TW420774B - Method and apparatus for automatically correcting documents in chinese language - Google Patents

Abstract

There are provided automatic correction method and apparatus for automatically detecting and correcting wrong characters and/or missed characters in documents written in chinese language. Character to reading conversion section 200 converts an input original document to strings of phonetic symbols. Candidate word detection section 300 divides the strings into syllables and detects possible candidate words and related information using syllables as search keys. Similar candidate word detection section 400 detects possible candidate word and related information using the string of phonetic symbols as searching key wherein similar bits have been masked with mask means. Most optimal candidate letter string determination section 500 links individual candidate words to form an oriented net by using start and end positions of each candidate word corresponding to the original document as search keys and takes out a most optimal route by the dynamic design method with use of a maximum integral value of weight for frequency of use plus weight for word length plus weight for similarity of original document plus weight for similarity of meaning. Matching section 600 performs matching between the letter string of the above optimal route and the original document to detect and mark different letters.

Description

Printed by the Central Standards Bureau of the Ministry of Economic Affairs, Shellfish Consumer Cooperative, 42C774 A, _________ B " ____ V. Invention Description (1) [Industrial Use Field] The present invention relates to a computer that automatically detects and corrects typos in Chinese documents or Method and device for automatically correcting Chinese documents with dropped characters. [Past technology] When ancient Chinese wrote literary documents, the causes of typos can be roughly divided into the following: (―) Homophonic or homonymous characters such as' [young seedlings look like pens] [Get], ordinary people It's easy to mistakenly write it as "·". In addition, [Extreme] of [Extremely Excited] can be mistakenly written as "Urgent" by accident. (2) Stroke error For example, in [Cap] of [Hat], accidentally "taking" can easily be mistakenly written as Chang et al. Or the word "目" is written as "Yue". In addition, complex characters such as “turtle” and “yu” are often written with fewer strokes. (3) The fonts are similar. For example, [Jai] in [Prime Minister], if people accidentally report that Gui Yi has neglected a point on Bao hijab or mistakenly wrote “Xin” as “Xing”. In addition, the "noisy" and "fighting" of "Noisy Fujian" are often mistakenly written as "Gate" or "Cat". The radicals are often written as "Dog", etc. When writing speed is too slow or careless, words are often lost, for example, [Tough Hard] is mistakenly written as [Tough Hard Paper Size Common Chinese National Standard (CNS > A4 Specification (2'297 G $)) ( Please read the notes on the back before filling in this page) Order I.4 Employee Cooperative Cooperative System of the Central Standards Bureau of the Ministry of Economic Affairs 420774 A 7 _____________________8, V. Description of Invention (2) ~】 etc. The word is misused. For example, [family] is often mistakenly written as [jiating] or [tingtingyuli] is often mistaken as [tingciyuli], etc. In recent ten years, due to the advancement and popularity of electronic computers, Chinese The input method also follows the controversy of hundreds of people. According to their editing methods, the Wenlun method can be roughly divided into: general keyboards and special input devices. The general key # types are as follows: (a) pronunciation marks (b) roots (c) ) Shape and sound combination (four) code (five) The first or the number of days, etc .; special input devices are as follows: (a) dedicated large keyboard (b) optical text recognition devices, etc. Although the Chinese input method relieves the Chinese from the tediousness of writing Chinese characters, but the computer is good for word input In the Chinese document building case, although some common mistakes can be avoided, such as stroke errors; other errors cannot be avoided. Generally speaking, the causes of typos in Chinese document files can be roughly divided into It is as follows: (1) The pronunciation marks and radical combinations of Chinese characters cannot be input correctly. The pronunciation marks of Chinese characters can be generally divided into four parts, including initials, finals, intermesons, and tones, as shown in the following example. Sound ·· 勹 （ b), female (P), door (m), · ·. Preposition: one (I) '乂 (u)' u (yu, iu) vowel: aya ⑷ '乙 (ο)' 死 (e), ..... Tone: ― (one beep) '/ (two beeps), M three beeps), (four beeps),' (light beeps) This paper size applies the Chinese National Standard (CNS) A4 specification (210X29? Mm) ) J ----------- install ------ order ------ line {Please read the notes on the back before filling (This page) 420774 A, B7_ V. Description of the invention. (3) '~ For example: [Shape] is pronounced as [丁 — 厶 /] [Word] is pronounced as [卩,] where' Chinese Valley is easy to be confused. The pronunciation pairs are as follows: Initial part: [corpse] and [二] or [<] and [丁] and other intermediary parts: [一] and [u] Final consonant part ... [B] and [―] Or [Bow] and [Tai] and other tonal parts: It is easy to be indistinguishable, especially for foreigners. For example, '[Interest] (Ding Yi Er, < u,) is easily misinterpreted as (Ding Yizhang, < —,), so they are often turned into "sexual organs, [Student] (TU ^ / Corpse-J-often with [Sketch]] (Ding Yigan V, Zhe Er, misused each other. In addition, when inputting glyphs, the input radical combination is similar or the combination is wrong may