<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-1117">
  <Title>A Large-Scale Semantic Structure for Chinese Sentences</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Method
</SectionTitle>
    <Paragraph position="0"> The analysis method that will be presented here is logically equivalent to the parsing of syntax and semantic dependency with feature constraints.</Paragraph>
    <Paragraph position="1"> The key idea in our method is to avoid the complexity hierarchical tree sturcture. We are concerned with building structures that reflect basic relationships between one word and other in a single sentence. We use methods developed for the analysis of semantic relationships to produce a framework based on the order link. We started from an initial analysis based on the surface syntactics, then we analyzed deep semantic relationships, and attempted to improve it by removing the semantic order from the syntactic structure and reconnecting them in different places. Since many word phrase patterns are difficult for computers to recognize, trying to compromise between linguistic correctness and engineering convenience, we link the difference semantic roles on the flat level, while employing a few template rules. All semantic words are linked on the same level. They are non-hierarchical constructs. This flatted representation allows access to various levels of syntactic description tree simultaneously.</Paragraph>
    <Paragraph position="2"> In fact, the purpose of generalization is to get a regular expression from the original sentence.</Paragraph>
    <Paragraph position="3"> We manually tagged two kind of relationship among our large-scale frameworks: 1. syntax-semantic relationship; 2. semantic feature relationship.</Paragraph>
    <Paragraph position="4"> Our framework consists of a set of nodes and a set of arcs that join the nodes, with each word or concept corresponding to a node and links between any two nodes that are directly associated. The basic links in the framework are between one word item to another based on immediate semantic deperdency order. We summarized the immediate semantic relationship through a variety of semantic relation features such as agent, reason, result and so on. The feature of relationship between two nodes are labeled on the arc.</Paragraph>
    <Paragraph position="5"> We developed the first fully instantiated semantic structure by manually labeling semantic representations in a machine-readable format. To make sure that our model can deal with various kinds of texts in real life situations, we have analysed 10000 sentences from large Web site corpora based on our formal model. Our aim is not to describe in detail any specific, but to capture at an abstract level the semantic relations between the direct components in a sentence. Our model's most important domain of application is to Chinese sentence analysis, but it may also be applicable to different languages. This semantic framwork constructs a model on the basis of a few rules.</Paragraph>
    <Paragraph position="6"> The present paper indicates how situation types are represented, how these representations are composed from semantic representations of linguistic constituents, and how these type differences affect the expression of sentences.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Syntax-Semantics Relationship Labeling
</SectionTitle>
      <Paragraph position="0"> This work flow includes linking and labeling of each relation between direct semantic items in single sentences, which reflects different semantic representation, and descriptions of the relations of each frame's basic conceptual structure in terms of semantic actions. A semantic representation is a feature that allows one word in the sentence to point at some other word to which it is related. A word in a sentence may have much direct representation, these are differentiated by the semantic action. By analyzing the direct se mantic representation, we can capture semantic relationships between words, reconstructing a framework for the order of Chinese sentences.</Paragraph>
      <Paragraph position="1"> In most cases syntactic relationships are consistent with semantic relationships. The following framework shows show some important similarities between the structure of syntactic and semantic structure. For example, in Wo Zai Kan Dian Shi . ('I am watching TV.') Syntactically, 'Wo '(I)is subject, directly relating to the verbal predicate 'Kan '(watch), 'Dian Shi '(TV)is object , also links to the verbal predicate directly. 'Zai '(be doing)as a adverb is an adjoined predicate 'Kan '(watch), there is a direct relationship between the two nodes.</Paragraph>
