Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/10—Text processing
  • G06F40/12—Use of codes for handling textual entities
  • G06F40/126—Character encoding
  • G06F40/129—Handling non-Latin characters, e.g. kana-to-kanji conversion
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/20—Natural language analysis
  • G06F40/205—Parsing
  • G06F40/216—Parsing using statistical methods
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/20—Natural language analysis
  • G06F40/232—Orthographic correction, e.g. spell checking or vowelisation
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/40—Processing or translation of natural language
  • G06F40/42—Data-driven translation
  • G06F40/44—Statistical methods, e.g. probability models
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/40—Processing or translation of natural language
  • G06F40/53—Processing of non-Latin text

Definitions

• the invention relates to a language input method and system. More particularly, the invention provides language input method and system that has error tolerance for both typographical errors that occur during text entry and conversion errors that occur during conversion from one language form to another language form.
• Language specific word processing software has existed for many years. More sophisticated word processors offer users advanced tools, such as spelling and grammar correction, to assist in drafting documents. Many word processors, for example, can identify words that are misspelled or sentence structures that are grammatically incorrect and, in some cases, automatically correct the identified errors.
• Word processors can offer suggestions to aid the user in choosing a correct spelling or phraseology.
• the second and more typical cause of errors is that the user incorrectly enters the words or sentences into the computer, even though he/she knew the correct spelling or grammatical construction. In such situations, word processors are often quite useful at identifying the improperly entered character strings and correcting them to the intended word or phrase. Entry errors are often more prevalent in word processors designed for languages that do not employ Roman characters.
• Language specific keyboards such as the English version QWERTY keyboards, do not exist for many languages because such languages have many more characters than can be conveniently arranged as keys in the keyboard. For example, many Asian languages contain thousands of characters. It is practically impossible to build a keyboard to support separate keys for so many different characters.
• language specific word processing systems allow the user to enter phonetic text from a small character-set keyboard (e.g., a QWERTY keyboard) and convert that phonetic text to language text.
• "Phonetic text” represents the sounds made when speaking a given language
• the "language text” represents the actual written characters as they appear in the text.
• Pinyin is an example of phonetic text
• Hanzi is an example of the language text.
• Word processors that require phonetic entry thus experience two types of potential entry errors.
• One type of error is common typing mistakes.
• event if the text is free of typographical errors another type of error is that the word processing engine might incorrectly convert the phonetic text to an unintended character text.
• a cascade of multiple errors may result.
• the typing induced errors may not be readily traced without a lengthy investigation of the entire context of the phrase or sentence.
• the invention described herein is directed primarily to the former type of entry errors made by the user when typing in the phonetic text, but also provide tolerance for conversion errors made by the word processing engine. To better demonstrate the problems associated with such typing errors, consider a Chinese- based word processor that converts the phonetic text, Pinyin, to a language text, Hanzi.
• N third reason for increased typing errors during phonetic text input is that many people speak natively in a regional dialect, as opposed to a standard dialect.
• the standard dialect which is the origin of phonetic text, is a second language.
• spoken words may not match corresponding proper phonetic text, thus making it more difficult for a user to type phonetic text.
• many Chinese speak various Chinese dialects as their first language and are taught Mandarin Chinese, which is the origin of Pinyin, as a second language.
• Mandarin as a second language may be prone to typing errors when attempting to enter Pinyin.
• Another possible reason for increased typing errors is that it is difficult to check for errors while typing phonetic text. This is due in part to the fact that phonetic text tends to be long, unreadable strings of characters that are difficult to read.
• entered phonetic text is often not "what you see is what you get.” Rather, the word processor converts the phonetic text to language text. Ns a result, users generally do not examine the phonetic text for errors, but rather wait until the phonetic text is converted to the language text.
• Pinyin character strings are very difficult to review and correct because there is no spacing between characters. Instead, the Pinyin characters run together irregardless of the number of words being formed by the Pinyin characters.