cause incorrect input. For example, [Yue] and [Yue] or [Jiu] and [Love] are very similar to the root code of the word combination. (2) Homonyms Wrong choice of homonyms or words printed by the staff consumer cooperative of the Central Standardization Bureau of the Ministry of Economic Affairs As a result, for example, [homonyms] are often incorrectly entered as [homonyms] or [homonyms]. (3) Errors in the reference dictionary No matter what input method is used, a reference dictionary must be used to perform the conversion. Therefore, if there is an error in the contents of the reference dictionary, the input result will also be wrong. For example, if [Shade Shadow] in the reference dictionary is registered as [Shade Shadow], as long as the above paper size is entered, the Chinese National Standard (CNS) A4 is displayed (210X297 male f). ^ 420774 Α7 _ B? 5. Description of the invention (4) The pronunciation of the narration phrase must have a wrong conversion result. (4) Incorrect input operation Generally, all kinds of document editing software are used to make documents. However, when performing editing functions such as [Insert] or [Delete], often one word (redundant) is often edited due to careless editing. One word or less. With the increase of typos in Chinese document files, which seriously affects the quality of the documents, how to effectively detect errors in Chinese document files and automatically correct them has become an important issue. A conventional method and device for automatically correcting Chinese typos, such as described in the Republic of China Patent Gazette Bulletin No. 260772. The system block diagram of the embodiment is shown in Figure Π, 100 is an input device for inputting a Chinese document to be processed; 110 Is a Chinese document file for storing Chinese documents to be processed; 12 is a comprehensive approximate glyph replacement device that replaces each character in the input Chinese document with similar characters with reference to the comprehensive approximate character set for combining into a plurality of candidate strings; 130 It is a comprehensive set of approximate characters that stores Chinese characters with glyphs, phonetics, meanings, or characters similar to the input code, for example, as shown below (s •• Font approximation, p: Fonal approximation, M: Finality approximation ,! approximate)

People: Enter S, force P, P, S, S, S, S, S, B, S, P, S, P, S, S, S, S, S, S, S, S, S, S, S, S, and S Specifications (2 丨 〇χ 297) (Please read the precautions on the back before filling in this page)

Printed by Cooperative of Staff of Central Standards Bureau of Ministry of Economic Affairs 420774

AT ~ ----- B7__ V. Description of the Invention (5) 140 is a s model model scoring device that scores each candidate string and finds the highest rated candidate string; 150 is a language model that contains (a) Statistics-the shell database, which records the frequency of occurrence of each language unit and the frequency of successive occurrences between language units, and can include a Chinese knowledge base that records the frequency of each word, (b) a scoring device, based on a string Contained language unit and language model statistics database, a scoring device that evaluates the score of the word lift and deducts points for non-original documents; 160 is the highest score candidate for searching the southernmost score candidate string by dynamic programming String search device; 17 is a typo judging device that compares the highest-scoring candidate string with the text in the document word by word and indicates that the different text is a typo; 80 is the marked The text string is output to the marked result in the marked file. 190 is the marked file that stores the marked text string. The processing action of the conventional Chinese typo detection and correction method is that the input device 100 inputs a Chinese document to be processed from the Chinese file U0 and divides it into a plurality of processing units according to the position of the punctuation marks, and then enters a comprehensive approximate glyph substitution. The device 120 then refers to the comprehensive approximate word set 130 with each processing unit, takes out all approximate words with similar glyphs, phonetic meanings, word meanings, or round-in codes, and combines them into a plurality of candidate words_ ', then enters the language model scoring device 140, A statistical language model is used to give a score to each candidate string in the device "50", where the language model score deducts the text score of the non-original document, and the dynamic string is used in the highest-scoring candidate string search device. 60 Search for the highest-scoring candidate string; then enter the typo judging device 170 'to compare the candidate string with the highest score with the input original document building word-by-word' and mark the different characters as typos, and then enter the scale of the paper Appropriate Financial Compensation Standard (CNS) A;! Specifications (2 丨 Qx 297) -9- ---------- I. Equipment -------- Order ----- -^ ** (Please read the notes on the back before filling out this page) Printed by the Central Standards Bureau of the Ministry of Economic Affairs and Consumer Cooperatives 420774 * A 7 -------______ B7 V. Description of Invention (6) -—— · Marking Results In the output 18, the marked result is output to the marked file slot ⑽ t. [Problems to be Solved by the Invention] The above-mentioned conventional techniques have the following problems: (1) As long as the comprehensive approximate word set is used, the words that are not included cannot be correctly corrected and corrected, so it must take a lot of manpower and Material resources to build and maintain this knowledge