Special Feature: system's slowest response time (0.24 sec) with the quickest manual response time, the system is at least 130 times

Transcription

1 Special Feature: An Automated Chinese Telephone Directory Y. H. Chin, J. W. Jou, W. H. Peng, and C. C. Yang Telecommunication Laboratories of Taiwan Introduction A real-time, on-line information storage and retrieval system for telephone information service has been designed at the Telecommunication Laboratories of Taiwan. The input and output of the system are processed in Chinese via a keyboard and a graphic display unit respectively. System functions include sorting, merging, updating, displaying, and printing of data. In order to avoid a serial, exhaustive search, the file is subdivided into blocks of length x. The optimal size of each block is determined so as to minimize the average number of comparisons (file accesses) to locate a record. The system has been tested on a telephone office of medium capacity-about 8000 subscribers. The number of daily queries is about 1600 and the updating rate is approximately 50. On the basis of a comparison of the May 1975 system's slowest response time (0.24 sec) with the quickest manual response time, the system is at least 130 times faster. The design philosophy stresses fast responses to queries and quick updating of the data. For a computer, a Chinese character is not a character but a pattern; therefore, a portion of the coding is related to Chinese. System Description The system hardware (see Figure 1) consists of an HP/2100A computer under the control of the Moving-Head Disk Operating System (DOS-III). Two secondary storage devices are attached: a moving head disk unit and a magnetic tape unit. The disk unit is used to store the system programs, utility programs, and data files. The data 49

2 PHONETIC SYMBOL KEYBOARD GRAPHIC DISPLAY TAPE READER ELECTROSTATIC PRINTER ~~~~~~~~~~~MAGNETIC DISK TAPE Figure 1. System Block Diagram of the Automatic Telephone Directory Inquiry Services System KEY TABLE key record number TELEPHONE DIRECTORY F ILE record 1 record CHINESE CHARACTER PATTERN FILE (dot matrix) record 1 record m record k Displaying (a) Data Base Structure and Searching Sequence of Telephone Directory Data Base LI user's name 16 s Tel. No. I pointer I 3 s 2 s (b) Record Format for Telephone Directory File I Figure 2. Logical Structure of Data Base files contain the telephone directory file and the Chinese consonants and 16 vowels. The phonetic spellings are character pattern file. The magnetic tape is used as a translated into a string of numerical codes which is used as backup unit. The addresses of the telephone subscribers are a search key to locate the desired record and to display it in stored on the tape for printing of the telephone book.* Chinese. Input devices consist of a paper tape reader, used as a main input device for reading utility programs and data *In general, an operator finds a user's phone number whenever a, keyboard, Muser's name is given in a query. Therefore, the operator at the X which is used to type thephonet icsym spellingseof the service center has no need to find a phone number through a home i t which is used to type the phonetic spellings of the input address of a particular subscriber. In fact, the subscriber's address Chinese characters,4 contains 37 phonetic symbols-21 is confidential at the Query Service Center. 50 COMPUTER

