62 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. abstraCt Considering the popularity of the Internet, an automatic interactive feedback system for E- learning websites is becoming increasingly desirable. However, computers still have problems understanding natural languages, especially the Chinese language, firstly because the Chinese language has no space to segment lexical entries (its segmentation method is more difficult than that of English) and secondly because of the lack of a complete grammar in the Chinese language, making parsing more difficult and complicated. Building an automated Chinese feedback system for special application domains could solve these problems. This paper proposes an interactive feedback mechanism in a virtual campus that can parse, understand and respond to Chinese sentences. This mechanism utilizes a specific lexical database according to the particular ap- plication. In this way, a virtual campus website can implement a special application domain that chooses the proper response in a user friendly, accurate and timely manner. Keywords: grammar; interactive feedback; lexical database; natural language; segmenta- tion method iNtroduCtioN The easiest way to communicate to users is to talk to them in their natural language. Considering the popularity of the Internet, an automated interactive feedback system for e-learning Web sites is becoming in- creasingly desirable. However, it still is difficult for a computer to understand the meaning of some natural languages. At pres- ent a three-year old child can understand and respond to languages better than a computer can. To understand the natural language, a computer must be trained to understand a single sentence. Then, it would need to be trained to analyze longer sentences or paragraphs. In principle, there are at least a Chinese interactive feedback system for a virtual Campus Jui-Fa Chen, Tamkang University, Taiwan Wei-Chuan Lin, Tak Ming College, Taiwan Chih-Yu Jian, Tamkang University, Taiwan Ching-Chung Hung, Tamkang University, Taiwan IGI PUBLISHING This paper appears in the publication, International Journal of Distance Education Technologies , Volume 6, Issue 4 edited by Shi-Kuo Chang &Timothy K. Shih © 2008, IGI Global 701 E. Chocolate Avenue, Suite 200, Hershey PA 17033-1240, USA Tel: 717/533-8845; Fax 717/533-8661; URL-http://www.igi-global.com ITJ 4443 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 63 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. two skills that a computer should be able to apply to a single sentence: 1. Defining the meaning of each word in the sentence. 2. Transforming the linear structure of a sentence into another structure that rep- resents the meaning of that sentence. The first step of processing a Chinese sentence is seeking the meaning of each lexicon in a dictionary. However, there can be many meanings for each lexicon, and the computer must have the ability to choose the right one. Even if that is accomplished, it is still difficult for the computer to pro- cess the Chinese sentence because there are no spaces used to segment the lexicon. Therefore, a segmentation method is needed before parsing the Chinese sentences. The second step of understanding a Chinese sentence is transforming the seg- mented lexicons into a structure that can be understood by a computer. In general, the transformation procedure can be divided into three parts: A. Syntactic analysis procedure: In this procedure, the input lexicon is trans- formed into a specific structure that represents the relationship between lexicons. However, not all the combi- nations of lexicons of a sentence are legal. The computer must eliminate the illegal combinations to ensure a correct performance. B. Semantic analysis procedure: This procedure obtains the meaning of the sentence from the established struc- ture. The obtained meaning is a unit of knowledge representation, which can be mapped to the corresponding object or event in the actual world. C. Pragmatics analysis procedure: This procedure determines the real purpose of the sentences and gives the appropri- ate response to users. The remainder of this paper is laid out as followed. The next section discusses the related works on syntax and semantic analysis, followed by a description of the proposed four subsystems of segmentation, syntactic analysis, semantic analysis, and the response subsystems. The next sec- tion provides some examples to show the implementation of the proposed method. Finally, there is conclusion and some fu- ture works. revieW of related Works link Grammar technology Most sentences in a natural language are structured so that arcs that connect words may not cross each other. This phenomenon is called planarity in the link grammar system (Sleator & Temperley, 1991). A link grammar consists of a set of words and has a linking requirement. The linking requirements of each word are contained in a dictionary. To illustrate the linking require- ments, Figure 1 shows a simple dictionary for the words “a,” “the,” “cat,” “mouse,” and “chased.” The linking requirement of each word is represented by the Figure 1 above the word. Each of the lettered boxes is a connec- tor which is satisfied when it is “plugged into” a compatible connector, as indicated by its shape. If the mating end of a con- nector is drawn facing to the right, then its mate must be to its right facing to the left. Exactly one of the connectors attached to a given black dot must be satisfied. Thus, the “cat” requires a D connector to its left 64 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. and either an O connector to its left or an S connector to its right. Plugging a pair of connectors together corresponds to drawing a link between that pair of words. Figure 2 is the simplified form of Figure 1 and shows that “the cat chased a mouse” is part of the language. Table 1 encodes the linking requirements of the example in Figure 2. The link grammar dictionary consists of a collection of entries, each of which defines the linking requirements of one or more words. These requirements are speci- fied by a formula of connectors combined by the binary associative operators & and or. Precedence is specified by parentheses. A connector is simply a character string ending in + or -. Memory-based Parsing system Most methods of semantic analysis first recognize the verb of a sentence and then determine the correctness on the semantics of lexical entries around the verb. Memory- based parsing (Chung, & Moldovan, 1993, 1994a, 1994b; Kim & Moldovan, 1993) also begins with the restrictions of a verb to determine the correctness of subject and object. The memory-based parsing system consists of four modules: • Concept sequence layer: Keeps the restrictions of the subject and object of each verb for both syntax and se- mantics. • Syntactic layer: Keeps all parts of speech for comparing the syntactic restrictions. • Semantic concept hierarchy: Defines the relationship of all nouns, and is Figure 1. Words and connectors in the dictionary the cat chased D D S S O a D mouse D O Words Formula a the cat mouse Chased D+ D- & (O- or S+) S- & O+ Table 1. The words and linking requirements in a dictionary Figure 2. The simplified form of Figure 1 The cat chased a mouse D S O D Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 6� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. used for verifying the semantic restric- tions. • Instance layer: Contains the lexical en- tries of a sentence typed by the user. Figure 3 shows an example of a memory-based parsing with a concept sequence [agent, MURDER, object] for murder-event. At the top of the knowledge base is the concept sequence layer, which consists of concept sequence roots and elements. The semantic concept hierar- chy and syntactic layer connect concept sequence elements with concept instances in the instance layer. Concept instances are produced from phrasal inputs and are connected to the corresponding syntactic category and semantic concept nodes. The result of parsing is represented by connecting instances of concept sequence roots and corresponding concepts in the instance layer. the ProPosed systeM overview There are many learners in a virtual cam- pus, and each learner has his or her own preference. Although the search goal can be found by a belief network, as considered in customization, using only the default category to analyze is insufficient. When a learner logs onto a virtual campus, if he or she is an existing learner, the system could load his learning profile to achieve the customization. If the learner is new, the system could administer a quiz to determine an initial learning profile. The flowchart is shown in Figure 4 The proposed Chinese interactive feedback system (Chen, Lin, & Jian, 2003a, 2003b, 2003c; Chen, Lin, Jian, & Hung, 2005) is divided into four sub-sys- tems: the segmentation system, syntactic analysis system, semantic analysis system, Figure 3. Part of knowledge base used for processing: “The Shining Path” Concept sequence layer [agent, MURDER, object] murder-event agent10 MURDER10 object10 first last next next Syntatic layer noun-group syntactic Instance layer shining-path#1 instance lexical “the shining path” Semantic concept hierarchy thing physical abstract is-a is-a animate inanimate is-a is-a human non- human is-a is-a human- group person is-a is-a terrorism- organization is-a nation is-a shining path FXLM is-a is-a is-a instance 66 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. and response system. Thus, learners can use Chinese sentences to interact with the virtual campus. When the learners input Chinese sentences, the segmentation system separates the learner’s input sentences and gives the appropriate part of speech for each segmented lexical entry. The syntactic analysis system parses these segmented lexical entries to judge whether the sentence is legal and gives the syntactic part of each lexical entry. The semantic analysis system judges the correctness of the semantics and provides a semantic learning method based on the learner’s habits. Finally, a response system gives the learner the response re- sult according to the encoding of the input sentence. segmentation system One difference between the Chinese and English language is that the Chinese lan- guage has no obvious separation to segment the lexical entry. Therefore a segmentation method to parse the Chinese language is necessary. Figure 5 shows the architecture of the Segmentation System. The segmentation system structure is divided into four sub-systems: the segmen- tation system, corpora-comparing system, keyword in context comparing system, and weighted calculation method (Chen, Lin, & Jian, 2003a). These subsystems are explained as follows: 1. Segmentation: Segmentation separates the user’s input sentences and compares the separated units with those obtained from the corpora-comparing system. 2. Corpora-comparing system: This system includes two steps: corpora- comparing and part-of-speech(POS) saving. It compares the receiving strings with those in the corpora and saves the results to build a segmenta- tion tree. 3. Keyword in context comparing system: After building a segmentation tree, the system compares the POS with the keyword in context according to the grammar rules and deletes the improper segmentation tree. This mechanism is divided into the Unknown Word Judg- ment System and Context-proofreading System. 4. Weighted calculation system: Because there may be more than one kind of segmentation result, each result’s weighted value is computed to find the most proper one. The segmentation result with the largest weighted value is the most suitable result. Figure 4. Flowchart of feedback system User login Reasoning system User’ s degree Response suitable teaching material User learning User logout User’ s profile Update First login ? NO Simple testYes Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 6� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. �e���e��t�t���� There is no space between lexical entries in the Chinese language to help segmenta- tion. Chinese characters are composed of two continuous bytes in representation. The system judges whether this word is Chinese code when segmenting sentences to put the pointer’s displacement in the best place. However, in transmitting, some special Chi- nese words have a special『\』inserted after transmitting through the network browser. The system would remove the『\』 prior to segmenting the sentences. The main functions of the segmenta- tion system are: 1. The consideration of special Chinese words in the user’s input sentences serves to avoid punctuation-transfer- ring mistakes 2. The transfer of punctuation in the user’s input sentences serves to obtain the same dividing code. The system splits the continuous Chinese words and numbers them into strings and adds a dividing code both in front of and be- hind the strings. The system also adds a dividing code behind the empty word of the user’s input sentence. In addition the system also splits the user’s input sentence and compares the segmented lexical entries with the corpora-com- paring system. Figure 6 shows the segmentation system process. Most Chinese lexicons possess at most six characters. The segmentation length of a Chinese sentence should be limited to avoid segmenting a sentence into many impossible ways. For example, a sen- tence composed of n words should have 2n-1 possible segmentations. A maximum matching (Chen & Liu, 1992) mechanism is used to segment a sentence. Basically, the maximum matching method compares a string started at the kth character with a lexical database and finds out all possible segmentations. If C(k), C(k)_C(k+1), C(k)_C(k+1)_C(k+2) are stored in the lexical database, the maximum matching Figure 5. The architecture of the segmentation system User's Client WWW Server HTML PageInput Chinese Language prehandle sentence segmenting lexical entries comparing corpora record POS establish Segmentation tree professional corpora comparing with keyword in context grama rule Computing weight form word mark rule Segmentation Tree wordbuilding rule and weight Segmentation tree and results 68 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. method