base. (2) In the language model scoring device order, only the frequency of occurrence of each candidate unit and the frequency of successive occurrences between candidate units are considered without using semantic information, and the detection rate and correction rate cannot be further improved. (3) The problem of missing words and redundant words in Chinese documents cannot be effectively solved. [Methods used to solve problem points] In order to solve the above problem points, the first invention of the application is an automatic correction method for Chinese documents using computer technology, which is characterized by including the following steps: v 羯 1 in computer memory Construct a broken dictionary that stores ordered list pronunciations, phonetic notation tables corresponding to word order tables, candidate words corresponding to phonetic notation tables, and their pronunciation marks; constructs a memory that stores text marks and their initial settings, and other possible pronunciation marks Phonetic dictionaries of pronunciation; Construct a phonetic dictionary that stores pronunciation marks and their corresponding all homonyms and their related frequency weights and semantic codes in memory; This paper scale applies Chinese National Standards (CNs.) A4 specifications (2iOX297 public Chu) (Please read the notes on the back before filling in this hundred c

Printed by the Employees' Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs 420774 V. Description of the Invention (7) ~~~~~-~ Step 2 The above-mentioned broken pronunciation dictionary and phonetic dictionary will be converted from the original document string input by the input device into a pronunciation symbol string The word-to-phonetic conversion of the column is almost a paragraph; after the syllables are extracted from the string of pronunciation marks obtained by the word-to-phonetic conversion method in Step 3, the syllables are referenced by the syllables as the search elements. Candidate word selection means detected by its related information; Step 4: U uses concealing means to mask the similar digits of consecutive early word candidate syllables as a search element, and refers to the phonetic word dictionary for possible candidate words. Words and their related information detected similar candidate word selection means; Step 5 uses the candidate text corresponding to the beginning and end of the original document string as the index element to link each candidate word into a directed network and use After calculating the similarity weighting and word length weighting of each candidate word, the sum of frequency weighting, word length weighting, and similarity weighting of the original document is used as the evaluation function. Method to determine the best candidate string for the path with the best evaluation score; Step 6 uses the text string on the best path to compare with the original document string to detect the different words and mark them. In addition, the invention in item 2 of the scope of patent application is an automatic correction device for Chinese documents using computer technology, which is characterized in that: the word order table reading should be stored, and the corresponding word order table should be stored. The phonetic notation table and the corresponding paper size are applicable to China National Standard (CNS) A4 specifications (210X 297 mm) -11------------------------ ^ --- 4 (Please read the notes on the back before filling this page) 420774

AT —__ ___ V. Description of the invention (8) A dictionary of candidates for the pronunciation sibling list and its pronunciation marks; a pronunciation dictionary containing text marks and their initial settings, and other possible pronunciation marks; A phonetic dictionary of all homonyms and their related frequency-weighted and semantic code; refer to the above-mentioned broken dictionary and phonetic dictionary to convert the original document string rotated by the input device into a phonetic transcription conversion unit ; After the pronunciation is performed using the word-to-phone conversion method, the pronunciation is executed, and the candidate word selection unit that detects possible candidate words and related information by referring to the syllable dictionary with reference to each syllable as a retrieval element; A sequence of pronunciation marks after concealing similar bits by using a masking method for consecutive single-character candidate syllables is used as a search element, and a similar candidate word selection unit that detects possible candidate words and related information by referring to the phonetic word dictionary; The word position corresponding to the start and end of the original document string is used as an index element to link each candidate word into one Directed network uses calculation methods to calculate the similarity weighting and word length weighting of each candidate word, and uses the sum of frequency weighting, word length weighting, and similarity weighting of the original document as the evaluation function to find the best by dynamic programming. The best candidate string determination section of the path for evaluating scores; a reference comparison section that compares the character string on the optimal path with the original document string and uses the comparison method to detect and mark different words. Furthermore, the invention in the third item of the patent application scope is an automatic correction device for Chinese documents according to the second item, which is characterized by: constructing a new paper ruler in the memory that has been learned through learning methods (CNS) ) Λ (21 OX 297 ^ ¾ Factory --- 420774 Printed by the Consumers Cooperative of the Central Standard Umbrella Bureau of the