3 Output devices include an HP/1331C X-Y graphic display, which is used to display the desired telephone record in Chinese, and an electrostatic printer (a VERSATEC plotter) for printing the telephone book in Chinese. The speed is about two lines/sec in Chinese characters. Depending on the size of the printed Chinese character, each line can contain 40 to 80 Chinese characters. Except the DOS-III system program, we developed our own utility programs for handling data processing. The functions of these programs are searching, sorting, merging, converting, updating, displaying, printing, and plotting. Data Base Structure In this system we have two data bases: a telephone directory data base and a Chinese character pattern data base. The telephone directory data base is a three-level indexed sequential file containing three files: (1) index table, (2) key table, and (3) telephone directory file (Figure 2a). The record format of the key table has 6 bytes for the search key, 2 bytes for the corresponding record number in the telephone directory file, and 2 bytes for a pointer used for updating. The variable-length record format of the telephone directory file (Figure 2b) consists of three fixed-length fields: a 32-byte name field, which is a variable-length field containing 1 to 16 Chinese character codes, a 6-byte telephone number field, and a 4-byte pointer field for printing the telephone- book. Since each sector (256 bytes) can store six 42-byte records, it is very easy to calculate the actual location in the disk. Each record is assigned an unique record number. Through this number, the physical address of a specified record can be identified with the following rule: sector = starting sector of the file + (S 1)/R (1) record number in the sector remainder of (S - 1)/R... (2) where S is the record number of the specified record and R is the number of records in each sector. In the key table file, the entries are stored in ascending order with respect to their numerical value of search key as well as record number. In order to minimize the mean number of comparisons to locate the desired record, the key table file is subdivided into blocks of length x. The optimal block size x is determined with respect to combined search methods so that the mean number of comparisons is minimized. In the key table file, each block is initially 90% filled, with the remainder reserved for overflow. After a period of time the key table is reorganized to avoid filling up the reserved overflow space. In case of filling up, an overflow signal is given, and a new key table and 10% reserved block space are automatically regenerated. Therefore, the logical structure of the data base does not need to be changed when the system is applied to a larger city like Taipei, which has 260,000 telephone subscribers. Every Chinese character is represented by an 18 X 15 dot matrix and is assigned an unique number as its code-i.e., its record number (see Figure 3). Hence each May 1975 w t v _ ~r I JL L I _-- _4 PI _I~_ Chinese character expressed in an 18 x 15 dot matr'ix Figure 3. record format: I -_-k dot matrix 36 bytes memory data of the dot matrix Chinese Character Pattern and the Record Format Chinese character needs 36 bytes for its dot matrix representation. There are 1400 Chinese characters in use in our present system. These are stored in ascending order with respect to frequency of usage, so that characters with high usage can be accessed (i.e., made core-resident) very rapidly.3 These 1400 character representations form a file called the Chinese character pattern file. To access a Chinese character, simply get this character's code and set R = 7 in formulas (1) and (2). Then, the location is found and the desired dot matrix of the character is displayed. This operation takes, at most, one disk access. The reasons for using a dot matrix rather than using other forms, such as the Chiao-Tung Radical System,5 are that it requires a smaller number of disk accesses and permits an elegant and symmetrical presentation of the Chinese character in display and print. (Typical video and hardcopy outputs are shown in Figures 4 and 5.) Relationship Between Search Methods and Optimal Block Size To locate a desired record Ri from the file, the search sequence is as follows. A user's name is first translated into phonetic spellings which are typed by an operator into the system through a keyboard; then the spellings are translated into a string of numerical codes which are used as a search key. (This input operation is illustrated in Figure 6.) The search key is first compared with each entry in the index table of n/x entries in order to locate the block, say Bj, where the desired record Ri is stored. In case of a match, each code of these Chinese characters in the Ri must search for the Chinese character pattern file in order to find out its corresponding dot matrix. From the dot matrix, this desired record Ri can be displayed and printed in Chinese. In case of mismatch, an error message will be sent to the operator. Depending on the received information, the operator can either retype the query or inform the requester to repeat his message. 51

4 E_ W _ - _ t ; GC a. b. c. d. Figure 4. The displayed result: (a) multiple-phone number record; (b) variable-length record; (c) subscribers with same name are distinguished by their addresses inside the parenthesis; (d) updated record shown in the second row. Among the various methods known on retrieving information records from a data base, we have chosen ISAM (indexed sequential access method) and binary search method in our application. For the time being, the other methods are unsuitable for our case. For example, the difficulties in using scatter storage techniques are (1) finding satisfactory hashing function so that, after hashing the phonetic spellings of Chinese characters, the search key can be uniformly'distributed in each block of the key table; and (2) estimating the suitable storage allocation scheme for the hashing table, because the number of telephone users is increased rapidly. This problem is particularly serious for a minicomputer with only 16K core memory. On the other hand, linear search is time-consuming. Therefore, it contradicts the requirement of real-time reply on inquiry. In order to minimize the mean number of comparisons to locate the record Ri from the block Bj, the problem of determining the optimal block size x with respect to search methods is solved as follows: (1) For a sequential search in the index block and the selected block Bj, the optimal block size is n1/2; (2) For a sequential search in the index table and binary search in the block Bj, the optimal block size is n.ln2/2; (3) For a binary search in the index table and the selected block Bj, the optimal block size is n/2; (4) For a binary search in the index table and sequential 52 search in the selected block Bj, the optimal block size does not exist; where n is the number of records in the telephone directory.** **Detailed proof: An ordered file of n records (either in fixed length or in variable length) is subdivided into blocks of length x. There are n/x blocks, each having x records. The optimal block size x with respect to various search methods is proved as follows: (1) For a sequential search in the index table and in the block 13the optimal block size is ny2. Proof: The mean number of comparisons to locate a record from the file is + t)(n/x)/(n/x) + l/2x(x + 1)/x /2((n/x) + 1) + /2(X +) =a 1/2((n/x) = di/dx = O e (-n/x2) x = n/2 d2i/dx2 2n/x3 = 2/nl/2 >0-O xmin x=n/2 = n'/2 (2) For a sequential search in the index table and binary search in the block Bi, the optimal block size is (n.ln2)/2. Proof: a =1/2(n/x + 1) + ((x + 1)(log2(x + 1))/x - 1) 1/2(n/x + 1) + (log2(x + 1) - 1) (See Reference I for x>10) da/dx= 0x =(n.ln2)/2 d2n/dx2 4/(ln2)3n2 >0 -xmin x= = (n.ln2)/2 (n.ln2)/2 COMPUTER