• Pinyin-to-Hanzi conversion often does not occur immediately, but continues to formulate correct interpretations as additional Pinyin text is entered.
• the single error may be compounded by the conversion process and propagated downstream to cause several additional errors. Ns a result, error correction takes longer because by the time the system converts decisively to Hanzi characters and then the user realizes there has been an error, the user is forced to backspace several times just to make one correction. In some systems, the original error cannot even be revealed.
• Language specific word processors face another problem, separate from the entry problem, which concerns switching modes between two languages in order to input words from the different language into the same text. It is common, for example, to draft a document in Chinese that includes English words, such as technical terms (e.g., Internet) and terms that are difficult to translate (e.g., acronyms, symbols, surnames, company names, etc.).
• Conventional word processors require a user to switch modes from one language to the other language when entering the different words.
• the user when a user wants to enter a word from a different language, the user must stop thinking about text input, switch the mode from one language to another, enter the word, and then switch the mode back to the first language.
• Ns an example, when a user types a string of Pinyin input text "woshiyigezhongguoren", the system converts this string into Chinese character: " (generally translated to "I am a Chinese”).
• a user instead of typing "woshiyigezhongguoren", a user types the following:
• wosiyigezhongguoren (the error is the "sh” and “s” confusion); woshiyigezongguoren (the error is the “zh” and “z” confusion); woshiygezhongguoren (the error is the "i” omission after "y”); woshiyigezhonggouren (the error is the “ou” juxtaposition); woshiyigezhongguiren (the error is the "i” and "o” confusion).
• the inventors have developed a word processing system and method that makes spell correction feasible for difficult foreign languages, such as Chinese, and allows modeless entry of multiple languages through automatic language recognition.
• a language input architecture converts input strings of phonetic text (e.g., Chinese Pinyin) to an output string of language text (e.g., Chinese Hanzi) in a manner that minimizes typographical errors and conversion errors that occur during conversion from the phonetic text to the language text.
• the language input architecture may be implemented in a wide variety of areas, including word processing programs, email programs, spreadsheets, browsers, and the like.
• the language input architecture has a user interface to receive in input string of characters, symbols, or other text elements.
• the input string may include phonetic text and non-phonetic text, as well as one or more languages.
• the user interface allows the user to enter the input text string in a single edit line without switching modes between entry of different text forms or different languages. In this manner, the language input architecture offers modeless entry of multiple languages for user convenience.
• the language input architecture also has a search engine, one or more typing models, a language model, and one or more lexicons for different languages.
• the search engine receives the input string from the user interface and distributes the input string to the one or more typing models.
• Each typing model is configured to generate a list of probable typing candidates that may be substituted for the input string based on typing error probabilities of how likely each of the candidate strings was incorrectly entered as the input string.
• the probable typing candidates may be stored in a database.
• the typing model is trained from data collected from many trainers who enter a training text. For instance, in the context of the Chinese language, the trainers enter a training text written in Pinyin. The observed errors made during entry of the training text are used to compute the probabilities associated with the typing candidates that may be used to correct the typing error. Where multiple typing models are employed, each typing model may be trained in a different language.
• the typing model may be trained by reading strings of input text and mapping syllables to corresponding typed letters of each string. A frequency count expressing the number of times each typed letter is mapped to one of the syllables is kept and the probability of typing for each syllable is computed from the frequency count.
• the typing model returns a set of probable typing candidates that account for possible typographical errors that exist in the input string.
• the typing candidates are written in the same language or text form as the input string.
• the search engine passes the typing candidates to the language model, which provides probable conversion strings for each of the typing candidates.
• the language model is a trigram language model that attempts to determine a language text probability of how likely a probable conversion output string represents the candidate string based on two previous textual elements.
• the conversion string is written in a different language or different text form than the input string.
• the input string might comprise Chinese Pinyin or other phonetic text and the output string might comprise Chinese Hanzi or other language text.