Ministry of Economic Affairs B7 V. Description of Invention (9) 'Word meaning code and its pre-word meaning code combination semantic dictionary for learning dictionary; above The best candidate string determination unit uses the position of each candidate word corresponding to the start and end of the original document string as an index element to link each candidate word into a directional network and calculates the similarity of each candidate word using calculation means After degree-weighting and word-length weighting, the sum of frequency-weighting, word-length weighting, original document similarity-weighting, and semantic similarity-weighting of the learning dictionary with reference to semantics is used as an evaluation function to find the path of the best evaluation score in a dynamic programming manner. [Function] According to the structure of the automatic correction device for a document of the present invention, the word-to-sound conversion means inputs the original document character _ from the input unit and refers to the broken character. Dian and Yuyin Dictionary convert the original document string into the pronunciation symbol string. 'Use the candidate word selection method to refer to the phonetic word dictionary to detect all possible candidate words. Then use similar candidate word selection method for continuous single-word syllables. The candidate uses similar sound masking means and refers to the phonetic dictionary to detect similar candidate words, then uses the best candidate string determination means to find the best path word _ for the learning dictionary with reference to semantic meaning, and then uses the reference comparison means to convert the original document word The typo in the string is marked and the result is rounded off. [Example] In the above, the semantic meaning is the meaning (or semantic code) of the morpheme itself. In this example, the dictionary used by Kakugawa Bookstore (1985) is used. The appropriate classification method is shown. Its classification method uses a hexadecimal four-digit code to represent a morpheme. The leftmost digit indicates a large classification, the second digit indicates a medium classification, and the third digit indicates a small classification. Classification, the rightmost digit is Ί * L ---------- Shang ------ Order ------ (Please read the precautions on the back before filling this page) 420774 Λ7 B7 2. Description of the invention (10) means fine classification. This lexical dictionary divides all morphemes into ten categories: nature, character, change, throbbing, mood, character, sexual orientation, society, academy, and objects. Classification. Before the implementation of the four digits of the fiscal reduction and so on, s' is shown in the following example. SO s02 s028 s028a (belonging to nature) (being natural to meteorology) (belonging to meteorological wind) (belonging to wind strength) For this hierarchical classification code, for example, as shown in Figure 13, the meaning range of the upper-level meaning code covers a wider range of meanings. Conversely, the meaning range of the lower-level meaning code is narrower. Therefore, we can cooperate The actual need to apply the meaning code 'does not have to-enumerate and save memory space. In addition, the n code 疋 is represented by a number ′. Therefore, mathematical operations such as 'set logical operation' can be used to process the meaning code, so that the meaning code month b generates more information with value. For a detailed description of the meaning code, refer to ',', the detailed description of the "machine translation device" of the Republic of China Patent Gazette No. 161238. The consumer cooperation of the Shiyang Standard Bureau of the Ministry of Economic Affairs and Du Duanyi, in addition, there are 1,300 pronunciations in Chinese, so as long as you use two digits mP, you can not represent all the pronunciations, of which there are η initials (consonants); η The 3 s members (vowels) have 14 and 5 tones. Therefore, the structure of the two-bit Chinese pronunciation is the first! As shown in the figure, bit W contains the initials (bit 2 ◦ spear; 丨 (bits 0 ～ 丨)), and bit 2 contains the tones (bits 4 to 6) and 4 mothers (bits το 0 to 3). Therefore, as long as column M covers the median region of byte 1 and uses the X set calculation method, g 卩 can find ^ having the same initials, finals, and original paper. The iron scales are applicable to China National Standards (CNS} -14- 420774 Jing Tao Ministry of Central Standards Bureau ua; Industrial and consumer cooperation Du printed A7 B * _______ V. Invention Description (11) ~ — " 一 ~~ Transcribed text》 In order to use the masking method to process similar sounds in each district, ten per district Elements in each group of similar sounds are assigned a bit pattern with a distance of 1, as shown in the example in Figure 2. For a detailed description of the Chinese pronunciation compression code and similar bit configuration, refer to the Republic of China Patent Gazette Bulletin No. 089477 "Chinese character conversion device (1) detailed instructions. In addition, in order to deal with the problem of redundant characters, missing characters, or missing characters caused by editing errors, the masking means of this embodiment can be used in addition to the above-mentioned bit masking. The entire text is obscured, for example, "六 一 厶 /," not On "liuyi 厶 / ， 为 — ,, · or it jig 厶 / \ (* 'stands for obscured text, that is, any text can be used) as the search key, you can check out" slim and slender "with reference to the phonetic dictionary. Ding In addition, referring to Yin Huamin Patent Gazette Bulletin No. 089476 "Chinese Character Conversion Device (2)", it was learned that the word length is an important price judgment when converting the phonetic symbol sequence to the literal sequence. Cai also uses word length weighting as an item of the evaluation function, and its calculation formula is as follows. For example, when the candidate word is "everyone", its word length weighting is (2-1) * 2 = 2. Word length weighting = (Word count of candidate word -1) * 2 Furthermore, in order to use the textual information of the original document order to find the best path effectively, this embodiment weights the similarity of the original book as an item of the evaluation function. 