Relation structure of sample (S4). 

Contexts in source publication

Context 1

... such favorable natural environment, man-made disasters still make the Khmer people unfortunate to suffer from the pain of war.”) The first clause in (S5) is positive (“favorable natural environment”), while the last two clauses are negative (“unfortunate to suffer from the pain of war”). Besides the connectives 儘管 “despite” and 還 是 “still”, the opposing polarity values between the first and the last two clauses is also a strong clue to the existence of a Comparison relation. In addition, the same polarity of the last two clauses is also a hint that no Comparison relation occurs between them. To capture polarity information, we estimate the polarity of each clause and detect the negations from the clause. The polarity score is a real number estimated by a sentiment dictionary-based algorithm. For each clause, the polarity score, and the existence of negation are taken as features. All the models in the experiments are evaluated by 5-fold cross-validation. The metrics are accuracies and macro-averaged F-scores. The t-test is used for significance testing. We firstly examine our model for the task of two-way classification. In this task, binary classifiers are trained to predict the existence of Contingency and Comparison relations in a given sentence. For meaningful comparison, a majority classifier is used as a baseline model, which always predicts the majority class. In the dataset, 72.6% of the sentences involve neither Contingency nor Comparison. Thus, the major class is “Nil”, and the accuracy and the F-score of the baseline model is 72.6% and 42.06%, respectively. The experimental results for the two-way classification task are shown in Table 3. In the table, the symbol † denotes the lowest accuracy which has a significant improvement over the baseline at p=0.05 for the two models. The symbol ‡ denotes the adding of a single feature yields a significant improvement for the model at p=0.005. The performance of the decision tree and the SVM are similar in terms of accuracy and F-score. Overall, the decision tree model achieves better accuracies. In the two-way classification task, the decision tree model with only the Word feature achieves an accuracy of 76.75%, which is significantly better than the baseline at p=0.05. For both the decision tree and the SVM, Connective is the most useful feature: performance is significantly improved with the addition of Connective. Besides the binary classification task, we extend our model to tackle the task of finer classification. In the second task, four-way classifiers are trained to predict a given sentence with four classes: existence of Contingency relations only, existence of Comparison relations only, existence of Both relations, and Nil. The experimental results of the four-way classification task are shown in Table 4. Consistent with the results of the two-way classification task, the addition of Connective to the SVM yields a significant improvement at p=0.005. The performance between the decision tree and the SVM is still similar, but the SVM achieves a slightly better F-score of 45.26% in comparison with the best F-score of 44.47% achieved by the decision tree. We further extend our model to predict the full relation structure of a given sentence as shown in Figure 1 and Figure 4. This is a 49-way classification task because there are 49 types of the full relation structures in the dataset. Not only as many as 49-ways, 72.6% of instances belong to the Nil relation, which yields an unbalanced classification problem. The experimental results are shown in Table 5. In the most challenging case, the SVM achieves a better F-score of 7.66% in comparison with the F-score of 4.90% achieved by the decision tree. Connective is still the most helpful feature. Comparing the F-scores of the SVM in the three tasks with the F-scores of the decision tree, it shows that the SVM performs better for predicting finer classes. We compare the performances between the explicit instances and the implicit instances for the three tasks with the decision tree model trained on all features. The results are shown in Table 6. The higher accuracies and the lower F-scores of the implicit cases are due to the fact that the classifier tends to predict the sentences as Nil when no connective is found, and most implicit samples are Nil. For example, the relation of Contingency in implicit sample (S6) should be inferred from the meaning of 帶給 “brought”. (S6) 得天獨厚的地理環境,的確帶給這個百 年港埠無窮的財富。 (“The unique geographical environment, it really brought the infinite wealth to this hundred-year port.”) In addition, some informal/spoken phrases are useful clues for predicting the relations, but they are not present in our connective dictionary. For example, the phrase 的 話 “if” implies a Contingency relation in (S7). This issue can be addressed by using a larger connective dictionary that contains informal and spoken phrases. (S7) 想要以自助旅行的方式進行的話,那麼 隨團旅遊呢? (“If you want to backpacking, how about an organized tour?”) We regard an instance as explicit if there is at least one connective in the sentence. However, many explicit instances are still not easy to label even with the connectives. As a result, predicting explicit samples is much more challenging than the task of recognizing explicit discourse relations in English. One reason is the ambiguous usage of connectives as shown in (S2). The following sentence depicts another issue. The word 但是 “however” in (S8) is a connective used as a marker of an inter-sentential relation. That is, the entire sentence is one of the arguments of an inter- sentential Comparison relation, but it does not contain any intra-sentential relation inside the sentence ...