would choose the longest word and continue with C(k+3). Because the length of most of the Chinese lexicons do not exceed six characters, the maximum length of a word in a sentence is set to six characters. C�rp�r�-C���p�r���� ��ste�� After the segmentation system separates the sentences, it compares string length and context. The system records all match- ing POSs and adds the information of the lexical entries to a segmentation tree. If no POSs match, the system feeds back a false value and recalls the unknown word judgment sub-system to determine whether this word is an unknown word. If the lexi- cal entry is determined by the system to be an unknown word, then the entry is saved into the segmentation tree. Because most new unknown words are proper names, the system often sets the POSs of these unknown words to be temporary nouns, and continues processing the following set of strings. If the system still cannot find the corresponding POS in the corpus database of one to six continuous words or cannot find the proper unknown word after processing by the unknown word judgment system, it views these six continuous words as an unknown word and adds the unknown word into the segmentation tree. The corpora structure used in the system is shown in Table 2. The saved data format contains the numbers of words, context, POS, types, and word probability. They are explained as follows: A. Numbers of words: To speed up the comparing of the corpora, the informa- tion of the numbers of lexical entries are recorded so that the system does not have to search the entire database, greatly improving the efficiency of the system. B. Context: Refers to the recorded context of the lexical entry. C. POS: Records the POS of the lexical entry. If the number of the POS is larger than one, the system separates the sen- tence with “,” as a divided symbol. D. Types: This paper is focused on mutual conversation segmentations in the basic Figure 6. Segmentation system’s flow chart source special chinese Y punctuation transform into divide char. Y N form word N insert divide char. Y send segmented words N divide char. N move pointer remove surplus 『/』 Numbers of words Context POS Types Word frequency Table 2. Corpora data structure Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 6� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. computer concept domain. Therefore, the type is used to mark the kind of special domain database that is used for the lexical entry. E. Word frequency: Shows how often the lexical entry has appeared in the equilibrium corpus database. The in- formation is used primarily for weight calculation. U��k������ W�rd Jud���e��t ��ste�� This system searches the segmented lexical entries for unknown words. After the sys- tem receives strings, it splits N continuous words continuously and compares them with the corpus database. If the proofread- ing is successful, the system feeds the first words of the lexical entry back to the posi- tion where the string engages. The system sets the string which is beyond the position of being an unknown word and saves it into a segmentation tree. If the system fails when compared, it feeds back 0 to show that this word is not an unknown word. Figure 7 shows the flow of the unknown word judgment system. T��e ��t� �tructure �f t��e �e���e��t�t���� Tree N�de The system adds the segmented lexical entry into the segmentation tree to speed up node searching. The segmentation tree structure can make data saving more flexible by increasing or decreasing segmentation nodes. The segmentation tree is a six node tree. The tree structure is shown in Figure 8. Every node in Figure 8 follows from zero to at most six sub-nodes which are added dynamically when compared with the corpora. The original input sentence connects the first node of the root to the following branches. In this way the system can dispose of space dynamically to save and display the segmentation results. Every node of the segmentation tree is composed of the following node structure as shown in Figure 9. Each node records the information after the system searches the database which is convenient for the context-proofreading system and the weighted-calculating system. The fields in Figure 9 are explained as follows: Figure 7. Unknown word judgment system process string and pointer i=1 comparing success Y return ptr i<= MaxLen Y i=i+1 N N return 0 �0 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. A. Context: The context of the lexical entry. B. Length: Records the length of the strings. C. Array of POS: Records the POS of a lexical entry (up to six) with a pre-set value of an empty string. D. Result of database searching: Re- cords the searching result of this lexical entry. If the searching result is found from the corpus database, it is recorded as true. However, if it is an unknown word, it is recorded as false. E. Down-connection and up-connec- tion: Records the number of the up- per or lower nodes, referring to the upward or downward lexical entry. If there is no up or down connection, it records 0. Because the system deals at most with six continuous words when segmenting sentences, the array size of the down-connecting is six and that of the up-connecting array is one. root three-word four-wordtwo-word five-wordone-word six-word Figure 8. Segmentation tree structure Context Result of database searching Length up-connectionArray of POS down-connect ion array Figure 9. Node data structure Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 �1 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. Context-Proofreading system This system uses the segmentation tree built in the corpora-comparing system to evaluate the context according to the grammar recorded in the grammar principle database. This system deletes segmentation sub-trees which are not matched with the grammar and decides the POS which the lexical entry belongs to. When proceeding with the grammar proofreading, the system only compares the POS of the front lexical entry rather than proofreading the whole article so that the system can determine the POS of the lexical entry in the oral language conversation more correctly while increasing speed and flexibility of the judgment. The main reason for adopting the method of judging the relationship of the grammar between the front and rear words is that grammar structure is usually not perfect in an oral language conversation. If the system uses only grammar rules it would find errors in determining the POS. In contrast, if the system checks only the relationship between the front and rear words, it would correctly determine the POS. The detailed procedure is shown in Figure 10. Weighted-Calculation system The weighted-calculation system is used to judge the correctness of segmentation results when there is more than one result after grammar analyzation, as a segmented Chinese sentence may have more than one suitable way for splitting. This system com- putes the weights according to the lexical entry building principle, with the segmenta- tion result having the largest weight being the correct one. The process of the weighted