5 BWt (E rr3 ft r-3 t ta pt Io1\j& OIV il IRil 1kAEf# ftgb r,s Eb!X ti',l4 31U %B , ID I f k I e+3 tgn,44 ES L r th t- " >: X-x PCI. Vcw*Yf.T re' I P~P-q3-,, ZM, L 7 NI- L.tL =Ee W i8 b5939 3S S i ? DU f39 293'(' 'ujl-h3 QI3 ES~i , Figure 5. The Printed Result. The 5-digit numbers represent continued phone numbers-e.g., represents two telephone numbers 3163 and (3) For a binary search in the index table and the block Bj, the optimal block size is n. Proof: a = (log2(n/x + - 1) 1) + (log2(x + 1) - 1) da/dx= 0 x = n/2 d2i/dx2 = 2/nt/2(nl/2 + 1)2 >0 1x=n1/2 Xmin- n/2 (4) For a binary search in the index table and sequential search in the block Bj, the optimal block size does not exist. Proof: a = (log2(n/x + 1) - 1) + 1/2(x + 1) da/dx=0 -> x=0or-n This is impossible since x is a positive integer. In all, of these four expressions, the first term and the second term represent the formula of the mean number of comparisons in May 1975 Figure 6. User's name in Chinese (given by an inguirer) Phonetic spellings (typed by an operator on a keyboard) S{ 7 2 T- 1 <-72 Translated numeric value (system generated) Actual search key (starting to search) Typical Example of Input Operations for Searching a Telephone Record the index table and the selected block Bj respectively. For large n (e.g., n > 1000), the value of a in (2) and (3) is always less than the value in (1) whenever x = xmin; therefore, the binary search should be adopted in either the index table or the block B- or in both. 53