• the search engine selects the associated typing candidate and conversion candidate that exhibits the highest probability.
• the search engine converts the input string (e.g., written in phonetic text) to an output string consisting of the conversion candidate returned from the language model so that the entered text form (e.g., phonetic text) is replaced with another text form (e.g., language text). In this manner, any entry error made by the user during entry of the phonetic text is eliminated.
• the output string may have a combination of the conversion candidate as well as portions of the input string
• the user interface displays the output string in the same edit line that continues to be used for entry of the input string. In this manner, the conversion is taking place automatically and concurrently with the user entering additional text.
• Fig. 1 is a block diagram of a computer system having a language-specific word processor that implements a language input architecture.
• Fig. 2 is a block diagram of one exemplary implementation of the language input architecture.
• Fig. 3 is a diagrammatic illustration of a text string that is parsed or segmented into different sets of syllables, and candidates that may be used to replace those syllables assuming the text string contains errors.
• Fig. 4 is a flow diagram illustrating a general conversion operation performed by the language input architecture.
• Fig. 5 is a block diagram of a training computer used to train probability- based models employed in the language input architecture.
• Fig. 6 is a flow diagram illustrating one training technique.
• Fig. 7 is a block diagram of another exemplary implementation of the language input architecture, in which multiple typing models are employed.
• Fig. 8 is a flow diagram illustrating a multilingual conversion process.
• the invention pertains to a language input system and method that converts one form of a language (e.g., phonetic version) to another form of the language (e.g., written version).
• the system and method have error tolerance for spelling and typographical errors that occur during text entry and conversion errors that occur during conversion from one language form to another language form.
• the invention is described in the general context of word processing programs executed by a general-purpose computer. However, the invention may be implemented in many different environments other than word processing and may be practiced on many diverse types of devices. Other contexts might include email programs, spreadsheets, browsers, and the like.
• the language input system employs a statistical language model to achieve very high accuracy.
• the language input architecture uses statistical language modeling with automatic, maximum- likelihood-based methods to segment words, select a lexicon, filter training data, and derive a best possible conversion candidate.
• the language input architecture includes one or more typing models that utilize probabilistic spelling models to accept correct typing while tolerating common typing and spelling errors.
• the typing models may be trained for multiple languages, such as English and Chinese, to discern how likely the input sequence is a word in one language as opposed to another language. Both models can run in parallel and are guided by the language model (e.g., a Chinese language model) to output the most likely sequence of characters (i.e., English and Chinese characters).
• Exemplary Computer System Fig. 1 shows an exemplary computer system 100 having a central processing unit (CPU) 102, a memory 104, and an input/output (I/O) interface 106.
• the CPU 102 communicates with the memory 104 and I/O interface 106.
• the memory 104 is representative of both volatile memory (e.g., RAM) and non-volatile memory (e.g., ROM, hard disk, etc.).
• the computer system 100 has one or more peripheral devices connected via the I/O interface 106.
• Exemplary peripheral devices include a mouse 110, a keyboard 112 (e.g., an alphanumeric QWERTY keyboard, a phonetic keyboard, etc.), a display monitor 114, a printer 116, a peripheral storage device 118, and a microphone 120.
• the computer system may be implemented, for example, as a general-purpose computer. Nccordingly, the computer system 100 implements a computer operating system (not shown) that is stored in memory 104 and executed on the CPU 102.
• the operating system is preferably a multi-tasking operating system that supports a windowing environment.
• An example of a suitable operating system is a Windows brand operating system from Microsoft Corporation.
• Fig. 1 the language input system may be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network (e.g., LAN, Internet, etc.).
• program modules may be located in both local and remote memory storage devices.
• a data or word processing program 130 is stored in memory 104 and executed on CPU 102. Other programs, data, files, and such may also be stored in memory 104, but are not shown for ease of discussion.