calculation is listed as follows: Weight = Weights of length * Weights of searching result * Word frequency A. Weights of length: The longer lexical entry has a higher priority according to the lexical entry-building principle. Therefore, the longer the length the larger the weight . B. Weight of searching result: The weight of the searching result changes based on whether this word is an unknown Figure 10. The flow chart of the keyword in context comparing system Segmentation tree move pointer to sub-tree fit in with grammar Y record analysis result delete this sub-tree N have un-search sub-tree Y finish analyze N have un-search sub-tree Y finish analyze N �2 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. word or not. In principal, the weight of a known word is larger than that of an unknown word. However, according to the principle of long lexical entry privacy, the system sets the weight of an unknown word to be the same as N-continuous words. Therefore, the weight of an unknown word is only slightly larger than a one-continuous known word. C. Word frequency: Shows how often the lexical entry appears in the equilibrium corpora. The more often the lexical entry appears, the higher frequency it has. After calculating the weighted sum of all nodes on every branch the system can find the segmented result that is the most suitable for the lexical entry-build- ing principle. syntactic analysis system The main function of the syntactic analysis system is to transform the lexical entries of the input sentence into a structure that can represent the relationship of these lexical entries. However, not all the input sentences are legal in syntax, and the system should provide a fault-tolerance mechanism. With a fault-tolerance mechanism, the system can tolerate common mistakes in general oral conversation and thereby increase the level of fluency in the conversation. Figure 11 shows the flowchart of the syntactic analysis system which utilizes the “Word-based Link Grammar” (Sleator & Temperley, 1991) as the parsing method of the syntax. Word-based link Grammar The method of the Word-based Link Gram- mar defines the linking rules on each lexical entry for making the link relations. The syntactic analysis system obtains the rela- tions as the syntactic parts of each lexical entry. Table 3 shows the linking rules of each part of speech. When the syntactic analysis system starts analyzing, it obtains the linking rules of each lexical entry from a dictionary and makes a link according to these linking rules. The parsing algorithm is shown as Algorithm 1. Figure 11. Flowchart of the syntactic analysis system Segmented sentence Fault-tolerance syntactic parser Special error rules Link grammar rules and general error rules Definition of general error rules Identify of user Legal in syntax yes no Error log Adjustment of error rules Semantic analysis system Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 �3 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. F�u�t-T��er���ce Mec������s�� The sentences that have a syntax error usually appear in oral conversations and those sentences that are difficult to parse. Therefore, it is necessary for the syntactic analysis system to provide a fault-toler- ance mechanism. The proposed syntactic analysis system provides the fault-tolerance mechanism by modifying the linking rules of interrelated lexical entries. Figure 12 shows an example of fault-tolerance pro- cessing by omitting the preposition . In the first block of Figure 12, after the segmentation system process, the correct Chinese sentence is segmented into , , and and their parts of speech are “Pa”, “D”, and “N” respectively. The system obtains the linking rules of each lexical entry from the dictionary and checks if the linkage of each lexical entry is correct. However, in the second block of Figure 12, because of the omission of preposition , the sentence can not make a connection between lexical entry and by means of the linking rules. Therefore, in the last block of Figure 12, with the defining of error linking rules “Err_D”, the lexical entry and can make a connection by linking rules “Err_D” so as to provide the fault-tolerance processing. semantic analysis system The semantic analysis system, as shown in Figure 13, transforms the structure of the sentence, as constructed by the syntactic analysis system, into the semantic meaning. part of speech linking rules Noun(N) (S+ or O-)&(Q- or())&(@Adj- or ())&(Do- or ())&(Ds+ or ())&(Cn1+ or ())&(Cn2- or ()) Personal pronoun(Pa) (S+ or O-)&(@Adj- or ())&(Ds+ or ())&(Cn1+ or ())&(Cn2- or ()) Demonstrative(Pb) (Bs+ or Pq+) Doubt pronoun(Pc) (S+ or O-) Quantifier (Q) (num- or Pq- or (Pq- & num-))&(Q+) Adjective(Adj) (Adj+ or Bj-)&(Adva- or ())&(Noj- or ())&(Ca1+ or ())&(Ca2- or ()) Adverb-decorate adjective(Adva) (Adva+) Adverb-decorate verb(Advb) (Advb+) Negation(No) (Noj+ or Nov+) Auxiliary verb(Hv) (Hv+) Transitive verb(Vt) (Hv- or ())&(S-)&(O+)&(Advb- or()) Intransitive verb(Vi) (Hv- or ())&(S-)&(Advb- or()) Preposition(D) (Ds- & Do+) Conjunction(C) (Ca1- & Ca2+)or(Cn1- & Cn2+) Indicative(Bv) (Bs- or S-)&(O+ or ())&(Bj+ or ()) Table 3. The linking rules of each part of speech �4 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. Comment: Sentence: sentence inputted by user Token: segmented lexical entry First_Token: first lexical entry of sentence Last_Token: last lexical entry of sentence Token_Link: flag of whether the lexical entry is linked or not Link_Grammar: linking rules of lexical entry Disjuncts: linking rules in disjunctive form Syntactic_Error: syntactic error flag Right_Links: right connectors of linking rules Left_Links: left connectors of linking rules Syntactic_Part: syntactic part Syntactic_Error_Procedure: procedure when errors exist on syntax BEGIN get Tokens of Sentence segmented by the segmentation system END BEGIN FOR(i=First_Token to Last_Token) BEGIN set Token_Link off get Link_Grammar of the ith Token from Dictionary make Disjuncts of the ith Token END set Syntactic_Error off FOR(i=First_Token to Last_Token) BEGIN FOR(j=next Token of the ith Token to the Last_Token and exist Right_Links) BEGIN IF(one of jth Token's Left_Links matches one of ith Token's Right_Links) THEN BEGIN 1.make a link between the ith and the jth Token and assign Syntactic_Part 2.set both ith and jth Token's Token_Link on 3.remove the Disjuncts of the ith Token and the jth Token that are without a link 4.remove this link from the Disjuncts of the ith Token END END IF(ith Token's Token_Link=off) THEN BEGIN set Syntactic_Error on END END IF(Syntactic_Error=on) THEN BEGIN call Syntactic_Error_Procedure() END END Algorithm 1. Syntactic analysis system Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 �� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. The system judges the correctness of the semantics and provides a semantic learning method based on the user’s oral habits. Because the judgment of semantics only determines the correctness of the subject and the object around the verb, the proposed system searches for the verb of a sentence in advance. If there is no verb in the sentence, the system will continue to the next sub-system after retaining the semantic meaning in the semantic network. The parsing algorithm of semantics is shown as Algorithm 2. Me���r�-B�sed P�rs���� ��ste�� The proposed system utilizes the “Memory- based parsing system” (Chung & Moldo- van, 1994b) as the parsing method of the semantics. There are three parts in the memory-based parsing system: the con- Figure 12. Fault-tolerance processing with omitting of the preposition cept sequence layer, the semantic concept hierarchy, and the instance layer. Concept sequence layer The concept sequence layer keeps both the syntactic and the semantic restrictions of the subject and the object around the verb. As shown in Figure 14, the concept sequence layer takes the verb as the principal element. The verb element links to both the subject and the object elements via the pointers to obtain their restrictions. The detailed contents are explained as follows: • Structure of the verb:  Lexical entry of the verb: Save the context of the verb.  