This paper presents a temporal relation identifier by the team NAIST.Japan. Our identifier has two char-actaristics: sequence labeling model and use of de­pendency parsed tree.

Firstly, we treated each problem a sequence la­beling problem, such that event/time pairs were or­dered by the position of the events and times in the document. This idea is for task B and C. In task B, the neighbouring relations between an EVENT and DCT-TIMEX3 tend to interact. In task C, when EVENT-a, EVENT-b, and EVENT-c are linearly or­dered, the relation between EVENT-a and EVENT-b tends to affect the one between EVENT-b and

EVENT-c.

Secondly, we introduced dependency features where each word was annotated with a label indi­cating its tree position to the event and the time, e.g. "descendant" of the event and "ancestor" of the time.

The dependency features are introduced for our ma­chine learning-based relation identifier. In task A, we need to label several different event-time pairs within the same sentence. We can use information from TIMEX3, which is a descendent of the target EVENT in the dependency tree.

Section 2 shows how to use a sequence labeling model for the task. Section 3 shows how to use the dependency parsed tree for the model. Section 4 presents the results and discussions.

Our approach to identify temporal relation is based on a sequence labeling model. The target pairs are linearly ordered in the texts.

Sequence labeling model can be defined as a method to estimate an optimal label sequence

over an observed sequence {xi, xiixn). We consider, w-parameterized function

Here, denotes all possible label combinations over y; <!>(x,y) denotes a feature expression over x, y. Introducing a kernel function:

we have a dual representation:

A dependency relation is a head-modifier relation on a syntactic tree. Figure 4 shows an example de­pendency parsed tree of the following sentence -"The warremts may be exercised uritil 90 days after their issue date". We parsed the TimeEval data us­ing MSTParser v0.2 (McDonald and Pereira, 2006), which is trained with all Penn Treebank (Marcus et al., 1993) without dependency label.

We introduce tree position labels between an tar­get node and another node on the dependency parsed tree: ANC (ancestor), DES (descendant), SIB (sib­ling), and TARGET (target word). Figure 5 shows the labels, in which the box with double lines is the target node. The tree position between the target EVENT and a word in the target TIMEX3 is used as a feature for our machine learning-based relation identifier.

We also use the words in the sentence including the target entities as features. Each word is anno-

TARGET node: "exercised^ TARGET nodes: "90" and "days" TARGET-A node: "exercised^

tated with (1) its tree position to the EVENT, (2) its tree position to the TIMEX3, and (3) the com­bination of the labels from (1) and (2). Fig. 6 shows the labels of tree positions. The left picture shows (1) EVENT-based labels of the tree position with the target EVENT "exercised". The center pic­ture shows (2) TIMEX3-based ones with the target

TIMEX3 "90 days". The right picture shows (3)

JOINT ones which are combinations of the relation label with the EVENT and with the TIMEX3. We perform feature selection on the words in the cur­rent sentence according to the tree position labels. Note that, when MSTparser outputs more than one trees for a sentence, we introduce a meta-root node to bundle the ones in a tree.

We use HMjVLSVM !as a sequence labeling model with features in Table 1, 2 and 3 for task A, B and C, respectively. The attributes value in TIMEX3 is encoded as the relation with DCT-TIMEX3: {BEFORE, OVERLAP, AFTER, VAGUE}.In task A, only words in the current sentence with JOINT relation labels "TARGET/*" or "ANC/*"or "*/DES" were used. In task C, attributes in the TIMEX3 are annotated with the flag whether the TIMEX3 entity is the highest (namely the nearest to the root node) in the tree. Some adverbs and con­junctions in the succeeding sentence help to deter­mine the adjacent two relations. Thus, we introduce all words in the succeeding sentence for Task A and B. These features are determined by our preliminary experiments with the trial data .

Table 4 is our results on the test data. Whereas, our system is average rank in task A and B, it is worst mark in task C. The features from dependency parsed trees are effective for task A and B. However, these are not for task C.

Now, we focus on what went wrong instead of what went right in our preliminary experiments in trial data. We tried point-wise methods with other _Table 1 : Features for Task A_

all attributes in the target EVENT all attributes in the target TIMEX3 —the attributes value is encoded as the relation with DCT-TIMEX3_ all words in the current sentence with TIMEX3-based label (2) of tree position words in the current sentence with JOINT label (3) of tree position — only relation label with "TARGET/*" or "ANC/*"or label (1) of tree position from the EVENT to the all words in the succeeding sentence

_Table 2: Features for Task B_

all attributes in the target EVENT all attributes in the target TIMEX3 of in the current sen­tence with EVENT-based label (1) of tree position all attributes in the target TIMEX3 of in the preceding and succeeding sentence all words in the current sentence with EVENT-based la­bel (1) of tree position all words in the succeeding sentence

_Table 3 : Features for Task C_

all attributes in the target two EVENTs (EVENT-1 and all attributes in the TIMEX3 in the sentence including EVENT-1 with the label (1) of tree position to EVENTall attributes in the TIMEX3 in the sentence including EVENT-2 with the label (1) of tree position to EVENTall words in the sentence including EVENT-1 with the label (1) of tree position to EVENT-1 all words in the sentence including EVENT-2 with the label (1) of tree position to EVENT-2

machine learners such as maximum entropy and multi-class support vector machines. However, se­quence labeling method with HMjVLSVM outper­formed other point-wise methods in the trial data.

We have dependency parsed trees of the sen­tences. Naturally, it would be effective to intro­duce point-wise tree-based classifiers such as Tree Kernels in SVM (Collins and Duffy, 2002; Vish-wanathan and Smola, 2002) and boosting for clas­sification of trees (Kudo and Matsumoto, 2004). We tried a boosting learner which enables us to perform subtree feature selection for the tasks. However, the boosting learner selected only one-node subtrees as useful features. Thus, we perform simple vector-based feature engineering on HMjVLSVM.

We believe that it is necessary for solving task C to incorporate knowledge of verb-verb relation. We also tried to use features in verb ontology such as VERBOCEAN (Chklovsky and Pantel, 2004) which is used in (Mani et al., 2006). It did not improved performance in our preliminary experiments with trial data.