1、Lexical Acquisition,Fu lei 2007-05-14,Outline,Introduction Evaluation measures Verb subcategorization Attachment ambiguity Selectional preferences Semantic similarity between words Significance and further reading,Outline,Introduction Evaluation measures Verb subcategorization Attachment ambiguity S

2、electional preferences Semantic similarity between words Significance and further reading,Introduction,General goal To develop algorithms and statistical techniques for filling the holes in existing machine-readable dictionaries by looking at the occurrence patterns of words in large text corpora.,I

3、ntroduction,Lexical acquisition problems Collocations Verb subcategorization The recipient of contribute is expressed as a prepositional phrase with to Attachment ambiguity The children ate the cake with their hands The children ate the cake with blue icing Semantic categorization What is the semant

4、ic category of a new word that is not covered in our dictionary Selectional preference The verb eat usually takes food items as direct objects,Outline,Introduction Evaluation measures Verb subcategorization Attachment ambiguity Selectional preferences Semantic similarity between words Significance a

5、nd further reading,Evaluation measures,Evaluation measures Precision Recall F-measure Fallout,Evaluation measures,Target=tp+fn Selected=tp+fp Total=tn+fp+tp+fn,Evaluation measures,Precision=tp/selected=tp/(tp+fp) Measure of the proportion of selected items that the system got right Recall=tp/target=

6、tp/(tp+fn) Measure of the proportion of the target items that the system selected F=1/(/P+(1-)/R) If set =0.5,F=2PR/(P+R) Fallout=fp/(fp+tn) Measure of the proportion of non-targeted items that were mistakenly selected,Evaluation measures,Acc=(tp+tn)/total,Outline,Introduction Evaluation measures Ve

7、rb subcategorization Attachment ambiguity Selectional preferences Semantic similarity between words Significance and further reading,Verb subcategorization,Definition We refer to the classification of verbs according to the types of complements they permit as subcategorization. we say that a verb su

8、bcategorizes for a particular complement.,Verb subcategorization,Subcategorization frame A particular set of arguments that a verb can appear with is referred to as a subcategorization frame Example,Verb subcategorization,Why called subcategorization We can think of the verbs with a particular set o

9、f semantic arguments as one category. Each such category has several subcategories that express these semantic arguments using different syntactic means. Example The class of verbs with semantic arguments theme and recipient has a subcategory that expresses these arguments with an object and a prepo

10、sitional phrase (for example, donate in He donated a large sum of money to the church), and another subcategory that in addition permits a double-object construction (for example, give in He gave the church a large sum of money),Verb subcategorization,Importance for parsing example,Verb subcategoriz

11、ation,Algorithm for learning subcategorization frames Lerner by Brent(1993) Two steps Cues Hypothesis testing,Verb subcategorization,Example greet-V Peter-CAP,-PUNC I came-V Thursday-CAP,-PUNC before the storm started,Verb subcategorization,Outline,Introduction Evaluation measures Verb subcategoriza

12、tion Attachment ambiguity Selectional preferences Semantic similarity between words Significance and further reading,Attachment ambiguity,Mainly discussed PP attachment Example The children ate the cake with a spoon,Attachment ambiguity,Model A simple model based on co-occurrence statistics Probabil

13、istic model,Attachment ambiguity,Probabilistic model By Hindle and Rooth(1993) Model introduction,Attachment ambiguity,example He put the book on World War II on the table Derivation,Attachment ambiguity,Derivation,Attachment ambiguity,Attachment ambiguity,Attachment ambiguity,Problems 該模型建模時只考慮了v,n

14、,pp，然而現(xiàn)實句子中其它的信息可能對確定pp的依附性也有很大的作用，例如hindle和rooth發(fā)現(xiàn)在名詞前出現(xiàn)形容詞最高級就很可能是名詞短語依附 該模型僅考慮了緊跟在np后面的pp是修飾緊在其前的np的還是vp的這一基本情況，實際上pp的依附性有很多種情況。例如：pp對復合名詞的依賴（door bell manufacturer）,Outline,Introduction Evaluation measures Verb subcategorization Attachment ambiguity Selectional preferences Semantic similarity be

15、tween words Significance and further reading,Selectional preferences,Definition Most verbs prefer arguments of a particular type. such regularities are called selectional preferences or selectional restrictions. Importance of the acquisition of SP Infer Susan had never eaten a fresh durian before Ra

16、nking the possible parses,Selectional preferences,Model By Resnik(1993,1996) In principle, the model can be applied to any class of words that imposes semantic constraints on a grammatically dependent phrase: verbsubject, verbdirect object, verbpp, adj.noun etc. Here, only consider the case verb-dir

17、ect object,Selectional preferences,Model Selectional preference strength (SPS): Measures how strongly the verb constrains its direct object. Selectional association The proportion that its summand contributes to the overall preference strength,Selectional preferences,Model How to estimate P(c|v),Sel

18、ectional preferences,SPS Example,Selectional preferences,Association Example,Outline,Introduction Evaluation measures Verb subcategorization Attachment ambiguity Selectional preferences Semantic similarity between words Significance and further reading,Semantic similarity,The holy grail of lexical a

19、cquisition is the acquisition of meaning. Semantic similarity Judgments of semantic similarity can be explained by the degree of contextual interchangeability or which on word can be substituted for another in context (Miller and Charles 1991),Semantic similarity,Application Generalization Under the

20、 assumption that semantically similar words behave similarly Susan had never eaten a fresh durian before. Query expansion KNN classification,Semantic similarity,Computing methods Vector space measures Probabilistic measures,Semantic similarity,Vector space measures Binary valued Matching coefficient

21、 Dice coefficient Jaccard (or Tanimoto) coefficient Overlap coefficient Cosine Real valued Cosine,Semantic similarity,Definition of similarity measures for binary vectors,Semantic similarity,Comparison Matching coefficient Simply counts the number of dimensions on which both vectors are non-zero. Di

22、ce coefficient Normalizes for length by dividing by the total number of non-zero entries. Range from 0 to 1 Jaccard coefficient Penalizes a small number of shared entries more than the Dice coefficient.,Semantic similarity,Comparison Overlap coefficient Has the flavor of a measure of inclusion. Cosi

23、ne Identical to the Dice coefficient for vectors with the same number of non-zero entries, but it penalizes less in cases where the number of non-zero entries is very different.,Semantic similarity,Real valued,Semantic similarity,Real valued cosine For normalized vector, cosine gives the same ranking of similarities as Euclidean distance does,Semantic similarity,Advantage of vector measures Simple representation Easy to compute Example,Semantic similarity,Probabilistic measures Why introduce this method vector space based measures is that, excep