'The calculation formula is as follows. Original document similarity weighting = (guarantee space π accounted for two 広, ^ (candidate sub 6 orders and original document corresponding words of the same number of words) / candidate word number of papers Standards and wealth management materials (CNS) Α4 secret (17〇 ^ 297 ^ 7 —r 衣 —-- Please read the back Precautions to fill out this page) book

In I. -15- 420774 Λ7 、 --------- Β7_ V. Description of the invention (12) For example, the corresponding text in the original book is "Tingyuli" and the candidate word is "Tingyuyuli" , The original document similarity of this candidate term is weighted 3/4 (0.75). Furthermore, in this embodiment, consecutive semantic information is also introduced. As shown in FIG. 12, this semantic meaning is obtained by automatically learning the semantic meaning of the previous and subsequent words in the large corpus after tagging, or it can also be learned by labeling documents in different fields. And get. Since semantic codes are defined in a hierarchical manner, the calculation of semantic similarity between consecutive words can be calculated using set intersection. For example, the result of the set intersection of semantic codes [7140] and [714a] is [714]. The matching three-code semantic similarity is defined as 3/4. By analogy, it is 1 when they are exactly the same, 1/2 when they match two yards successfully, 1/4 when they match one yard, and 0 if there is no match at all. FIG. 3 is a system block diagram of an automatic document correction method and device in an embodiment of the present invention. 250 is a memory that uses a word sequence table as an index key and its corresponding pronunciation mark table 'and uses the above-mentioned pronunciation mark as an index key. The broken dictionary part of the corresponding candidate word and pronunciation mark, its schematic structure is shown in Figure 9; 260 is a word dictionary that stores the initial setting pronunciation and other possible pronunciation marks with the text mark as the index key The schematic structure is shown in FIG. 10; 450 is a vocabulary dictionary section that memorizes pronunciation marks as index keys and their corresponding homonyms and related frequency-weighted (long-term learning) and semantic codes. The structure diagram is shown in FIG. 11; 550 is a semantic dictionary for learning a semantic pair of word meaning codes as index keys and corresponding word meaning codes obtained after learning from a large tagging corpus. The schematic structure diagram is shown in FIG. As shown in Figure 12: 350 is temporarily memorized with paper in the middle. The paper ruler is applicable to the Chinese National Standard (CNS) Λ4 specification (2 丨 0X297 mm) (Please read the notes on the back first (Please fill this page again) Pack. Printed by the Employees' Cooperatives of the Central Bureau of Standards, Ministry of Economic Affairs -16-

A 420774 ______B7 V. Buffer of the description of the invention (100), 100 is the input part for inputting original document strings by input devices, such as magnetic butterfly machines, keyboards, etc .; 200 is a reference to the dictionary of broken sounds 250 and the dictionary of phonetic sounds 260: The word-to-phonetic conversion unit that converts the inputted original document string into a pronunciation symbol string; 300 is a reference to the phonetic word dictionary portion 45. After performing candidate syllable cuts on the pronunciation symbol string, all candidate words and related Candidate word selection section for information extraction; 400 is a similar candidate for similar single words and related information, which is used for contiguous single syllable candidates, using similar masking methods or vocal masking, and referring to the vocabulary dictionary section 4 5 Word selection unit; 500 is based on the start and end positions of each candidate word on the original document string as an index, and establishes a directed network of candidate words, and then refers to the semantic dictionary for learning dictionary 550 to use frequency weighting , Word length weighting, original document similarity weighting, and semantic similarity weighting sum are used as the evaluation function, and the best candidate word for the best path is found using dynamic programming from back to front Decision section: 6Q〇 is to take out the string on the best path, use the comparison method and the original document string stored in the buffer 350 to perform a comparison, and then mark the reference of the different text on the original document string Comparison section; 700 is a round-out section that rotates the best string and the original document string after marking. The processing flow chart of the word-to-sound conversion unit 200 is shown in FIG. 4. In step s201, the original document string is input by the input unit 丨 00 and stored in the buffer 350, and then proceeds to step 522 with reference to the word-sound dictionary unit 26. 〇 After performing syllable cutting on the original document string, proceed to step s203 with reference to the phonetic dictionary 26. Execute the pronunciation conversion of non-broken character syllables', and then proceed to step 32. 