                                                                                                ex aape
                                                                                                               eexa pe
                                                                                                                xaam
                                                                                                                     m
                                                                                                                    mi
                                                                                                                        n




               Baseline
                                                                                                                       in e
                                                                                                                          e




           Learners Voting
                                                                                                                            ffiillll




         Learners & & Voting
                                                                                                                            ffiixx
                                                                                                                           llo
                                                                                                                      fflfoo
                                                                                                                               www
                                                                                                                      gr
                                                                                                                     gg ra
                                                                                                                        raab
                                                                                                                                bb
                                                                                                                    g
                                                                                                                   ggrra
                                                                                                                     raas
                                                                                                                             ssppp
                                                                                                                        ki
                                                                                                                      kick ic
                                                                                                                               ck
                                                                                                                    kn kk
                                                                                                                    kn
                                                                                                                   kn
                                                                                                                       ooco
                                                                                                                             cckkk




                                             Baseline
                                          SuperTags
                                          TroFi Base
                                          TroFi Base
                                                                                                                      lee
                                                                                                                     l len




                                             TroFi Base
                                                                                                                             nndd



                                                                                               Target Words
                                                                                                Target Words
                                                                                                Target Words
                                                                                               Target Words
                                                                                                                      m
                                                                                                                      m
                                                                                                                     mi
                                                                                                                              iss
                                                                                                                            isss
                                                                                                                                 s
                                                                                                                     pa
                                                                                                                    papa
                                                                                                                               ss
                                                                                                                             ssss
                                                                                                                          re
                                                                                                                      rrees
                                                                                                                               sstt
                                                                                                                         ririid
                                                                                                                      r dd
                                                                                                                                 e
                                                                                                                                ee
                                                                                                                         rrro
                                                                                                                ssm ollll
                                                                                                                 sm o
                                                                                                                  sm
                                                                                                                  mo
                                                                                                                      ooo
                                                                                                                        oottth
                                                                                                                            hh
                                                                                                                       s
                                                                                                                      ssttte
                                                                        Context




                                                                                                                          eepp
                                                                                                                             p
                                                                                                                       st
                                                                                                                      sstic
                                                                                                                        ticc
                                                                                                                           ik
                                                                                                                    st kk
                                                                                                                   sstrr
                                                                                                                      sr i
                                                                                                                      t i ik
                                                                                                                          kke
                                                                        Sum Similarities




                                                                                                                             e
                      Baseline TroFi Base SuperTags Sum of Similarities Sum of Similarities




                                                                                                                    ttto e
                                                                                                                   toou
                                                                       Sum ofof Similarities




                                                                                                                       uuc
                                                                TroFi Base Learners & Voting




                                                                                                                Av h
                                                                                                               AAve cch
                                                                                                                Av h
                                                                                                                 veerera
                                                                                                                      raag
                                                                                                                          ggee
                                                                                                                             e
25 25
                                                                                                                                                                                                   53.8%




                                                                                                                                                                     46.3%
                                                                                                                                                                                      48.9%
                                                                                                                                                                             48.4%




                                                                                                                                           29.4%
                                                                                                                                                        36.9%
TroFi Example Base –
Iterative Augmentation
   Purpose
       cluster more target sentences for given target word after
        initial run using knowledge gained during initial run
       improve accuracy over time
   Method
       use TroFi with Active Learning
       after each run, save weight of each feedback set sentence
       for each feedback set sentence,
        weight = highest similarity to any target sentence
       newly clustered sentences added to feedback sets with
        weight = 1
       in subsequent runs for same target word, use saved
        weighted feedback sets instead of building new ones

                                                                    26
                                                        absorb
                                                        assault
          100
                                                        attack
                                                        besiege
          90TroFi Example Base                          cool
          80                                            dance
                                                        destroy
          70
               Two runs, one regular, one iterative    die
                                                        dissolve    63.9%
          60
                augmentation, for 50 target words       drag
f-score




          50
                                                        drink
          40
               Uses optimal Active Learning model      drown
                                                        eat
          30
               Easy to expand clusters for current target
                                                        escape
          20
                words further using iterative augmentation
                                                        evaporate   26.6%

          10
                                                        examine

           0
               Also possible to add new target words, but
                                                        fill
                                                        fix




                 vaporize
                  dissolve
                 examine




                        plow
                   destroy



                evaporate




                        play
                     drown




                     wither
                   escape



                         flow




                        cool




                      sleep
                        lend




                  stumble
                       pass


                           roll
                           die




                      touch
                         ride
                    absorb




                   smooth
                        step




                 Average
                     dance




                     pump
                         rain
                       miss
                   assault




                            fix




                            fly


                       plant
                  besiege




                          eat




                           kill
                         rest




                        melt



                        pour
                            fill




                       flood
                        drag




                        grab




                        stick




                   flourish
                     grasp




                      strike

                     attack




                       drink




                     target
                         kick
                     knock




                requires new feedback sets              flood
                                                        flourish
                                       Target Words
                                                        flow
                                                        fly
                                High Baseline   TroFi
                                                        grab
                                                                    27
                                                        grasp
                                                        kick
TroFi Example Base
   Literal and nonliteral clusters of WSJ sentences
***pour***                          back to WSJ files
                                RefNonliteral label –
*nonliteral cluster*                    Literal label –
                                    eitherUnannotated –
                                             testing legacy
                                        either testing legacy
wsj04:7878 N As manufacturers get bigger , they are likely to pour more money
                                           from iterative
                                    or active learning augmentation run
                                        or active learning
                  into the battle for shelf space , raising the ante for new players ./.
wsj25:3283 N Salsa and rap music pour out of the windows ./.
wsj06:300 U Investors hungering for safety and high yields are pouring record
                  sums into single-premium , interest-earning annuities ./.
*literal cluster*
wsj59:3286 L Custom demands that cognac be poured from a freshly opened
                  bottle ./.


