Trigger Detection

A trigger is any word in a sentence that serves as an indicator to an event. Due to the nature of event annotation of the biological literature, the trigger detection cannot be simply reduced to a dictionary lookup for trigger expression for certain reasons. First, a number of common textual expressions act as event triggers in some cases, but fail to

12

act as triggers in other cases. Second, a single expression may be associated with various event classes. For example, the instances of the token “over expression”, are evenly distributed among “positive regulation,” “transcription” and the negative classes. In light of these challenges, we approach trigger detection with a multi-class support vector machine (SVMLight multiclass) classifier that assigns event classes to individual tokens, one at a given time instance.⁷ This is in contrast to the sequence-modeling problem, which applies a sequential model.

The study implements the trigger detection as a token labeling problem, by assigning an event class to tokens or a negative class if it cannot be assigned to any event class. This implies that for a given token, if it is assigned to a class, it is therefore a trigger; else, it is a non-trigger. In some instances, adjacent tokens belong to the same class. We will assign a single trigger to such adjacent tokens if and only if one of the token occurs in the training data. It is observe that different event classes could share the same token as a trigger. Therefore, such token belongs to more than one event classes. Hence, the study uses the multi-class classification approach for our trigger detection. We use the combined class to assign detected triggers to a single event class (McClosky et al., 2012). For instance, “transcription/regulation” denotes a token that acts as a trigger to two events of the mentioned classes. This implies that, our trigger detection produces a single event node for each detected trigger.

The classifier is trained on gold standard triggers from the training data and incorporates several arrays of the token to be classified. Table 1 below shows features used in the trigger detector.

7 http://www.cs.cornell.edu/people/tj/svm_light/svm_multiclass.html

13

Table 1. Features of trigger candidate for trigger detection

Type Features

Token Token features are listed in Table 2 An example is illustrated in Table 3 Candidate in the parser

output

Token features of word with dependencies from candidate

Bigrams of dependencies

Bigram and Trigram of words (base form + POS) An example is illustrated in Table 4

Shortest path Shortest path features listed in Table 5 between candidate and their closest entity

Length of the path between candidates and entities An example is illustrated in Table 6

Token features include features of the individual token. They are testing for token capitalization, testing for punctuation and numeric characters, testing for stemmed tokens and token frequency. In trigger detection, the state (i.e., presence or absence) of the token in known trigger expressions and their classes extracted from the training corpus are used. In general, Table 2 shows our token features. Table 3 shows the result of token features for “IL-4 gene regulation involves NFAT1 and NFAT2”.

Table 2. List of token features

 Token is an upper case form

 Token is in our trigger dictionary

 Token has a numeric character

 Token is an entity

 Token is after “_”

 Token is a stemmed word

 Token frequency

14

Table 3. Token features for “IL-4 gene regulation involves NFAT1 and NFAT2”

Token Features

Table 4 gives the example of candidate in the parser output. In Table 4, the token features are the same features described in Table 2. Column “Bigram dependency” is the sum of the bigrams of candidate token and all its dependency tokens in the parse tree normalized by the total token size. Column “Bigram and trigram of words” is the sum of the bigram and trigram of tokens in the input data normalized by the total token size.

Bigram dependency Bigram and trigram of words Dictionary Frequency

regulation 1 1 0.356 0.524

involves 1 1 0.271 0.445

Dependency chains are built up to the depth of three starting from the token word to be classified. The study uses a set of features for finding the shortest path and

15

builds a dependency chain considering the three tokens with the shortest path from our candidate token. The sentences are parsed using a dependency parser (Stanford full parser) to identify argument-predicate roles. Words in the sentence and the relationships between these words form the dependency parse tree of the sentence. We use typed-dependency representation to represent our dependency as a simple tuple.

For example:

reln(gov, dep)

where reln is the dependency relation, gov is the governor word and the dep is the dependent word. Table 5 shows the features for our shortest path. Table 6 shows the shortest path features for “IL-4 gene regulation involves NFAT1 and NFAT2”.

Table 5. Features for the shortest path between candidate trigger and its

Table 6. Shortest path features for “IL-4 gene regulation involves NFAT1 and NFAT2”

16

In Table 6, the vertex walk is the sequence of nodes from the candidate node and their substructures. From Figure 2, the vertex walk from the candidate node

“regulation” has a score of four (three “theme” and one “cause” respectively).

Similarly, the edge walk of the shortest path features accounts for the “theme” and

“cause” pair attributed to each candidate mode. As shown in Figure 2, the candidate node “regulation” three pairs attributed to it, and the candidate node “involves” has a single pair attributed to it. The N-gram of dependency and N-gram of words are the sums of bigram and trigram of the candidate token and its dependency words in the dependency tree and the tokens in the input sentence normalized by the token size in the background data.