• the word processing program 130 is configured to receive phonetic text and convert it automatically to language text. More particularly, the word processing program 130 implements a language input architecture 131 that, for discussion purposes, is implemented as computer software stored in memory and executable on a processor.
• the word processing program 130 may include other components in addition to the architecture 131, but such components are considered standard to word processing programs and will not be shown or described in detail.
• the language input architecture 131 of word processing program 130 has a user interface (UI) 132, a search engine 134, one or more typing models 135, a language model 136, and one or more lexicons 137 for various languages.
• the architecture 131 is language independent.
• the UI 132 and search engine 134 are generic and can be used for any language.
• the architecture 131 is adapted to a particular language by changing the language model 136, the typing model 135 and the lexicon 137.
• the search engine 134 and language module 136 together form a phonetic text-to-language text converter 138. With the assistance of typing model 135, the converter 138 becomes tolerant to user typing and spelling errors.
• text means one or more characters and/or non-character symbols.
• Phonetic text generally refers to an alphanumeric text representing sounds made when speaking a given language.
• a "language text” is the characters and non- character symbols representative of a written language.
• Non-phonetic text is alphanumeric text that does not represent sounds made when speaking a given language. Non-phonetic text might include punctuation, special symbols, and alphanumeric text representative of a written language other than the language text.
• phonetic text may be any alphanumeric text represented in a Roman-based character set (e.g., English alphabet) that represents sounds made when speaking a given language that, when written, does not employ the Roman-based character set.
• Language text is the written symbols corresponding to the given language.
• word processor 130 is described in the context of a Chinese-based word processor and the language input architecture 131 is configured to convert Pinyin to Hanzi. That is, the phonetic text is Pinyin and the language text is Hanzi.
• the language input architecture is language independent and may be used for other languages.
• the phonetic text may be a form of spoken Japanese, whereas the language text is representative of a Japanese written language, such as Kanji.
• Kanji a Japanese written language
• Phonetic text is entered via one or more of the peripheral input devices, such as the mouse 110, keyboard 112, or microphone 120.
• the computer system may further implement a speech recognition module (not shown) to receive the spoken words and convert them to phonetic text.
• a speech recognition module not shown
• the UI 132 displays the phonetic text as it is being entered.
• the UI is preferably a graphical user interface. N more detailed discussion of the UI 132 is found in co-pending application Serial No. , entitled "LANGUAGE
• the user interface 132 passes the phonetic text (P) to the search engine 134, which in turn passes the phonetic text to the typing model 137.
• the typing model 137 generates various typing candidates (TCj, ..., TC N ) that might be suitable edits of the phonetic text intended by the user, given that the phonetic text may include errors.
• the typing model 137 returns multiple typing candidates with reasonable probabilities to the search engine 134, which passes the typing candidates onto the language model 136.
• the language model 136 evaluates the typing candidates within the context of the ongoing sentence and generates various conversion candidates (CQ, ..., CC N ) written in the language text that might be representative of a converted form of the phonetic text intended by the user.
• the conversion candidates are associated with the typing candidates.
• Conversion from phonetic text to language text is not a one-for-one conversion.
• the same or similar phonetic text might represent a number of characters or symbols in the language text.
• the context of the phonetic text is interpreted before conversion to language text.
• conversion of non-phonetic text will typically be a direct one-to-one conversion wherein the alphanumeric text displayed is the same as the alphanumeric input.
• the conversion candidates (CQ, ..., CC N ) are passed back to the search engine 134, which performs statistical analysis to determine which of the typing and conversion candidates exhibit the highest probability of being intended by the user.
• the search engine 134 selects the candidate with the highest probability and returns the language text of the conversion candidate to the UI 132.
• the UI 132 then replaces the phonetic text with the language text of the conversion candidate in the same line of the display. Meanwhile, newly entered phonetic text continues to be displayed in the line ahead of the newly inserted language text.