   Resource for nonliteral language research
                                                                                    28
Conclusion
   TroFi – a system for nonliteral language recognition
   TroFi Example Base – an expandable resource
    of literal/nonliteral usage examples for the nonliteral
    language research community
   Challenges
       Improve the algorithm for greater speed and accuracy
       Find ways of using TroFi and TroFi EB for interpretation

   TroFi – a first step towards an unsupervised,
    scalable, widely applicable approach to nonliteral
    language processing that works on real-world data
    for any domain in any language
                                                                   29
Questions?




             30
Extras
Not part of defense




                      31
32
The Long-Awaited Formulas
  affn(W, S) = maxWi  S simn(W, Wi)



  affn(S, W) = maxSj  W simn(S, Sj)



  simn+1(S1, S2) = W  S1 weight(W, S1) · affn(W, S2)



  simn+1(W1, W2) = S  W1 weight(S, W1) · affn(S, W2)



                                                         33
Types of Metaphor
   dead (fossilized)
     „the eye of a needle‟; „the are transplanting the
        community‟
   cliché
     „filthy lucre‟; „they left me high and dry‟; „we must
        leverage our assets‟
   standard (stock; idioms)
     „plant a kiss‟; „lose heart‟; „drown one‟s sorrows‟
   recent
     „kill a program‟; „he was head-hunted‟; „she‟s all that
        and a bag of chips‟; „spaghetti code‟
   original (creative)
     „A coil of cord, a colleen coy, a blush on a bush turned
        first men‟s laughter into wailful mother‟ (Joyce)
     „I ran a lawnmower over his flowering poetry‟           34
Conceptual Metaphor
     „in the course of my life‟
     „make a life‟
     „build a life‟
     „put together a life‟
     „shape a life‟
     „shatter a life‟
     „rebuild a future‟
                        (Lakoff & Johnson 1980)
                                                  35
Anatomy of a Metaphor

            metaphor
                           object
                                              target
   source         a sunny smile

                            sense (tenor)
     image (vehicle)
     = ‘sun’                = ‘cheerful’, ‘happy’,
                            ‘bright’, ‘warm’


                                                       36
Traditional Methods
     Metaphor Maps
         a type of semantic network linking sources to
          targets


     Metaphor Databases
         large collections of metaphors organized
          around sources, targets, and psychologically
          motivated categories


                                                      37
Metaphor Maps

            Action                        Event
   Actor
                       Kill
            Killing                    Death Event
                      Result                                     Patient

   Killer

                       Kill
                                              Dier
                      Victim




            Animate            Living-Thing



                                                Killing (Martin 1990)
                                                                           38
Metaphor Databases
 PROPERTIES ARE POSSESSIONS
 She has a pleasant disposition.
 CHANGE IS GETTING/LOSING
 CAUSATION IS CONTROL OVER AN OBJECT RELATIVE TO A POSSESSOR
 ATTRIBUTES ARE ENTITIES
 STATES ARE LOCATIONS and PROPERTIES ARE POSSESSION.
 STATES ARE LOCATIONS
 He is in love.
 What kind of a state was he in when you saw him?
 She can stay/remain silent for days.
 He is at rest/at play.
 He remained standing.
 He is at a certain stage in his studies.
 What state is the project in?
 It took him hours to reach a state of perfect concentation.
 STATES ARE SHAPES
 What shape is the car in?
 His prison stay failed to reform him.
 This metaphor may actually be more narrow:
 STATES THAT ARE IMPORTANT TO PURPOSES ARE SHAPES.
 Thus one can be 'fit for service' or 'in no shape to drive'
 It may not be a way to talk about states IN GENERAL.
 This metaphor is often used transitively with SHAPES ARE CONTAINERS.
 He doesn't fit in
 She's a square peg

                                                                        39
Attempts to Automate
   Using surrounding context to interpret
    metaphor
       James H. Martin and KODIAK
   Using word relationships to interpret
    metaphor
       William B. Dolan and the LKB




                                             40
Metaphor Interpretation as an
Example-based System

                    kick the bucket
                    bite the dust     ins Grass
                    pass on            beissen
                    croak             entweichen
                    cross over to     hinueber treten
  kick the bucket                                       ins Grass beissen
                     the other side   dem Jenseits
                    go the way of      entgegentreten
                     the dodo         abkratzen
                    die               sterben
                    decease
                    perish


                                                                     41
Word-Sense Disambiguation
#1
  An unsupervised bootstrapping algorithm for word-
    sense disambiguation (Yarowsky 1995)
       Start with set of seed collocations for each sense
       Tag sentences accordingly
       Train supervised decision list learner on the tagged set
        -- learn additional collocations
       Retag corpus with above learner; add any tagged
        sentences to the training set
       Add extra examples according to „one sense per
        discourse constraint‟
       Repeat

                                                             42
Problems with Algorithm #1
                                    Hard to define for Metaphor
     Need clearly defined           vs Literal
      collocation seed sets         Difficult to determine what
     Need to be able to extract     those features should be
      other features from            since many metaphors are
      training examples              unique
                                    People will often mix literal
     Need to be able to trust       and metaphorical uses of a
      the one sense per              word
      discourse constraint



                                                                 43
Similarity-based Word-Sense
Disambiguation

     Uses machine-readable dictionary definitions as
      input
     Creates clusters of similar contexts for each
      sense using iterative similarity calculations
     Disambiguates according to the level of attraction
      shown by a new sentence containing the target
      word to a given sense cluster


                                                     44

								
To top