      <Paragraph position="2"> Semantically, 'Wo '(I)is the agent and 'Dian Shi ' (TV)is the recipients, both of them have a direct relationship with the activity 'Kan '(watch). So we link the different nodes as follows: In cases where the relationship between syntax and semantics is inconsistent, by syntactic analysis, if there are multiple syntactic analyses among a sentence, we always choose the analysis</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 'Head' Determination relationship that is consistent with the semantic
</SectionTitle>
      <Paragraph position="0"> relationship. For example, the Chinese sentence The basic link is the direct link between two semantic units. In addition, a set of general rules for determining the directions has been identified.</Paragraph>
      <Paragraph position="1"> Jie Bian Zuo Zhao Xu Duo Ren .</Paragraph>
      <Paragraph position="2"> (many people sit beside the street.) The above sentence can be analyzed either of the following two syntactic structures.</Paragraph>
      <Paragraph position="3">  1. That between Head and Its Modifier as a Case of Direct Relationship type 1:  The head (see below), and the modifiers that come before it, constitute a type of modification relationship, which is one of the typical cases of direct relationships, e.g, A. Gao zige de ren type 2: tall body DE person the person with tall body B. (to be compared with the above sentence) ren de gezi gao person DE body tall The two syntactic structures are analyzed with difference in the first node and the second node. In type 1, 'Jie Bian '(beside of the street)is analyzed as subject, for type2, the linguist also analyzed it as adverb modifier, adjuncting to the predicate 'Zuo ' (seat). But when this sentence is analyzed in terms of semantics, there is only one relationship structure similar as type 2. 'Ren '(people)is analyzed as agent, 'Jie Bian '(beside of the street) as localizer, attached to the activity 'Zuo '(seat). This semantic structure is consistent with the syntactic structure type 2. Only one structure can display both syntax and semantic relationship simultaneously. So we choose the second analysis. 'The person's body is tall.' In the above sentence, ren 'person' and gezi 'body' hold a modification relationship, but gao 'tall' and ren'person' are related indirectly as the relationship between the two words is realized through that of gezi 'body'. Therefore, we say that the relationship that ren 'person' holds with gezi 'body' is a direct one, but that with gao is a rather indirect one.</Paragraph>
      <Paragraph position="4"> 2. That between An Action Verb and Its Patient as a Case of a Direct relationship In case a head noun is an AGENT of an action verb within a modifying phrase, then the relationship between the Head none and the action verb is a direct one. The following sentences illustrate the point.</Paragraph>
      <Paragraph position="5"> If the syntactic relationship is different from the semantic relationship, we take no account of the syntactic order. In the Chinese sentence C. chi pingguo de nuhai.</Paragraph>
      <Paragraph position="6"> Eat apples DE girl Ta Ku Hong Liao Yan Jing .</Paragraph>
      <Paragraph position="7"> 'the girl who is eating apples.' (she cry so much that her eyes become red.) D. (to be compared with the above sentence) Within the surface syntactic structure, adjective 'Hong ' (red)will be analyzed as complementation and directly associated with main verb 'Ku ' (cry) , which indicate result of predicate. Underlying the syntactic structure, 'Hong ' (red) actually point to 'Yan Jing '(eyes)in semantic representation. There is no direct semantic relationship between 'Ku ' (cry)and 'Hong ' (red). The semantic network can be analyzed as: she cry + her eyes become red, the immediate relationship between 'he' as a possessor and 'belly' as a possession and that between 'belly' as entity and 'painful' as description. In this case we link the node 'Hong ' (red)to 'Yan Jing '(eyes) directly based on semantic relationship.</Paragraph>
      <Paragraph position="8"> nuhai chi pingguo girl eat apples 'The girl is eating apples.' In the above sentence, nuhai 'girl' is an AGENT of the action verb chi 'eat', the two words have a direct semantic relationship, therefore we link them directly and annotate 'girl' as a head. In contrary, the relationship between nuhai 'girl' and pingguo 'apples' is of an indirect type.</Paragraph>
      <Paragraph position="9">  3. Other Cases of Direct Relationships  In case there is neither a modification nor an AGENT/PATIENT relationship, the whole phrase, which is still directly related to a following describing phrase, has to be embedded. E.g., E. ban shiqing yinggai guquan daju.</Paragraph>