S: Link to the restriction of the subject.  O: Link to the restriction of the object. �6 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. Figure 13. Flowchart of the semantic analysis system Comment: Token: lexical entry First_Token: first lexical entry of asentence Last_Token: last lexical entry of a sentence Syntactic_Part: syntactic part Verb: the verb of a sentence Subjective_Token: the lexical entry of a subject Objective_Token: the lexical entry of a object Verb_Token: the lexical entry of a verb Semantic_Error_Procedure: the procedure when an error exist in the semantics Semantic_Network: semantic network BEGIN FOR(i=First_Token to Last_Token) BEGIN IF(exist Token which Syntatic_Part is a Verb) THEN BEGIN search Subjective_Token and Objective_Token that is related to this Verb_Token check semantics between Subjective_Token and Verb_Token check semantics between Verb_Token and Objective_Token IF(semantics is not illegal) THEN BEGIN call Semantic_Error_Procedure() return END according to Subjective_Token, Verb_Token and Objective_Token create Semantic_Network END END FOR(i=First_Token to Last_Token) BEGIN IF(Token isn't in Semantic_Network) THEN BEGIN insert the Token into the Semantic_Network according to the link END END END Algorithm 2. Semantic analysis system Result of syntactic analysis Is there any verb Get the verb Get subject and object around the verb Semantic parsing Is legal in semantic Concept sequence database Semantic concept hierarchy Error lognoyes yesnoResponsion system Semantic meaning Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 �� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. Figure 14. Structure of the concept sequence layer  Time: Link to the parts of the sen- tence regarding time.  Place: Link to the parts of sentence regarding place.  Sentence: Link to the sub-sen- tence. • Structure of the subject or the object:  Syntactic restriction: Record the restriction of the syntax.  Semantic restriction: Record the restriction of the semantics.  Ellipsis flag: For judging whether the subject or the object element can be omitted.  Lexical entry or sentence: Link to the actual lexical entry or sub-sen- tence of the subject or the object. Based on the structure of the concept sequence layer, the system provides not only verification of semantic restrictions but also a basis of encoding of the input sentence. semantic Concept hierarchy The semantic concept hierarchy defines the relation of the nouns according to the meaning of the nouns. The semantic restric- tions of subject and object in the concept sequence layer directs them to get their restrictions via the pointers. The structure of the semantic concept hierarchy is shown in Figure 15. The detailed contents are explained as follows: • ID: Store the identity number of the noun. • Lexical entry: Store the context of the noun. • ID of parent: Keep up-link of the parent’s ID. instance layer In the instance layer, the proposed system records the lexical entries and obtains the semantic restrictions from the concept sequence layer by comparing the lexical entry of the verb with the verb elements in the concept sequence layer. Figure 16 takes Verb Subject Object Structure of concept sequence layer Structure of verb Lexical entry of verb Link S O Time Place Syntactic restriction Semantic restriction Ellipsis flag Link Lexical entry or sentence Sentence Structure of subject or object �8 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. as an example for parsing by the above three layers. Because each lexical entry of subject and object could find a path to reach their semantic restric- tions, the example sentence is legal. The parsing path of the subject is “ 「 」「 」「 」” and the path of the object is “「 」「 」「 」”. learning Mechanism of semantics In the procedure of semantics processing there could be some inconsistency between the system and the user. Because of this problem, the system should provide a learn- ing mechanism (Kim & Moldovan, 1995) for reducing the differences between the system and the user. The proposed learn- ing mechanism is divided into two parts: generalization and specialization. Ge��er���z�t���� Generalization of the learning mechanism loosens the semantic restrictions. Figure 17 shows examples of generalization: There are two conditions of generaliza- tion when differences appear between the system and the user: • The restriction layer of the system is greater and equal to the restriction layer considered by the user, as shown on the left side of Figure 17. Should the user decide that one of the oblique nodes should be corrected, the system would find the lowest common parent node (meshed node) of the oblique node and the restriction node (black node) as the new restriction. • The restriction layer of the system is less than the restriction layer considered by the user as shown on the right side of Figure 17. If the user decides that the meshed node on top of the restriction node (black node) should be corrected, this meshed node would become the new restriction node. specialization Specialization of the learning mechanism shrinks the semantic restrictions. Figure 18 shows an example of specialization. If a user decides that one of the nodes (oblique node) under the restriction node is illegal in semantics, the system would change the restriction by eliminating all the illegal nodes under the restriction node Figure 15. Structure of semantic concept hierarchy Object Entity Abstract Creature Non-Creature Animal Plant Human Non-Hunan You Me Teacher DogCat GrassFlower Stone Semantic concept hierarchy database ID Lexical entry ID of parent Structure of node Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 �� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. Figure 16. Example of semantic verification Figure 17. Generalization Figure 18. Specialization or Old restriction New restriction Old restriction New restriction New restriction Old restriction New restriction and the meshed nodes would become the new restriction nodes, as shown in Figure 18. If the illegal node considered by the user is above or the