computing word-pair antonymy *saif mohammad *bonnie dorr φ graeme hirst *univ. of maryland φ univ....
TRANSCRIPT
Computing Word-Pair Antonymy
*Saif Mohammad*Bonnie Dorr
φGraeme Hirst
*Univ. of MarylandφUniv. of Toronto
EMNLP 2008
Introduction
• Antonymy: pair of semantically contrasting words.
• Ex: Strongly antonymous: HotCold
Semantically contrasting: EnemyFanNot antonymous: PenguinClown
Usage
• Detecting contradictions• Detecting humor• Automatic creation of thesaurus
Problem Definition
• Given a thesaurus, find out the antonymous category pairs.
• Assign the degree of antonymy to each pair of antonymous categories.
Hypothesis(1)
• The Co-occurrence Hypothesis of Antonyms– Antonymous word pairs occur together much
more often than other word pairs.
Hypothesis(1)
• Empirical proof:– 1,000 antonymous pairs from Wordnet– 1,000 randomly generated word pairs– Use BNC as corpus, set window size 5.– Calculate the MI for each word pairs and average
itAverage Standard deviation
Antonymous pair 0.94 2.27
Random pair 0.01 0.37
Hypothesis(2)
• The Distributional Hypothesis of Antonyms– Antonyms occur in similar contexts more often
than non-antonymous words– Ex work: activity of doing job
play: activity of relaxation
Hypothesis(2)
• Empirical proof:– Use the same set of word pairs in hypothesis(1)– Calculate the distributional distance between their
categories
Average Standard deviation
Antonymous pair 0.30 0.23
Random pair 0.23 0.11
Distributional Distancebetween Two Thesaurus Categories
c1,c2: thesaurus categoryI(x,y):pointwise mutual information between x and yT(c):the set of all words w such that I(c,w)>0
Method
• Determine pairs of thesaurus categories that are contrasting in meaning
• Use the co-occurrence and distributional hypotheses to determine the degree of antonymy of word pairs
Method•16 affix rules were applied to Macquarie Thesaurus •2,734 word pairs were generated as a seed set.
•Exceptions: sectXinsect• Relatively few
Method
• 10,807 pairs of semantically contrasting word pairs from WordNet
Method
• If any word in thesaurus category C1 is antonymous to any word in category C2 as per a seed antonym pair, then the two categories are marked as contrasting.
• If no word in C1 is antonymous to any word in C2, then the categories are considered not contrasting
Method
• Degree of antonymy----category level– By distributional hypothesis of antonyms, we
claim that the degree of antonymy between two contrasting thesaurus categories is directly proportional to the distributional closeness of the two concepts
Method
• Degree of antonymy----word level– target words belong to the same thesaurus
paragraphs as any of the seed antonyms linking the two contrasting categories highly antonymous
– target words do not both belong to the same paragraphs as a seed antonym pair, but occur in contrasting categories medium antonymous
– target words with low tendency to co-occur lowly antonymous
Method
• Adjacency Heuristic– Most thesauri are ordered such that contrasting
categories tend to be adjacent
Evaluation
• 1,112 Closest-opposite questions designed to prepare students for GRE(Graduate Record Examination)– 162 questions as the development set– 950 questions as the test set
Evaluation
• Closest-opposite questions– Ex:
adulterate: a. renounce b. forbid c. purify d. criticize e. correct
Evaluation
• Closest-opposite questions– Ex:
adulterate: a. renounce b. forbid c. purify d. criticize e. correct
摻雜的
純淨的 批評
正確
禁止聲明放棄
Evaluation
Discussion
• The automatic approach does indeed mimic human intuitions of antonymy.
• In languages without a wordnet, substantial accuracies may be achieved.
• Wordnet and affix-generated seed are complementary.
Conclusion
• Proposed an empirical approach to antonymy that combines corpus co-occurrence statistics with the structure of a thesaurus.
• The system can identify the degree of antonymy between word pairs.
• An empirical proof that antonym pairs tend to be used in similar contexts.
Thanks