      <Paragraph position="10"> Handle problem should care-about overall situation 'People should care about the overall situation when they handle problems.' F. chouyan hai shenti.</Paragraph>
      <Paragraph position="11"> Smoke harm health 'Smoking harms health.' G. ta neng daying de shiqing wo ye neng daying. He can accept DE event I also can accept The above three head semantic structures clearly show us the different relationships among sentence and noun phrases with different meaning. The head words are connected to their modifier through arrow arcs. The first SVO relationship is also represented by non-head tagging.</Paragraph>
      <Paragraph position="12"> 'The event that he can accept are also acceptable to me.'</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.3 'Head' Determination
</SectionTitle>
      <Paragraph position="0"> Since Chinese lacks morphological cues, the grammatical markers (such as De ,Ba ,Bei ) and word order are comparatively important cues for the relationship determination. We have to rely on grammatical and semantic knowledge to guide role assignment.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.4 Feature Abstracting and Labeling
</SectionTitle>
      <Paragraph position="0"> Based on the analysis of semantic relationships, we have been parsing feature structures to express dependencies between semantic features. In our analysis model, semantic feature means a variety of detailed semantic relationships. Most of the time, semantic features are not so easy to define.</Paragraph>
      <Paragraph position="1"> Some feature typologies have been provided, but there is still much discussions about the nature of a the confusion of feature cation, we proposed a method to abstract the semantic feature directly from se contain the natural feature word. For those sentences without semantic features in labeling the semantic categorys include in other sentences, attached on the relationship arcs. T semantic framework based on dimension. For exam In this study, we have proposed an approach that combines 'basic conceptual structure' and our 'Head-Driven Principle'. According to the 'Head-Driven Principle', most structures are analyzed as having a 'Head' which is connected to various types of modifiers, such as Head-NP (adjectivenoun, noun-adverbial pairsWo Men Du ), Head-VP (adverbial-verb, verb-adverbial, adjective-verb...). In our framework, modification is represented by attaching tags with arrows to the core semantic item whereve the type of modification can be clearly identified. Since the SVO is the basic order in Chinese, there is no modifier relationship among the level of SVO. In our model, 'Subject-Predicate Structures' and 'Verb-Object Structures' are represented as non-head. In above example, the relation linking the 'core' noun and verb with their 'adjunct' is tagged with an arrow to indicate that it is a 'head'. Both A and B label the 'head' as the core noun. E labels the 'head' as the core verb.</Paragraph>
      <Paragraph position="2"> Employing the 'Head-Driven Principle' for the construction of semantic models. Some ambiguous sentences can be clearly represented. The different meaning among sentence or phrase containing same words can also be described . Conside the following sentence and phrases: Ta gezi bu gao.</Paragraph>
      <Paragraph position="3"> His stature isn't tall.</Paragraph>
      <Paragraph position="4"> He isn't tall.</Paragraph>
      <Paragraph position="5"> stature he In traditional analy constituent in a sentence.</Paragraph>
      <Paragraph position="6"> meaning of the sentence i is semantic feature linking 'he' and 'tall' thus in our semantic analy 'tall' semantically, 'sta marking a semantic relationship, immediate constituent.</Paragraph>
      <Paragraph position="7"> structure, after feature abstraction, is to its English counterpart. It facilitates the translation from one language into anot Xue Sheng Xi Huan Lao Shi .</Paragraph>
      <Paragraph position="8"> (The students like the teachers.) Xi Huan Lao Shi De Xue Sheng (the students who like the teachers) Xue Sheng Xi Huan De Lao Shi (the teachers who the students like) All of above examples containing same meaning words can have very different meaning, depending on the different word order and grammatical marker 'De '(DE). We use head tagging to construct different frameworks for these structures: In some sentences , there features but also their particular values included. Similarly we abstracted the features. Thus we can exp to express this level of detail. For exam Ta liang mi he two meters tall features refer to the hus we con ple: not sis, 'stature' is just a syntactic However, the essential s 'he is not tall', 'stature sis we link onl ture' is taken as feature rath This Chinese semantic are not only values attach and the feature structure gao.</Paragraph>