4 After referring to the broken character code 250, generate the pronunciation of the broken character syllables. Proceed to step s205 to use the grammar to correct the pronunciation marks. For example, the converted pronunciation of "Mom" is "The standard of Chinese paper (CNS) Α4 standard (2 丨 0x297 mm) for the paper scale. (Please read the notes on the back before filling This page} iL .. Printed by the Consumer Cooperatives of the Central Press Bureau of the Ministry of Economic Affairs -17- 420774 V. Description of Invention (μ) Yaya — '' But the second syllable at this time should be pronounced as a soft D Therefore, the pronunciation shown above after the execution of this step will be corrected to "门 丫 _ 门 丫." After the execution of step s205, the processing is terminated. The processing flowchart of the candidate word selection unit 300 is shown in FIG. s 301 The sequence of pronunciation marks completed by the word-to-speech conversion unit in the 200 round conversion. After the reference syllable dictionary unit 450 executes possible syllable cut-outs, it proceeds to step s302 and uses the pronunciation of each cut-out syllable as the search element reference. The phonetic word dictionary 45 takes out all the candidate words or words and their frequency weighting and semantic codes, and then proceeds to step s303 to store the candidate words or words and their related information in the buffer 350, and then ends the process. Printed by the employee's consumer cooperative (please read the notes on the back before filling in this page) The process flow similar to the candidate word selection section 400 is shown in Figure 6. Step s401 is input by the candidate word selection section 300. After the pronunciation symbol string and all candidate sub-words and their related information, proceed to step s402 to determine whether there are unprocessed consecutive transcripts for the syllable candidates, and if so, proceed to step S403 to use the pronunciation and Similar sound masking means, after referring to the phonetic word dictionary section 4 50 to retrieve similar candidate words and related information, execute step s 4 〇04 for the extracted similar candidate words, refer to the pronunciation marks After calculating the start and end positions of the similar candidate temples in series, the above-mentioned step s402 is executed; if the judgment result of step s402 is not, the process is terminated. The flow chart of the process of determining the best candidate string 500 is as the first step. As shown in Figure 7, after step s501, the similar candidate word selection unit 400 turns in similar candidate words and their related information, and then executes step S502 to retrieve the candidate words and their related information from the buffer 350, and then proceeds to step S503 uses the start and end positions of each candidate word as the index element. After establishing a directed network of candidate words, the steps are performed. The scale of this paper applies the Chinese Standard (CNS) Standard 8 (210 parent 297). — '• 18- Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs 420774 ~ ^ _______-------- B7__ V. Description of the invention (15) ~~-— One s504 is taken out from the buffer 350 In the original text, you start and end the book, using the starting and ending positions of each candidate word as the index element. After calculating the similarity weighting and word length weighting of the original document, perform step s505 to use frequency weighting, word length weighting, Original instrument similarity plus And semantic similarity degree evaluation function is a weighted sum, dynamic programming in a forward manner to identify the best path #, then, after step s5〇6 candidate of the optimal path and outputting the extracted terms &, the process ends. The processing flowchart of the reference comparison unit 600 is shown in FIG. 8. After step s60i is inputted by the best candidate string determination unit 500, the string eight on the best path is executed, and step S602 is executed to retrieve the original from the buffer 35o. After the document string 6, execute step S603 to compare the string eight and ugly by means of comparison, and mark the wrong word or word in the original document, and then execute step s604 to output the labeled original document string and the best path word After being connected to the output unit 700, the processing is terminated. Taking the automatic correction of Chinese documents as an example, the input unit 100 enters the original document string multilingual database system, and then enters the character-to-speech conversion unit 200 ′ with reference to the phonetic dictionary 260 and the broken dictionary 250 to convert the original characters The string conversion text "Xi 乂 B—UV 卩 I is a 乂 \ 万 乂, Ding Yi, Liu 乂, and the original document string is stored in the buffer 35. Then, after entering the candidate word selection section to perform possible syllable cutting, Taking these possible syllables as the search key, referring to the phonetic dictionary 450, all possible word or word candidates and related information are picked up, as shown in (3) of FIG. I4; then, a similar candidate word selection section 400 is entered, because Under the syllable "Xi 乂 r-uv", there are only candidate words but no candidate words. Therefore, for the two syllables, the above-mentioned syllable compression and masking methods are used. With reference to the syllable dictionary 450, it can be found as shown in (4) of Figure 14 After entering similar candidate words and related information, 'enter the best candidate string selection section 500, and first use the Chinese National Standard (CNS) A4 specification (210X297 mm) for each paper size ---.---- --¾ .