• the user interface 132 presents a first list of other high probability candidates ranked in order of the likelihood that the choice is actually the intended answer. If the user is still dissatisfied with the possible candidates, the UI 132 presents a second list that offers all possible choices.
• the second list may be ranked in terms of probability or other metric (e.g., stroke count or complexity in Chinese characters).
• Fig. 2 illustrates the language input architecture 131 in more detail.
• the architecture 131 supports error tolerance for language input, including both typographical errors and conversion errors.
• search engine In addition to the UI 132, search engine
• the architecture 131 further includes an editor 204 and a sentence context model 216.
• a sentence context model 216 is coupled to the search engine 134.
• the user interface 132 receives input text, such as phonetic text (e.g. Chinese Pinyin text) and non-phonetic text (e.g., English), from one or more peripheral devices (e.g., keyboard, mouse, microphone) and passes the input text to the editor 204.
• the editor 204 requests that the search engine 132, in conjunction with the typing model 135 and language model 136, convert the input text into an output text, such as a language text (e.g. Chinese Hanzi text).
• the editor 204 passes the output text back to the UI 132 for display.
• the search engine 134 Upon receiving a string of input text from the user interface 132, the search engine 134 sends the string of input text to one or more of the typing models 135 and to the sentence context model 216.
• the typing model 135 measures a priori probability of typing errors in the input text.
• the typing model 135 generates and outputs probable typing candidates for the input text entered by the user, effectively seeking to cure entry errors (e.g., typographical errors).
• the typing model 135 looks up potential candidates in a candidate database 210.
• the typing model 135 uses statistical-based modeling to generate probable candidates for the input text.
• the sentence context model 216 may optionally send any previously input text in the sentence to the search engine 132 to be used by the typing model 135. In this manner, the typing model may generate probable typing candidates based on a combination of the new string of text and the string of text previously input in the sentence.
• typing errors may be interchangeable to refer to the errors made during keyed entry of the input text. In the case of verbal entry, such errors may result from improper recognition of the vocal input.
• the typing model 135 may return all of the probable typing candidates or prune off the probable typing candidates with lower probability, thereby returning only the probable typing candidates with higher probability back to the search engine 134. It will also be appreciated that the search engine 134, rather than the typing model 135, can perform the pruning function.
• the typing model 135 is trained using real data 212 collected from hundreds or thousands of trainers that are asked to type in sentences in order to observe common typographical mistakes.
• the typing model and training are described below in more detail under the heading
• the search engine 134 sends the list of probable typing candidates returned from the typing model 135 to the language model 136.
• a language model measures the likelihood of words or text strings within a given context, such as a phrase or sentence. That is, a language model can take any sequence of items (words, characters, letters, etc.) and estimate the probability of the sequence.
• the language model 136 combines the probable typing candidates from the search engine 134 with the previous text and generates one or more candidates of language text corresponding to the typing candidates.
• Corpus data or other types of data 214 are used to train the trigram language model 136.
• the training corpus 214 may be any type of general data, such as everyday text such as news articles or the like, or environment-specific data, such as text directed to a specific field (e.g., medicine).
• Training the language model 136 is known in the word processing art and is not described in detail.
• the language input architecture 131 tolerates errors made during entry of an input text string and attempts to return the most likely words and sentences given the input string.
• the language model 136 helps the typing model 135 to determine which sentence is most reasonable for the input string entered by the user.
• the two models can be described statistically as the probability that an entered string s is a recognizable and valid word w from a dictionary, or P(w
• the denominator P(s) remains the same for purposes of comparing possible intended words given the entered string. Nccordingly, the analysis concerns only the numerator product P(s
• ⁇ ) can be restated as P(H
• H represents a Hanzi string
• P represents a Pinyin string.
• the goal is to find the most probable Chinese character H', so as to maximize P(H
• P) is the likelihood that an entered Pinyin string P is a valid Hanzi string H. Since P is fixed and hence P(P) is a constant for a given Pinyin string, Bayes formula reduces the probability P(H
• P), as follows: H' arg maxH P(H
• P) arg maxH P(P
• H) represents the spelling or typing model.