      <Paragraph position="9"> ntences that feature in a text. To avoid  'He is two meters tall.' Several different sentences which should be analyzed as having the same syntactic structure may have fundamentally different semantic structures. The following three sentences S1, S2and S3, for example, should be analyzed as having the syntactic structure, but their semantic structures are nevertheless represented as S1', S2' and S3' respectively in our framework. he tall two meters In the above framework, 'tall' is the semantic feature describing staturs of the agent 'he' , and 'two meters' express the value of the feature. They provid different information at different level, constructing a feature sturcture. NP + V + Adj + NP S1=Ta xiao-tong le duzi</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 The Advantages of Our Semantic Model
</SectionTitle>
    <Paragraph position="0"> he laugh-painful ASP belly In developing our semantic frameworks, we also have articulated a framework of 'Noun-Centrality' as a supplement to the widely assumed 'Verb-Centrality' practice. We can successfully disambiguate some troublesome sentences, and minimize the redundancy in language knowledge description for natural language processing. We automatically learn a simpler, less redundant representation of the same information.</Paragraph>
    <Paragraph position="1"> 'He laughed so much that his belly was painful.' S2=Wo kan-tou le ni I see- through ASP you 'I understand you thoroughly.' S3=Ta da po-le beizi She broke up the cup First, comparing syntactic order and semantic order, we used the reconstructed original order, giving some different order sentences similar results. Thus, variations of order in the same sentence can reveal the same relationships. She broke up the cup.</Paragraph>
    <Paragraph position="2"> S1': NP V Adj NP One semantic structure may correspond to more syntactic structures in Chinese, and this correspondence can be made specifically clear using our approach.</Paragraph>
    <Paragraph position="3">  The above three sentences, their syntactic structures are clearly different from each other. That is, the direct object wo 'me' appears right after the main verb in (1) whereas the same logical object has moved to a pre-verbal position with the help of a special Chinese preposition BA in (2) and to a sentence-initial position with the help of BEI in (3). But underlying the difference syntactic structures, they share the same basic semantic structure, using semantic represented expression, the three sentences of above example can be described in below.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
AGENT Ta 'she'
PATIENT Wo'me'
ACTION Da 'beat'
S2':NP V Adj NP
S3':NP V Adj NP
</SectionTitle>
    <Paragraph position="0"> On the other hand, many structural ambiguities in Chinese sentences are one of the major problems in Chinese syntactic analyses. One syntactic structure may correspond to two or more semantic structures, that is, various forms of structural ambiguity are widely observed in Chinese.</Paragraph>
    <Paragraph position="1"> Disregarding the semantic types will cause syntactic ambiguity. If this type of information is not available during parsing, important clues will be missing, and loss of accuracy will result.</Paragraph>
    <Paragraph position="2"> Consider the Chinese sentence Ta de yifu zuo de piaoliang.</Paragraph>
    <Paragraph position="3"> Her cloth do DE beautiful Reading 1: 'She has made the cloth beautifully b) minimal redundancy in language knowledge description for natural language processing.</Paragraph>
    <Paragraph position="4"> Reading 2: (Somebody) has made her cloth beautifully.' We hope to use the minimum analysis method to find the semantic order with equal relationship among new sentence. We then used the partition relationship as a training database to recognize new order as similar as these order structures. Syntactically, the sentence, with either one of the above two semantic interpretations, should be analyzed as S We also have been creating feature sets parsing feature structures to expressing dependencies between semantic features. Furthermore, we abstracted the values attached to the features. Thus we can expand the feature structures to express this level of detail.</Paragraph>
  </Section>
class="xml-element"></Paper>