6 System Testing To insert a new record into the system, the whole information of a telephone record is first stored at the bottom of the telephone directory file. Then the record's search key (the phonetic spellings of the selected characters from the user's name) is stored at the appropriate location in the ordered key table file. This operation takes two disk accesses: one for writing the search key into the key table and one for writing the record in the telephone directory file. It takes an average of 10 seconds to key in a telephone record of 8 Chinese characters, which includes the character codes, the phonetic spellings of the search key, and the telephone number. Hence, the oiine insertion operation is not used in the testing system with one CPU during the busy hours (9-12 A.M. and 2-5 P.M. daily2). The modification is simple: the record to be modified is first fetched and displayed on the graphic display (see Figure 5); then the information needed to be added or deleted can be processed similarly to an edit operation. The system's regenerating time, defined as the time of regenerating the key table, equals b(2dt + pt)/n where b is the number of blocks in the key table, dt is the average disk access time, Pt is the CPU time, and n is the total number of records in the telephone directory file. Now, the system's regeneration time (i.e., rebuilding cost) is approximately 0.7 milliseconds per record. The total size of our utility programs is about 6K bytes, and the size of the data file is about 313K bytes in which 210K bytes are used for the telephone directory file, 50K bytes for the key table, and 53K bytes for the Chinese character pattern file. The total size of the system is about 320K bytes. The response time to display a record in Chinese depends on the number of Chinese characters in the retrieved record. On the average, it takes 30 milliseconds (the average access time of an HP/7900 disk) to access a Chinese character. In the worst case, it takes 0.24 sec to display a record of 8 Chinese characters on the HP/1331C graphic display. By comparison of the slowest system's response time (0.24 sec) with the quickest manual response time, when some of the most frequently used Chinese characters (about 100) are made core-resident, the system's response time will be improved to 200 times faster than that of a manual system.3 Acknowledgment The authors would like to acknowledge the corrections and suggestions of Dr. R. C. T. Lee and Mr. Pierre Loisel. They are also indebted to the referees' useful comments and colleagues' discussions in the Laboratories. References 1. G. Salton, Automatic Information Organization and Retrieval, McGraw-Hill Book Company, N. Y., C. C. Yang, "Statistic Analysis on the Traffic of Taipei Query Service Center," Quarterly Report of the Telecommunication Labs, Vol. 3, No. 4, pp , October C. V. Ramamoorthy and Y. H. Chin, "An Efficient Organization of Large Frequency Dependent Files for Binary Searching," IEEE Trans. on Computers, October 1971, pp S. K. Chang, C. S. Chiu, M. H. Yang, and B. S. Lin, "PEACE-A Phonetic Encoding and Chinese Editing System," Proceedings of the First International Symposium on Computers and Chinese I/O Systems, Academia Sinica, Taipei, Taiwan, R. 0. C., Aug , 1973, pp C. C. Hsieh, M. W. Du, et al., "The Chiao-Tung Radical System," Proceedings of the First International Symposium on Computers and Chinese I/O Systems, pp Yeh-hao Chin is an assistant professor in the Computer Sciences Department at Northwestern University. Earlier, he was with the Telecommunication Laboratories in Chung-Li, Taiwan, where he was responsible for the design and development of software systems for telephone service systems and electronic switching systems. During this period he was also a,,v! X i research fellow at the Electronic Laboratory of UC Berkeley, and an adjunct associate professor in the Computer Science Department of the National Chiao-Tung University. He received the BSEE from the National Taiwan University in 1966 and the MS and PhD degrees in electrical engineering from the University of Texas in 1970 and Dr. Chin's research and teaching interests are in the design and development of data base systems for Chinese input and output. t Ld+_ Jun Wun Jou is a research scientist at the Telecommunication Laboratories in Taiwan where he leads a group in designing an automatic telephone directory inquiry service system for the Taipei area. Jou received his MS degree from National Chiao-Tung University, Hsinchu, Taiwan, in He is a member of the Institute of Electrical Engineering of the Republic of China. System Improvement At the present, each operation takes a certain amount of disk accesses. The disk access includes not only record retrieval but also location of the Chinese characters in retrieved record-especially the time spent in locating the Chinese characters. In order to reduce the disk access time, the usage frequency of each Chinese character is being monitored so that those characters used most often can be made core-resident. Also, some coding methods such as threshold functions for query will be adopted to facilitate a user's query. The "best" threshold value is studied so that whenever a communication ax4iguity occurs, the misspelled phonetic input will still " e a correct answer (i.e., get the desired record). In summary, we plan to do the following in the near future: (1) make the input query format more flexible, (2) reduce the number of disk accesses, (3) modify and extend the operating system when the system is working in a time-sharing environment with multiple terminals. u 54 W. H. Peng works for the Telecommunication Laboratories in Chung-Li, Taiwan, where he is engaged in research programs concerning information storage and retrieval. He received the BSEE and MSEE from the National Chiao Tung University, Hsinchu, Taiwan, in 1968 and Chen Chau Yang is a research scientist in the Computer Scientist Group of the Telecommunication Laboratories, where he is in charge of the design and implementation of an system at the Taipei Query Station, Yang received the BSEE and MSEE from National received the BSEE and MSEE from National Chiao-Tung University, Hsinchu, Taiwan, in 1969 and 1971, and is a member of the Institute of Electrical Engineering of the Republic of China. His current research interests are in the field of information storag(e and retrieval. COMPUTER