• the Hanzi string H can be further decomposed into multiple words Wi, W 2 , W 3 , ..., W M, and the probability P(P
• P f( , ) is the sequence of Pinyin characters that correspond to the word W,.
• W, is set to 1 if P f(l) is an acceptable spelling of word W, and is set to 0 if P f(l) is not an acceptable spelling of word W,.
• Ns a result conventional systems provide no tolerance for any erroneously entered characters.
• Some systems have the "southern confused pronunciation" feature to deal with this problem, alghough this also employs the preset values probabilities of 1 and 0.
• such systems only address a small fraction of typing errors because it is not data- driven (learned from real typing errors).
• the language architecture described herein utilizes both the typing model and the language model to carry out a conversion.
• the typing model enables error tolerance to erroneously input characters by training the probability of P(P f( I W,) from a real corpus.
• W,) can be trained; but in practice, there are too many parameters.
• one approach is to consider only single-character words and map all characters with equivalent pronunciation into a single syllable.
• the probability P(H) represents the language model, which measures the a priori probability of any given string of words.
• a common approach to building a statistical language model is to utilize a prefix tree-like data structure to build an N- gram language model from a known training set of text.
• One example of a widely used statistical language model is the N-gram Markov model, which is described in "Statistical Methods for Speech Recognition", by Frederick Jelinek, The MIT Press, Cambridge, Massachusetts, 1997.
• the use of a prefix tree data structure a.k.a. a suffix tree, or a PAT tree
• the N-gram language model counts the number of occurrences of a particular item (word, character, etc.) in a string (of size N) throughout a text. The counts are used to calculate the probability of the use of the item strings.
• a trigram model considers the two most previous characters in a text string to predict the next character, as follows:
• characters (C) are segmented into discrete language text or words (W) using a pre-defined lexicon, wherein each W is mapped in the tree to one or more C's;
• P( ) represents the probability of the language text
• W n _ 2 is the word previous to W n _ !
• Fig. 3 illustrates an example of input text 300 that is input by a user and passed to the typing model 135 and the language model 136.
• the typing model 135 segments the input text 300 in different ways to generate a list of probable typing candidates 302 that take into account possible typographical errors made during keyboard entry.
• the typing candidates 302 have different segmentations in each time frame such that the end-time of a previous word is a start-time of a current word. For instance, the top row of candidates 302 segments the input string 300 "mafangnitryyis -- as “ma”, “fan”, “ni”, “try”, “yi”, and so on.
• the second row of typing candidate 302 segments the input string “mafangnitryyis" differently as “ma”, “fang”, “nit”, “yu”, “xia”, and so on.
• the candidates may be stored in a database, or some other accessible memory. It will be appreciated that Fig. 3 is merely one example, and that there might be a different number of probable typing candidates for the input text.
• the language model 136 evaluates each segment of probable typing candidates 302 in the context of the sentence and generates associated language text. For illustration purposes, each segment of the probable typing text 302 and the corresponding probable language text are grouped in boxes. From the candidates, the search engine 134 performs statistical analysis to determine which of the candidates exhibit the highest probability of being intended by the user. The typing candidates in each row have no relation to one another, so the search engine is free to select various segments from any row to define acceptable conversion candidates. In the example of Fig. 3, the search engine has determined that the highlighted typing candidates 304, 306, 308, 310, 312, and 314 exhibit the highest probability. These candidates may be concatenated from left to right so that candidate 304 is followed by candidate 306, and so on, to form an acceptable interpretation of the input text 300.
• the search engine 134 selects the candidate with the highest probability.
• the search engine then converts the input phonetic text to the language text associated with the selected candidate. For instance, the search engine converts the input text 300 to the language text illustrated in boxes 304, 306, 308, 310, 312, and 314 and returns the language text to the user interface 132 via the editor 204.