------ ίτ ------ Λ t 锖 Read first Please note this item and fill in this page again) 420774 A7 Seal of the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs. 5. Description of Invention (l6) Candidate words or words are index elements at the beginning and end of the original document string. Candidate words or words are linked into a directed network. As shown in (5) of Figure 15, the learning dictionary is referred to by semantics, using frequency weighting, word length weighting, original document similarity weighting, and semantic similarity weighting. The sum is an evaluation function. In the dynamic programming method from back to front, the best path as shown in (5) of FIG. 15 can be found, and then the reference and comparison unit 600 is entered, and the original document is taken out from the buffer area 35. The strings are compared using the matching method to compare the strings on the best path with the original document strings and mark the differences with the mark symbol, as shown in the figure "Figure ⑹; drop words, []: typo. Finally, The wheel-out part side outputs the above-mentioned optimal path string and the labeled string. The present invention is not limited to this embodiment, and can be implemented as long as the gist is not changed. For example, the pronunciation symbol of the dictionary t can directly use two bytes. Compressed notes ^ Come Or the combination of the phonetic dictionary and the broken phonetic dictionary should be a modification of the financial embodiment. [Effects of the Invention] As described above, the present invention solves the problems of the conventional examples. In other words, the effects are: (1) It can be effective Error detection and correction of Chinese documents. The large tens of thousands of characters were selected from the textbook corpus of the elementary school as the experimental fund, and the error of the document was recorded by the person I arbitrarily and the location was recorded. After the document automatic correction device of the invention was implemented, The rate of anger and correction is about 87%. # J 阿 (2) = Any language model and special knowledge base are needed, so there is no need to spend manpower and material resources on collecting and maintaining the knowledge base. Its material is large (please read first Note on the back, please fill in this page again.) The size of the paper used for packing is in the Chinese country; quasi (CNs) A4 size (2! {) Κ297 公 爱 -20-

A U774 ^ _______ _B7 V. Description of the invention (17) —- (3) It can be used as the post-processing of Chinese input method or text recognition device, for example, it can effectively solve the problem of word grabbing. Wait, the present invention is highly practical. [Brief description of the figure] Figure 1 shows a 2-bit Chinese pronunciation structure diagram; Figure 2 shows an example of a bit pattern with a distance of 1 assigned to elements within each group of similar sounds in each zone; Figure 3 FIG. 4 is a block diagram of an embodiment of the present invention; FIG. 4 is a processing flowchart of a word-to-sound conversion unit in the embodiment of the present invention; FIG. 5 is a candidate word selection unit in the embodiment of the present invention FIG. 6 is a processing flowchart of a similar candidate word selection unit in the embodiment of the present invention; FIG. 7 is a processing flowchart of the best candidate string determination unit in the embodiment of the present invention; FIG. 8 is a processing flowchart of a reference comparison unit in the embodiment of the present invention; FIG. 9 is a schematic structural diagram of a broken dictionary in a processing flow in the embodiment of the present invention; FIG. 10 is an embodiment of the present invention Figure 11 shows the schematic structure of the phonetic dictionary. Figure 11 shows the structure of the phonetic dictionary in the embodiment of the present invention. The paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X297 public waste). (Please read the precautions on the back first. (Fill in this page again) -t Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs-21-420774 A7 B: V. Description of the invention (18) Figure 12 is a schematic diagram of the semantic dictionary for learning dictionary in the embodiment of the present invention: Figure 13 Concept of semantic hierarchy classification in the embodiment of the present invention FIG. 9 FIG. 14 is a schematic diagram of a processing example in the embodiment of the present invention: FIG. 15 is a continuous explanatory diagram of FIG. 14; FIG. 16 is a diagram of FIG. 14 Continuous explanatory diagram; FIG. 17 is a system block diagram of a conventional example. Component label comparison (please read the notes on the back before filling in this page) Binding 100... Input section 200... Word-to-sound conversion section 250. Broken dictionary section 260... Phonetic word section 300 ... Candidate word Selection unit 350 ... buffer 400 .. Similar candidate word selection unit 450 ... phonetic dictionary Zou 500 ... best candidate string determination unit 550 ... semantic pair learning dictionary 600 .. reference comparison unit 700 ... output unit Printed in New Zealand by the Consumers' Cooperative of the Shiyang Standard Bureau of the Ministry of Economic Affairs

Claims (1)