equal to the restriction node, the system would ask the user what the restriction should be. semantic Network The purpose of the semantic network (Quillian, 1968) is to store and represent the meaning of semantics so as to apply it in the inference mechanism. The semantic network describes the relationship between an object and an event. There are three ele- 80 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. ments in a semantic unit: entity, attribute, and value: • Entity: The principal part of a seman- tic unit that represents an object or event. • Attribute: An arc that describes the attribute of the entity. • Value: The result of the attribute that describes the entity. F i g u r e 1 9 u s e s t h e s e n t e n c e as an example. The meshed node is one of the nodes in the concept hierarchy layer, and the actual lexical entry in the instance layer uses the arc of the attribute ‘instance-of’ to form a semantic unit of semantic network with the concept node . With the link of the attribute ‘instance-of’ the actual lexical entry inherits the property or the capability from the concept node. The system transforms the linkage into the semantic unit subject . Consequently, the unit of semantic network is established through linkage by the syntactic parser. response system In the past, learners have had to know ex- act keywords when using a search engine. However, two identical keywords in dif- ferent contexts produce different meanings and therefore return different, sometimes undesirable, results. For example, “tell me about the specifications of Bluetooth” and “tell me about the applications of Blue- tooth,” although having the same keyword, have different meanings. The first sentence concerns hardware specifications of Blue- tooth, and the second, software applications of Bluetooth. Learners then would have to filter the data from the results by themselves. The Bayesian Network(BN) and Natural Language Understanding(NLU) (Arai, Wright, Riccardi, & Gorin, 1998; Carpenter & Chu-Carroll, 1998; Kuhn & De Mori, 1995; Miller & Bobrow, 1994; Pieraccini & Levin, 1992) are used to decipher ambigu- ous sentences and evaluate the searching preferences of different learner. Before attempting to develop a system, the application domain should be defined as it is very difficult to solve problems of uncertain domain. The application domain of the proposed system is set to the basic computer concept. As shown in Figure 20, understanding natural language queries for a specific application domain involves parsing the input query into a series of domain-specific keywords and searching for the goal of the learner’s query. A search- ing goal assumes that within a restricted application domain, there is a finite set of semantic keywords (M) as well as a finite set of searching goals (N). The searching goals SGi and keywords Ki are all binary decisions, and the keyword Ki is true if it appears in the speech. In this way, the Figure 19. An example of semantic network 我 石頭subject object 石頭 instance-of Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 81 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. proposed system can formulate the NLU problem by making M binary decisions with N BNs. The BN for the searching goal SGi takes the input as a set of keywords K extracted from the learner’s query. The BN then gives a posterior probability P(SG | K) for the binary decision. The connection of the BN assumes conditional independence among the set of keywords K, meaning that there are direct links between the goal and the concept nodes and no linkages among the concepts nodes. This is equivalent to a naïve Bayes formulation. When applying the Baysian rule the proposed system assumes that the searching goal SGi is present if P(SG | K) is greater than a threshold θ or absent otherwise. θ may be set to 0.5 for simplicity because P(SGi = 1 | K) + P(SGi = 0 | K) is equal to 1. This formula provides a method to reject out-of-domain queries(OODQ). A query is classified as OODQ when all BNs vote negative for their corresponding goals. As- suming that the searching sentence contains i keywords denoted by KWi , every keyword affects the searching goal. The searching sentence can be represented as Figure 20 and Formula 1. According to the Minimum Description Length (MDL) principle, every node in the BN provides the complexity of the network by a magnitude of Lnetwork. Lower values for the Lnetwork reflect lower network complexi- ties. Each node also provides the accuracy in modeling the data by a magnitude of Ldata. Lower values for Ldata reflect higher accuracy. In this way, the total description length Ltotal provided by the given node is defined by Ltotal = Lnetwork + Ldata. The total description length of a network is the sum of all the concept nodes in the network. The trained BN topologies is shown in Figure 21. There are two keyword groups (maximal sets of nodes that are all pairwise linked)—(SG, K1, K2) and (SG, K3) which show that the keyword groups can com- municate through the separator node SG. Each keyword group Ki relates to a joint probability P(SG, Ki). The keyword group (SG, K1, K2) relates to the joint probability P(SG, K1, K2), and the keyword group (SG, K3) relates to the joint probability P(SG, K3). Given a learner’s query, the proposed system derives the presence and absence of the various keywords K and updates the joint probability according to Formula 2. The updated joint probability is eventually marginalized to produce a probability for the searching goal P*(SG). Figure 20. Belief network of keywords and searching goal Keyword_1 Keyword_2 Keyword_n……… Searching Goal 82 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. The BN framework from NLU is ex- tended to mixed-initiative dialog modeling. The idea is to enable BNs to automatically detect missing or spurious keywords ac- cording to domain-specific constraints captured by their probabilities. If a missing keyword is detected, the BN prompts the dialog to display the necessary informa- tion to the learner. If a spurious keyword is detected, the BN prompts the dialog to notify the learner regarding the unnecessary information. Automatic detection of miss- ing and spurious keywords is achieved by the technique of backward inference which involves probability propagation within the BN. Considering the inferred searching goal SGi for a given learner’s query, the goal node of the corresponding BN is instanti- ated (to either 1 or 0) to test the network’s reliability in each of the input keywords. If the BN topology is assumes conditional independence among the keywords, the up- dated probability of the concepts would be P(Kj | SG). However, in the proposed BN in which the keywords depend on each other, the updated searching goal probability P*(Ki) would propagate to update the joint probabilities of each keyword group P*(Ki , SGi). In this way, each P *(Kj) can be ob- tained by marginalization. This procedure is described by Formula 3 and is similar to the procedure described by Formula 2 for updating concept probabilities. Based on the value of P*(Kj), the system makes a binary decision (by thresholdθ) regarding whether Kj should be present Figure 21. Trained topology of BN and keyword groups 1 2 1 2 1 2 1 2 2 1 3 2 1 ( | , ,...., ) ( , ,...., , ) ( , ,...., ) ( , ,...., ) ( | ) ( | ) ( | ) n n n i n n P SG KW KW KW P KW KW KW SG P KW KW KW P KW KW KW P KW KW P KW KW P KW KW (1) KW1, KW2, …., KWn : search keywords SG : Searching Goal Formula 1. Keyword_1 Keyword_2 Keyword_3 Searching Goal SG, K1, K2 SG, K3SG Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 83 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. or absent. This decision is compared with the actual occurrence of Kj in the learner’s query. Should the binary decision indicate that Kj is absent and it appears in the input query, the keyword is labeled spurious and the dialog would invoke an explanation. If the binary decision indicates that Kj should be present but it is absent from the query, the keyword is labeled missing and the dialog would invoke the prompting act. Using keywords and searching goals to categorize knowledge domains is not al- ways desirable as some learners just browse casually in which case a wider variety of search results is preferred. Therefore, the system records browsing habits as another factor that affects the knowledge domain. The corresponding BN is shown as Fig- ure 22 and the probability is described by Formula 4. The last part is to determine the content and degree of the response to learner. The learning results could be judged by some type of test where a high score indicates a good learning effect and a low score indi- cates a poor learning effect. The modified BN is shown as Figure 23 and the probability is described by Formula 5 and 6. eXPeriMeNtal results The implementation uses the following ex- ample sentence to describe the process of each step. Formula 2. * * * ( )( , ) ( | ) ( ) ( , ) ( )i i i P K P SG K P SG K P K P SG K P K (2) P*(K): Initialized by the presence or absence of the concepts in the learner’s query. P(SGi, K) : Joint probability obtained from the training set. P*(SGi, K) : Updated joint probability. * : Denotes an updated probability with knowledge about the presence or absence of the various concepts in the learner’s query. * * * ( )( , ) ( | ) ( ) ( , ) ( ) i i i i i i P SG P K SG P K SG P SG P K SG P SG (3) P*(SGi) : Updated from instantiating the searching goal node. P(K, SGi): Joint probability of the keyword group obtained from the training set. P*(K, SGi): Updated joint probability of the keyword group. Formula 3. 84 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. Figure 22. Bayesian network of knowledge domain Figure 23. Bayesian network of teaching material difficulty Keyword_1 Keyword_1 Keyword_n…… Knowledge Domain Hardware Software Network. . . Browse record Searching Goal Examination Result Excellent Good Poor Teaching Material Difficulty Basic Normal Hard Keyword_1 Keyword_1 Keyword_n…… Knowledge Domain Hardware Software Network. . . . Browse Record Searching GoalExamination Difficulty Formula 4. ( | , ) ( | ) ( | , ) ( | ) P BR KW SG P KW SG P KW BR SG P BR SG (4) P(KD|BR,SG): Probability of knowledge domain P(BR): Probability of learner’s browse record. P(SG): Probability of searching goal. Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 8� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. segmentation system The segmentation system divides the example sentence into lexical entries and gives each lexical entry a suitable part of speech as follows. Figure 24 shows the list branching from the first to the fourth layer. Table 4 shows the segmentation table of sentence in Figure 24 The segmentation system divides the example sentence into lexical entries and gives each lexical entry a suitable part of speech as follows: : Demonstrative pronoun [Pb] : Quantifier [Q] : Noun [N] : Transitive verb [Vt]: Transitive verb [Vt] : Adjective [Adj] : Adjective [Adj] : Noun [N] 1. : (Bs+ or Pq+) (()(Bs)) (()(Pq)) 2. : (num- or Pq- or (Pq- & num- ))&(Q+) ((num)(Q)) ((Pq)(Q)) ((Pq,num)(Q)) 3. : (S+ or O-)&(Q- or())&(@ Adj- or ())&(Do- or ())&(Ds+ or ())&(Cn1+ or ())&(Cn2- or ())(1) (({Q},{@Adj},{Do},{Cn2})(S,{Ds},{C n1})) ((O,{Q},{@Adj},{Do},{Cn2})({Ds},{ Cn1})) 4. : (Hv- or ())&(S-)&(O+)&(Advb- or()) ((S)(O)) Formula 5. ( | , ) ( | ) ( | , ) ( | ) P ED ER KD P ER KD P ER ED KD P ED KD (5) P(ER|ED,KD): Probability of examination result. P(ED): Probability of examination difficulty. P(KD): Probability of knowledge domain. Formula 6. ( | , ) ( | ) ( | , ) ( | ) P ER TMD KD P TMD KD P TMD ER KD P ER KD (6) P(TMD|ER,KD): Probability of teaching material difficulty. P(ER): Probability of examination result. P(KD): Probability of knowledge domain. 86 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. ((Hv,S)(O)) ((S,Advb)(O)) ((Hv,S,Advb)(O)) 5. : (Adj+ or Bj-)&(Adva- or ())&(Noj- or ())&(Ca1+ or ())&(Ca2- or ()) (({Adva},{Noj},{Ca2})(Adj,{Ca1})) ((Bj,{Adva},{Noj},{Ca2})({Ca1})) 6. : (Adj+ or Bj-)&(Adva- or ())&(Noj- or ())&(Ca1+ or ())&(Ca2- or ()) (({Adva},{Noj},{Ca2})(Adj,{Ca1})) ((Bj,{Adva},{Noj},{Ca2})({Ca1})) 7. : (S+ or O-)&(Q- or())&(@Adj- or ())&(Do- or ())&(Ds+ or ())&(Cn1+ or ())&(Cn2- or ()) (({Q},{@Adj},{Do},{Cn2})(S,{Ds},{C n1})) ((O,{Q},{@Adj},{Do},{Cn2})({Ds},{ Cn1})) After obtaining the above linking grammars, the system begins to parse the sentence according to the above algorithm. The linking process of the first and second lexical entries are shown in Figure 25. Because the first lexical entry contains only its relation to the second lexical entry with the linking requirement ‘Pq’, the link- ing results of the first and second lexical entries are set to true and it records the linkage ‘(Pq,1,2)’ in the linking table. The linkage ‘(Pq,1,2)’ denotes that the first lexi- cal entry connects leftward to the second lexical entry via the connector ‘Pq’. The linking result of the third lexical entry is still false as a result of having no relation- ship with the first lexical entry. The final Figure 24. List of separating from the first layer to the fourth layer Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 8� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. START STRING WORD RESULT Start at C(1) C(1) O C(1)_ C(2) X C(1)_C(2)_C(3) X C(1)_C(2)_C(3)_C(4) X C(1)_C(2)_C(3)_C(4)_C(5) X C(1)_C(2)_C(3)_C(4)_C(5)_C(6) X Start at C(2) C(2) O C(2)_ C(3) X C(2)_C(3)_C(4) X C(2)_C(3)_C(4)_C(5) X C(2)_C(3)_C(4)_C(5)_C(6) X C(2)_C(3)_C(4)_C(5)_C(6)_C(7) X Start at C(3) C(3) X C(3)_ C(4) X C(3)_C(4)_C(5) O C(3)_C(4)_C(5)_C(6) X C(3)_C(4)_C(5)_C(6)_C(7) X C(3)_C(4)_C(5)_C(6)_C(7)_C(8) X     Table 4. The segmentation table of sentence 88 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. result of the syntactic analysis is shown in Figure 26. semantic analysis system After the syntactic analysis procedure, the system can determine that the fourth lexical entry is a verb and that the third and seventh lexical entries are the subject and object, respectively, according to linkages (S,3,4) and (O,4,7). The system determines the correctness of the semantics, and the sen- tence is classified as legal if it conforms to semantic restrictions. Finally, it transforms these