• the typing model 135 begins operating on the new string of text in the new sentence.
• Fig. 4 illustrates a general process 400 of converting phonetic text (e.g.,
• the user interface 132 receives a phonetic text string, such as
• the input text string contains one or more typographical errors.
• the UI 132 passes the input text via the editor 204 to the search engine 134, which distributes the input text to the typing model 135 and the sentence context model 216.
• the typing model 135 generates probable typing candidates based on the input text.
• One way to derive the candidates is to segment the input text string in different partitions and look up candidates in a database that most closely resemble the input string segment. For instance, in Fig. 3, candidate 302 has a segmentation that dictates possible segments "ma”, "fan”, and so forth.
• the probable typing candidates are returned to the search engine 134, which in turn conveys them to the language model 136.
• the language model 136 combines the probable typing candidates with the previous text and generates one or more candidates of language text corresponding to the typing candidates. With reference to candidate 302 in Fig. 3, for example, the language model returns the language text in boxes 302a-j as possible output text.
• the search engine 134 performs statistical analysis to determine which of the candidates exhibit the highest probability of being intended by the user. Upon selecting the most probable typing candidate for the phonetic text, the search engine converts the input phonetic text to the language text associated with the typing candidate. In this manner, any entry error made by the user during entry of the phonetic text is eliminated. The search engine 134 returns the error- free language text to the UI 132 via the editor 204. At step 408, the converted language text is displayed at the UI 132 in the same in-line position on the screen that the user is continuing to enter phonetic text.
• the typing model 135 is based on the probability P(s
• the typing model computes probabilities for different typing candidates that can be used to convert the input text to the output text and selects probable candidates. In this manner, the typing model tolerates errors by returning the probable typing candidates for the input text even though typing errors are present.
• One aspect of this invention concerns training the typing model P(s
• the typing model is developed or trained on text input by as many trainers as possible, such as hundreds or preferably thousands.
• the trainers enter the same or different training data and any variance between the entered and training data is captured as typing errors.
• the goal is to get them to type the same training text and determine the probabilities based on the numbers of errors or typing candidates in their typing. In this way, the typing model learns probabilities of trainers' typing errors.
• Fig. 5 shows a training computer 500 having a processor 502, a volatile memory 504, and a non- volatile memory 506.
• the training computer 500 runs a training program 508 to produce probabilities 512 (i.e., P(s
• Training computer 500 may be configured to train on data 510 as it is entered on the fly, or after it is collected and stored in memory.
• Training computer 500 may be configured to train on data 510 as it is entered on the fly, or after it is collected and stored in memory.
• a typing model tailored for the Chinese language wherein Chinese Pinyin text is converted to Chinese character text.
• several thousands of people are invited to input Pinyin text.
• several hundred sentences or more are collected from each person, with the goal of getting them to make similar types and numbers of errors in their typing.
• the typing model is configured to receive Pinyin text from the search engine, and provide probable candidates that may be used to replace characters in the input string.
• the typing model is trained directly by considering a single character text and mapping all equivalently pronounced character text to a single syllable. For example, there are over four hundred syllables in Chinese Pinyin.
• the probability of phonetic text given a syllable e.g.. P(Pinyin text
• each character text is mapped to its corresponding syllable.
• Fig. 6 shows the syllable mapping training technique 600.
• the training program 508 reads a string of text entered by trainer.
• the text string may be a sentence or some other grouping of words and/or characters.
• the program 508 aligns or maps syllables to corresponding letters in the string of text (step 604). For each text string, the frequency of letters mapped to each syllable is updated (step 606). This is repeated for each text string contained in the training data entered by the trainers, as represented by the "Yes” branch from step 608. Eventually, the entered text strings will represent many or all syllables in Chinese Pinyin. Once all strings are read, as represented by the "No" branch from step 608, the training program determines the probability P(Pinyin text
• Each syllable can be represented as a hidden Markov model (HMM).