ABCD 420774 6. Scope of patent application 1 'An automatic correction method for Chinese documents, which is a method for the computer to automatically correct electronic Chinese documents, including the following steps: Step 1 Constructing a stored word order table pronunciation and corresponding word order in computer memory The phonetic notation table of the table and the corresponding dictionary of the phonetic notation and the pronunciation dictionary of the phonetic notation; construct and store in the memory order a character dictionary containing the text notation and its initial setting pronunciation and other possible pronunciation notations; construct and store in the memory Phonetic dictionaries with pronunciation marks and their corresponding all homonyms and their associated frequency weighting and semantic codes; Step 2 convert the original document string input by the input device into a phonetic symbol sequence with reference to the above-mentioned broken pronunciation dictionary and phonetic dictionary. Step 3: Perform syllable cut-outs on the pronunciation symbol string obtained by using the word-to-tone conversion method, and use each syllable as the retrieval element to refer to the phonetic dictionary to check the possible candidate words and related information. The candidate word selection means; step 4 The syllable uses a masking method to mask the similarity of the phonetic symbol string as a search element, and refers to a similar candidate word selection method that picks out possible candidate words and related information with reference to the phonological exhaustion; step 5 uses each candidate word The word position corresponding to the beginning and end of the original document string is used as an index element to link each candidate word into a directional network and use calculation methods to calculate the class A4 appearance of each candidate postword (210X297 mm > (please first (Read the note $ on the back side and fill out this page). Installation ·-, 11 Printed by the Consumer Standards Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs 420774 AS BS C8 D8 Printed by the Consumer Standards Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs After weighting with the word length, using the sum of frequency weighting, word length weighting, and similarity weighting of the original document as an evaluation function, a dynamic programming method is used to find the best candidate string determination method for the path of the best evaluation score; step 6 The text sequence on the best path and the original document word _ use the comparison means' to detect and mark the different words as a reference comparison means. 2. —kind A device for automatically correcting a document using computer technology, comprising: storing a word sequence table pronunciation, a phonetic notation table corresponding to the word sequence table, a candidate dictionary corresponding to the syllabary number list, and a broken pronunciation dictionary thereof; Phonetic dictionaries with text marks and their initial pronunciation, and other possible pronunciation marks; Phonetic dictionaries that store phonetic marks and their corresponding all homonyms and their related frequency weighting and semantic stele; refer to the above-mentioned broken pronunciation dictionary and phonetic dictionary The word-to-speech conversion unit that converts the original document string rotated by the input device into a pronunciation-symbol string; performs syllable cut-outs on the pronunciation-symbol string obtained by using the word-to-speech conversion method, and uses each syllable as the search element reference The phonetic dictionary lists possible candidate words and the candidate word detection unit related to the information > as a retrieval element, using a series of pronunciation marks after concealing similar bits by using a masking method for consecutive single-word candidate syllables. This paper size applies to the Chinese national standard (CNS > A4 size (210 X 297 public director)) Note on the back page, please fill in this page again) Installation · -Order _ < k -24- Printed by Employee Consumer Cooperative of Central Bureau of Standards, Ministry of Economic Affairs 420774 ίΐ C8 _______ D8 VI. Candidates with patent scope 3 and related information Detected similar candidate characters 々§1 selection unit; use the candidate text corresponding to the beginning and end of the original document string as the index element to link each candidate word into a directed network and use special means to calculate each After candidate word similarity weighting and word length weighting, the sum of frequency weighting, word length weighting, and original document similarity weighting is used as an evaluation function to find the best candidate string for the path of the best evaluation score in a dynamic programming manner. Decision department; a reference comparison department that uses the text string on the best path and the original document string to compare and find out the different words and marks them. 3. The automatic correction device for Chinese documents according to item 2 of the scope of the patent application, which is characterized by: constructing a combination of the post-word meaning code and the pre-word meaning code learned through learning means in the s-memory body. Semantic pair learning dictionary > The above-mentioned best candidate string determination unit uses the position of each candidate word corresponding to the start and end of the original document string as an index element to link each candidate word into a directed network and uses calculation means After calculating the similarity weighting and word length weighting of each candidate, the sum of frequency of weighting of learning subscripts using semantic meaning, weighting of word length, similarity weighting of original documents, and weighting of semantic similarity is used as the evaluation function. Those who get the best evaluation score. This paper size applies to Chinese National Standard (CNS) A4 specification (2l0x297) J1 [T-! I I In I—- I · ici I —I— ---- _ _--I ^ -Ψ. --- ^^ 1 _ I n 1 N U3 i Hu— (Please read the precautions on the back before filling this page) -25-