Context 2

... office tried to make the Yangmingshan area a more natural environment as the long-term garden of Taipei”) ,但隨著週休二 日 、 經 濟 環 境 改 善 (“But due to the two-day weekend and the improved economic conditions”) (“The issues of tourists parking, garbage, and other indirect effects become more serious”) 。 In (S1), the long sentence consists of three clauses, and such a Chinese sentence is expressed as multiple short sentences in English. Figure 1 shows that a Comparison relation occurs between the first clause and the last two clauses, and a Contingency relation occurs between the second clause and the third clause. An explicit paired discourse marker 雖 (although) ... 但 (but) denotes a Comparison relation in (S1), where the first clause is the first argument of this relation, and the second and the third clauses are the second argument of this relation. In addition, an implicit Contingency relation also occurs between the second and the third clauses. The second clause is the cause argument of this Contingency relation, and the third clause is its effect. It shows a nested relation, which makes relation labeling and relation structure determination challenging. In Chinese, an explicit discourse marker does not always uniquely identify the existence of a particular discourse relation. In sample (S2), a discourse marker 而 “moreover” appears, but neither Contingency nor Comparison relation exists between the two clauses. The discourse marker 而 has many meanings. Here, It has the meaning of “and” or “moreover”, which indicates an Expansion relation. In other usages, it may have the meaning of “but” or “however”, which indicates a Comparison relation. (S2) 而大陸經濟開放10年以來,其進步更 令 人 刮 目 相 看 。 (“Moreover, the progress of mainland is more impressive due to its economic openness for the last 10 years.”) Note that the relation structure of a sentence cannot be exactly derived from the parse tree of the sentence. Shown in Figure 2 is the structure of sample (S3) based on the syntactic tree generated by the Stanford parser. However, it is clear that the correct structure of (S3) is the one shown in Figure 3. (S3) 目前雖然還只能在圖片上讓女性露露臉 (“Although women only appear in the pictures”) , 但 未 來 女 性 的 貢 獻 (“The contribution of women”) ,將是教科書另一個著墨的重點 (“Will be another major focus in textbooks in the future”) This shows that the Stanford parser does not capture the information that the last two clauses form a unit, which in turn is one of the two arguments of a Comparison relation. In this work, we investigate intra-sentential relation detection in Chinese. Given a Chinese sentence, our model will predict if Contingency or Comparison relations exist, and determine their relation structure. In Section 2, the development of a corpus annotated with Contingency and Comparison relations is presented. The methods and the features are proposed in Section 3. In Section 4, the experimental results are shown and discussed. Finally, Section 5 concludes this paper. The corpus is based on the Sinica Treebank (Huang et al., 2000). A Total of 81 articles are randomly selected from the Sino and Travel sets. All the sentences that consist of two, three, and four clauses are extracted for relation and structure labeling by native Chinese speakers. A web-based system is developed for annotation. The annotation scheme is designed as follows. An annotator first signs in to the annotation system, and a list of sentences that are assigned to the annotator are given. The annotator labels the sentences one by one in the system. A sentence is split into clauses along commas, and all of its feasible binary tree structures are shown in the interface. The annotator decides if a Contingency/Comparison relation occurs in this sentence. The sentence will be marked as “Nil” if no relation is found. If there is at least one relation in this sentence, the annotator then chooses the best tree structure of the relations, and the second page is shown. The previously chosen tree structure is presented again, and at this time the annotator has to assign a suitable relation type to each internal node of the tree structure. The relation type includes Contingency “ 因果 ”, Comparison “ 轉折 ”, and Nil. For example, in sample (S4), its three internal nodes are annotated with three relation types as shown in Figure 4. (S4) 即使沒有傳承的使命感 (“Even without the sense of mission of the heritage”) ,為了尋求 更 好 的 治 療 方 式 (“In order to seek better treatments”) ,也會驅使這些醫學工作者跨越領 域區隔 (“These medical workers will be driven crossing domain areas”) ,去尋找資源 (“To find resources”) 。 