linkages into semantic meanings as shown in Figure 27: Figure 25. Process of syntactic analysis Figure 26. Result of the syntactic analysis • (S,3,4)subject • (O,4,7)object • (Adj,5,7)characteristic • (Adj,6,7)characteristic CoNClusioN aNd future Work This paper applies the linking of grammar to describe the syntactical construction of sentences and proceeds to verify and record the semantics according to the construction. At the end, the proposed system replies to the user by finding the suitable response derived from the response database. Ac- cording to the implementation results, the proposed system could correctly describe I D Lexical entry Linking result Disjunctions (()( Pq )) I D Lexical entry Linking result Disjunctions 1 這 True 2 個 True (( Pq )( Q )) (( Pq , num )( Q )) I D Lexical entry Linking result Disjunctions 3 處 理 器 False (( { Q } ,{ @ Adj } ,{ Do } ,{ Cn 2 } )( S ,{ Ds } ,{ Cn 1 } )) (( O ,{ Q } ,{ @ Adj } ,{ Do } ,{ Cn 2 } )( { Ds } ,{ Cn 1 } )) Pq Linking table ( Pq , 1 , 2 ) Linking result : The used disjunctions of each lexical entry : 這 個 處 理 器 有 許 多 新 的 功 能 Pb Q S O Adj Adj (()(Pq )) ((Pq )(Q )) (()(S)) ((S)(O )) (()(Adj )) (()(Adj )) ((O,@Adj )()) （Pq,1,2）、（Q,2,3）、（S,3,4）、（O,4,7）、（Adj,5,7）、（Adj,6,7） Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 8� Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. the relationship between lexicons. Fur- thermore, with the use of memory-based parsing, the proposed system can check the correctness of the semantics and provide a learning mechanism by changing the se- mantic restrictions of the concept sequence layer. Finally, by classifying the response databases and using the verb of the sentence as the search key, the response system greatly reduces the amount of response results and increases speed and accuracy. Of the numerous possible applications in the future, one could further develop the interaction between computer and user by utilizing user characteristics and habits, allowing the proposed method to generate a variety of improved responses. With the continual progression of the computing age and the increasing need for improved methods of automated communication, the future of the proposed system remains both worthwhile and practical. refereNCes Arai, J., Wright, G., Riccardi, & Gorin, A. (1998). Grammar fragment acquisition using syntactic and semantic clustering. The 4th International Conference on Spoken Language Processing. Carpenter, B. & Chu-Carroll, J. (1998). Natural language call routing: A robust, self-organizing approach. The 4th International Conference on Spoken Language Processing. Chen, J.-F., Lin, W.-C., Jian, C.-Y. (2003a). Using the keyword in context segmentation method for collaborative design in a Chinese Web site. The 10th ISPE International Confer- ence on Concurrent Engineering: Research and Applications, pp. 967–975. Chen, J.-F., Lin, W.-C., Jian, C.-Y., Ho, T-.Y., & Dai, S.-Y. (2003b). Using the keyword in context segmentation method for a Chinese Web site. 2003 International Conference on Computer- Assisted Instruction, National Taiwan Normal University, Taipei, Taiwan, pp. 74-80. Chen, J.-F., Lin, W.-C., Jian, C.-Y., & Hung, C-.C. (2003c). A Chinese automatic interactive feedback system for applying in a Web site. The Second International Human.Society@Internet Conference, pp.238-248. Chen, J.-F., Lin, W.-C., Jian, C.-Y., & Hung, C-.C. (2005). A Chinese interactive feedback system for an e-learning Web site. Journal of Information Science and Engineering, 21(5), 929-957. Chen, K. J., & Liu, S. H. (1992). Word iden- tification for mandarin Chinese sentences. Proceedings of the Fifteenth International Conference on Computational Linguistics, Nantes, pp.101-107. Chung, M. & Moldovan, D. (1993). Parallel memory-based parsing on SNAP. Parallel Processing Symposium, Proceedings of Seventh International Conference, pp. 680-684. Chung, M. & Moldovan, D. (1994a). Applying parallel processing to natural-language process- Figure 27. Semantic network of example sentence 有處理器 功能subject object 許多,新的 characteristic �0 Journal of Distance Education Technologies, 6(4), 62-�0, October-December 2008 Copyright © 2008, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. ing. IEEE Expert [see also IEEE Intelligent Systems], 9(1), 36-44. Chung, M. & Moldovan, D. (1994b). Memory- based parsing with parallel marker-passing. Pro- ceedings of the Tenth Conference on Artificial Intelligence for Applications, pp. 202-207. Kim, J-.T. & Moldovan, D.I. (1993). Acquisition of semantic patterns for information extraction from corpora. Proceedings of Ninth Confer- ence on Artificial Intelligence for Applications, pp.171-176. Kim, J-.T. & Moldovan, D.I. (1995). Acquisi- tion of linguistic patterns for knowledge-based information extraction. IEEE Transactions on Knowledge and Data Engineering, 7(5), 713-724. Kuhn, R. & De Mori, R. (1995). The applica- tion of semantic classification trees for natural language understanding. IEEE Trans. Pattern Anal. Machine Intell., 17, 449-460. Miller, S. & Bobrow, R. (1994). Statistical language processing using hidden understand- ing models. The Human Language Technology Workshop, 278-282. Pieraccini, R. & Levin, E. (1992). Stochastic representation of semantic structure for speech understanding. Speech Communication, 11, 283-288. Quillian, M.R. (1968). Semantic Memory. In Semantic information processing, pp. 216-270. Cambridge, MA: MIT Press. Sleator, D. & Temperley, D. (1991). Parsing English with a link grammar. Carnegie Mellon University Computer Science technical report CMU-CS-91-196. Jui-Fa Chen (陳瑞發) received his PhD, MS, and BS degrees in the department of computer science and information engineering from TamKang University (TKU), Danshui, Taipei, Taiwan, in 1998, 1992, and 1990, respectively. He is an assistant professor in the Department of Informa- tion Technology in TamKang University (TKU). His research interests include intelligent avatar, peer-to-peer communication, and software engineering. Wei-Chuan Lin (林 偉 川) received his PhD, MS, and BS degrees in the department of computer science and information engineering from TamKang University (TKU), Danshui, Taipei, Taiwan, in 1998, 1986, and 1984, respectively. After graduated from TKU, he worked in the Institute of Information Industry until 1993. He is an assocated professor in the Department of Information Technology in Takming College, Nei-Hoo District, Taipei, Taiwan, since 1993. His research interests include intelligent avatar, peer-to-peer communication, and software engineering. Chih-Yu Jian (簡志宇) eceived his PhD, MS, and BS degrees in the Department of Computer Science and Information Engineering from TamKang University (TKU), Danshui, Taipei, Taiwan, in 2007, 2001, and 1999, respectively. His research interests include intelligent avatar, peer-to- peer communication, and software engineering. Ching-Chung Hung (洪慶全) received his MS and BS degrees in the Department of Computer Science and Information Engineering from TamKang University (TKU), Danshui, Taipei, Tai- wan, in 2002, 2000. After graduated from TKU, he worked in Internet Information Corporation until now.