The WSD module is based on a simple heuristic rule by which we can extract sense-labeled instances automatically. The heuristic rule assumes that one sense per 3-gram which is proposed by us initially through investigating a Chinese sense-tagged corpus STC (Wu et al., 2006). The assumption is similar to the celebrated one sense per collocation supposition (Yarowsky, 1993), whereas ours has more expansibility. STC is an ongoing project which is to build a sense-tagged corpus containing sense-tagged 1, 2 and 3 months of People's Daily 2000 now. According to our investigation, given a specific 3-gram (w_iWverbWi) to any target verb, on average, we expect to see the same label 95.4% of the time. Based on this observation, we consider one sense per 3-gram (w_iwverbwi) or at least we can extract instances with this pattern.

For all the 27 multiple-sense target verbs in the official training data, we found their 3-gram (w_iwverbwi) and extracted the instances with the same 3-gram from a Chinese monolingual corpus - the 2001 People's Daily (about 116M bytes). We consider the same 3-gram instances should have the same label. Then an additional sense­labeled training corpus is built automatically in expectation of having 95.4% precision at most. And this corpus has 2145 instances in total (official training data have 4608 instances).

We build four systems to investigate the effect of our instances expansion using the Naïve Bayesian classifier. System configuration is shown in Table 1. In column 1, BL means baseline, X means instance expansion, 3 and 15 means the window size. In column 2, wi is the i­th word relative to the target word, wi-1wi is the 2-gram of words, w/j is the word with position information (je[-3,+3]). In the last column, 'O' means using only the original training data and 'O+A' means using both the original and additional training data. Syntactic feature and parameter optimizing are not used in this module.

We use a feature-based machine learning method to implement the SRL module in which three tags are labeled, namely the semantic role tag, the event role tag and the phrase type tag. We consider the SRL task as a four-step pipeline: (1) parsing which generates a constituent parse tree for the input sentence; (2) pruning which filters out many apparently impossible constituents (Xue and Palmer, 2004); (3) semantic role identification (SRI) which identifies the constituent that will be the semantic role of a predicate in a sentence, and (4) semantic role classification (SRC) which determines the type of identified semantic role. The machine learning method takes PAS as the classification unit which consists of a target predicate and an argument candidate. The SRI step utilizes a binary classifier to determine whether the argument candidate in the PAS is a real argument. Finally, in the SRC step, the semantic role tag and the event role tag of each identified argument can be obtained by two multi-value classifications on the SRI results. The remaining phrase type tag can be directly extracted from the constituent parsing tree.

The selection of the feature set is the most important factor for the feature-based SRL method. In addition to constituent-based features and dependency-based features, we also consider WSD-based features. To our knowledge, the combined use of constituents-based syntactic features and dependency-based syntactic features is the first attempts to use them both on the feature level of SRL. As a prevalent kind of syntactic features for SRL, constituent-based features have been extensively studied by many researchers. In this module, we use 34 constituent-based features, 35 dependency-based features, and 2 WSD-based features. Among the constituent-based features, 26 features are manually selected from effective features proven by existing SRL studies and 8 new features are defined by us. Firstly, the 26 constituent-based features used by others are:

• predicate (c1), path (c2), phrase type (c3), position (c4), voice (c5), head word (c6), predicate subcategorization (c7), syntactic frame (c8), head word POS (c9), partial path (c10), first/last word (c11/c12), first/last POS (c13/c14), left/right sibling type (c15/c16), left/right sibling head (c17/c18), left/right sibling POS (c19/c20), constituent tree distance (c21), temporal cue words (c22), Predicate POS (c23), argument's parent type(c24), argument's parent head (c25) and argument's parent POS (c26).

And the 8 new features we define are:

• Locational cue words (c27): a binary feature indicating whether the constituent contains location cue word.

• POS pattern of argument (c28): the left-to-right chain of POS tags of argument's children.

• Phrase type pattern of argument (c29): the left-to-right chain of phrase type labels of argument's children.

• Type of LCA and left child (c30): The phrase type of the Lowest Common Ancestor (LCA) combined with its left child.

• Type of LCA and right child (c31): The phrase type of the LCA combined with its right child.

• Three features: word bag of path (c32), word bag of POS pattern (c33) and word bag of type pattern (c34), for generalizing three sparse features: path (c7), POS pattern argument (c28) and phrase type pattern of argument (c29) by the bag-of-words representation.

