Argument coercion (a type of metonymy) occurs when the expected semantic class (relative to the a predicate) is substituted for an object of a dif­ferent semantic class. Metonymy is a pervasive phenomenon in language and the interpretation of métonymie expressions can impact tasks from se­mantic parsing (Scheffczyk et al, 2006) to ques­tion answering (Harabagiu et al, 2005). A seminal example in metonymy from (Lakoff and Johnson, 1980) is:

(1) The ham sandwich is waiting for his check.

The Arg 1 for the predicate wait is typically an animate, but the "ham sandwich" is clearly not an animate. Rather, the argument is coerced to ful­fill the predicate's typing requirement. This coer­cion is allowed because an object that would nor­mally fulfill the typing requirement (the customer) can be uniquely identified by an attribute (the ham sandwich he ordered).

SemEval-2010 Task 7 ("Argument Selection and Coercion") (Pustejovsky and Rumshisky, 2009) was designed to evaluate systems that de­tect such coercions and provide a "compositional history" of argument selection relative to the pred­icate. In order to accomplish this, an argument is annotated with both the semantic class to which it belongs (the "source" type) as well as the class ex­pected by the predicate (the "target" type). How­ever, in the data provided, the target type was un­ambiguous given the lemmatized predicate, so the remainder of this paper discusses source type clas­sification. The detection of coercion is then sim­ply performed by checking if the classified source type and target type are different.

In our system, we explore two approaches with separate underlying assumptions about how arbi­trary semantic classes can be learned. In our first approach, we assume a semantic class can be de­fined a priori from a set of seed terms and that WordNet is capable of defining the membership ofthat semantic class. We apply the PageRank al­gorithm in order to weight WordNet synsets given a set of seed concepts. In our second approach, we assume that arguments in the same semantic class will be selected by similar verbs. We apply a statistical test to determine the most representative predicates for an argument. This approach benefits from a large corpus from which we automatically extracted 200 million predicate-argument pairs.

The remainder of this paper is organized as fol­lows. Section 2 discusses our WordNet-based ap­proach. Section 3 describes our corpus approach. Section 4 discusses our experiments and results. Section 5 provides a conclusion and direction for future work. Due to space limitations, previous work is discussed when relevant.

Our first approach assumes that semantic class members can be defined and acquired a priori.

Our second approach assumes that semantic class members are arguments of similar predicates. As (Pustejovsky and Rumshisky, 2009) elaborate, predicates select an argument from a specific se­mantic class, therefore terms that belong in the same semantic class should be selected by simi­lar predicates. However, this assumption is often violated: type coercion allows predicates to have arguments outside their intended semantic class. Our solution to this problem, partially inspired by (Lapata and Lascarides, 2003), is to collect statis­tics from an enormous amount of data in order to statistically filter out these coercions.

The English Gigaword Forth Edition corpuscontains over 8.5 million documents of newswire text collected over a 15 year period. We processed these documents with the SENNA (Collobert and Weston, 2009) suite of natural language tools, which includes a part-of-speech tagger, phrase chunker, named entity recognizer, and PropBank semantic role labeler. We chose SENNA due to its speed, yet it still performs comparably with many state-of-the-art systems. Of the 8.5 million doc­uments in English Gigaword, 8 million were suc­cessfully processed. For each predicate-argument pair in these documents, we gathered counts by argument type and argument head. The head was determined with simple heuristics from the chunk parse and parts-of-speech for each argument (ar­guments consisting of more than three phrase chunks were discarded). When available, named entity types (e.g., Person, Organization, Location) were substituted for heads. This resulted in over 511 million predicate-argument pairs for argument types ArgO, ArgI, and Arg2. For this task, however, we chose only to use ArgI argu­ments (direct objects), which resulted in 210 mil­lion pairs, 7.65 million of which were unique. The Arg 1 argument was chosen because most of the arguments in the data are direct objects .

The "best" predicates for a given argument are defined by a ranking based on Fisher's exact test (Fisher, 1922):

where a is the number of times the given argument was used with the given predicate, b is the number of times the argument was used with a different predicate, c is the number of times the predicate was used with a different argument, d is the num­ber of times neither the given argument or predi­cate was used, and n = a+b+c+d. The top ranked (lowest p) predicates for the most common argu­ments in the training data are shown in Table 2.

We have conducted several experiments to test the performance of the approaches outlined in Sections 2 and 3 along with additional features commonly found in information extraction liter­ature. All experiments were conducted using the $-yy[muiticiass SUpp0rt vector machine library.

We have presented two approaches for determin­ing the semantic class of a predicate's argument. The two approaches capture different information and combine well to classify the "source" type in SemEval-2010 Task 7. We showed how this can be incorporated into a system to detect coercions as well as the argument's compositional history rel­ative to its predicate. In future work we plan to extend this system to more complex tasks such as when the predicate may be polysemous or unseen predicates may be encountered.

The authors would like to thank Bryan Rink for several insights during the course of this work.