The number of feasible relation structures of a sentence may be very large depending on the number of clauses. For a sentence with n clauses, the number of its feasible structures is given as the recursive function f ( n ) as follows, and the number of its feasible relation structures is 3 ! ! ! f n . For a two-clause sentence, there are only one tree structure and three possible relation tags (Contingency, Comparison, and Nil) for the only one internal node, the root. For a three-clause sentence, there are two candidate tree structures and nine combinations of the relation tags. For a four-clause sentence, there are five candidate tree structures and 27 combinations of the relation tags. There are theoretically 3, 18, and 135 feasible relation structures for the two-, three-, and four- clause sentences, respectively, though only 49 types of relations structures are observed in the dataset. Each sentence is shown to three annotators, and the majority is taken as the ground-truth. The Fleiss-Kappa of the inter-annotator agreement is 0.44 (moderate agreement). A final decider is involved to break ties. The statistics of our corpus are shown in Table 1. The explicit data are those sentences which have at least one discourse marker. The rest of the data are implicit. A total of 11 explicit sentences which contain both Contingency and Comparison relations form complex sentence compositions. The implicit samples are relatively rare. To predict the intra-sentential relations and structures, two learning algorithms, the modern implementation of the decision tree algorithm, C5.0 , and the support vector machine, SVMlight , are applied. The linguistic features are the crucial part in the learning-based approaches. Various features from different linguistic levels are evaluated in the experiments as shown below. Word : The bags of words in each clause. The Stanford Chinese word segmenter 3 is applied to all the sentences to tokenize the Chinese words. In addition, the first word and the last word in each clause are extracted as distinguished features. POS : The bags of parts of speech (POS) of the words in each clause are also taken as features. All the sentences in the dataset are sent to the Stanford parser 4 that parses a sentence from a surface form into a syntactic tree, labels POS for each word, and generates all the dependencies among the words. In addition, the POS tags of the first word and the last word in each clause are extracted as distinguished features. Length : Several length features are considered, including the number of clauses in the sentence and the number of words for each clause in the sentence. Connective : In English, some words/phrases called connectives are used as discourse markers. For example, the phrase “due to” is a typical connective that indicates a Contingency relation, and the word “however” is a connective that indicates a Comparison relation. Similar to the connectives in English, various words and word pair patterns are usually used as discourse markers in Chinese. A dictionary that contains several types of discourse markers is used. The statistics of the connective dictionary and samples are listed in Table 2. An intra-sentential phrase pair indicates a relation which occurs only inside a sentence. In other words, a relation occurs when the two phrases of an intra-sentential pair exist in the same sentence no matter whether they are in the same clause or not. In contrast, an inter- sentential connective indicates a relation that can occur across neighboring sentences. Some connectives belong to both intra-sentential and inter-sentential types. Each connective in each clause is detected and marked with its corresponding type. For example, the phrase 相對 的 “In contrast” will be marked as a connective that belongs to Comparison relation. The number of types and scopes of the connectives in a sentence are used as features. Dependency : The dependencies among all words in a sentence are used as features. The Stanford parser generates dependency pairs from the sentence. A dependency pair consists of two arguments, i.e., the governor and the dependent, and their types. We are interested in those dependency pairs that are across two clauses. That is, the two arguments of a pair are from different clauses. In our assumption, the clauses have a closer connection if some dependencies occur between them. All such dependency pairs and their types are extracted and counted. Structure : Recent research work reported improved performance using syntactic information for English discourse relation detection. In the work of Pilter and Nenkova (2009), the categories of a tree node, its parent, its left sibling, and its right sibling are taken as ...