Secondly, the selection of dependency-based features is similar to that of constituent-based features. But dependency parsing lacks constituent information. If we want to use dependency-based features to label constituents, we should map a constituent to one or more appropriate words in dependency trees. Here we use head word of a constituent to represent it in dependency parses. The 35 dependency-based features we adopt are:

• Predicate/Argument relation (d1/d2), relation path (d3), POS pattern of predicate's children (d4), relation pattern of predicate's children (d5) , child relation set (d6), child POS set (d7), predicate/argument parent word (d8/d9), predicate/argument parent POS (d10/d11), left/right word (d12/d13), left/right POS

left/right sibling word (d18/d19), left/right sibling POS (d20/d21), left/right sibling relation (d22/d23), dep-exists (d24) and dep- (d27), relation path length (d28), high/low support verb (d29/d30), high/low support noun (d31/d32) and LCA's word/POS/relation (d33/d34/d35).

In this work, the dependency parse trees are generated from the constituent parse trees using a constituent-to-dependency converter (Marneffe et al., 2006). The converter is suitable for semantic analysis as it can retrieve the semantic head rather than the general syntactic head.

Lastly, the 2 WSD-based features are:

• Situation description formula (s1): predicate's situation description formula generated by the

WSD module.

• Natural explanation (s2): predicate's natural explanation generated by the WSD module.

Table 2 shows the official result of the WSD system. BL_3 with window size three using the original training corpus achieves the best result in our submission. It indicates the local features are more effective in our systems. There are two possible reasons why the performances of the X system with instance expansion are lower than the BL system. First, the additional instances extracted based on 3-gram provide a few local features but many topical features. But, local features are more effective for our systems as mentioned above. The local feature related information that the classifier gets from the additional instances is not sufficient. Second, the granularity of the WSD module is too small to be distinguished by 3-grams. As a result, the additional corpus built upon 3-gram has more exceptional instances (noises), and therefore it impairs the performance of X_3 and X_15. Taking the verb ' If ' (belong to ) as an example, it has two senses in the task, but both senses have the same natural explanation: '!Ê|— Ä7:M^ÄÄ7:^W (part of or belong to), which is always considered as the sense in general SRL. The difference between the two senses is in their situation description formulas: 'partof (x,y)+NULL' vs. 'belongto (x,y)+NULL'.

In the SRL module, we use the training data provided by SemEval-2010 to train the SVM classifiers without any external resources. The training data contain 4,608 sentences, 100 target predicates and 13,926 arguments. We use the SVM-Light Toolkit (Joachims, 1999) for the implementation of SVM, and use the Stanford Parser (Levy and Manning, 2003) as the parser and the constituent-to-dependency converter. We employ the linear kernel for SVM and set the regularization parameter to the default value which is the reciprocal of the average Euclidean norm of the training data. The evaluation results of our SRL module on the official test data are shown in Table 3, where 'AB', 'SR', 'PT' and 'ER' represent argument boundary, semantic role tag, phrase type tag, and event role tag.

It is clear that 'AB' plays an important role as the labeling of the other three tags is directly based on it. Through analyzing the results, we find that errors in the recognition of 'AB' are mainly caused by two factors: the automatic constituent parsing and the pruning algorithm. It is inevitable that some constituents and hierarchical relations are misidentified in automatic parsing of Chinese. These errors are further enlarged by the heuristic-based pruning algorithm because the algorithm is built upon the gold-standard paring trees, and therefore a lot of real arguments are pruned out when using the noisy automatic parses. So the pruning algorithm is the current bottleneck of SRL in the evaluation.

From the fact that the results of 'AB+SR' and 'AB+ER' are close to that of 'AB', it can be inferred that the SR and ER results should be satisfactory if the errors in 'AB' are not propagated. Furthermore, the result of 'AB+PT' is low as the phrase types here is inconsistent with those in Stanford Parser. The problem should be improved by a set of mapping rules.

Finally, in the ED module, we combine the results of WSD and SRL by filling variables of the situation description formula obtained by the WSD module with the arguments obtained by the SRL module according to their event role tags. Table 4 shows the final results which are generated by combining the results of WSD and SRL. Obviously the reduced overall ranking comparing to WSD is due to the SRL module.