• HMM hidden Markov model
• Each input key can be viewed as a sequence of states mapped in HMM. The correct input and actual input are aligned to determine a transition probability between states.
• Different HMMs can be used to model typists with different skill levels.
• To train all 406 syllables in Chinese a large amount of data is needed. To reduce this data requirement, the same letter in different syllables is tied as one state. This reduces the number of states to 27 (i.e., 26 different letters from 'a' to 'z', plus one to represent an unknown letter).
• This model could be integrated into a Viterbi beam search that utilizes a trigram language model.
• training is based on the probability of single letter edits, such as insertion of a letter (i.e., ⁇ rf ⁇ x), deletion of a letter (i.e., xri ⁇ ⁇ ), and substitution of one letter for another (xci ⁇ y).
• the probability of such single letter edits can be represented statistically as:
• Each probability (P) is essentially a bigram typing model, but could also be extended to a N-gram typing model that considers a much broader context of text beyond adjacent characters. Accordingly, for any possible string of input text, the typing model has a probability of generating every possible letter sequence - by first providing the correct letter sequence, and then using dynamic programming to determine a lowest-cost path to convert the correct letter sequence to the given letter sequence. Cost may be determined as the minimal number of error characters , or some other measure. In practice, this error model can be implemented as a part of the Viterbi Beam searching method.
• Another annoying problem that plagues language input systems is the requirement to switch among modes when entering two or more languages. For instance, a user who is typing in Chinese may wish to enter an English word.
• the language input architecture 131 (Fig. 1) can be trained to accept mixed- language input, and hence eliminate mode shifting between two or more languages in a multilingual word processing system. This is referred to as "modeless entry”.
• the language input architecture implements a spelling/typing model that automatically distinguishes between words of different languages, such as discerning which word is Chinese and which word is English. This is not easy because many legal English words are also legal Pinyin strings. Additionally, since there are no spaces between Pinyin, English and Chinese characters, more ambiguities can arise during entry. Using Bayes rule:
• H' arg max H P(H
• P) arg max H P(P
• One way to handle mixed-language input is to train thelanguage model for a first language (e.g., Chinese) by treating words from a second language (e.g., English) as a special category of the first language. For instance, the words from the second language are treated as single words in the first language.
• a first language e.g., Chinese
• a second language e.g., English
• the typing model employed in the Chinese-based word processing system is a Chinese language model that is trained on text having a mixture of English words and Chinese words.
• a second way to handle mixed-language input is to implement two typing models in the language input architecture, a Chinese typing model and an English typing model, and train each one separately. That is, the Chinese typing model is trained a stream of keyboard input, such as phonetic strings, entered by trainers in the manner described above, and the English typing model is trained on English text entered by English-speaking trainers.
• the English typing model may be implemented as a combination of:
• An English spelling model of tri-syllable probabilities This model should has non-zero probabilities for every 3 -syllable sequence, but also generates a higher probability for words that are likely to be English-like. This can be trained from real English words also, and can handle unseen
• Fig. 7 illustrates a language input architecture 700 that is modified from the architecture 131 in Fig. 2 to employ multiple typing models 135(1)-135(N). Each typing model is configured for a specific language. Each typing model 135 is trained separately using words and errors common to the specific language. Accordingly, separate training data 212(1)-212(N) is supplied for associated typing models 135(1)-135(N). In the exemplary case, only two typing models are used: one for English and one for Chinese. However, it should be appreciated that the language input architecture may be modified to include more than two typing models to accommodate entry of more than two languages. It should also be noted that the language input architecture may be used in many other types of multilingual word processing systems, such as Japanese, Korean, French, German, and the like.
• the English typing model operates in parallel with the Chinese typing model.
• the two typing models compete with one another to discern whether the input text is English or Chinese by computing probabilities that the entered text string is likely to be a Chinese string (including errors) or an English string (also potentially including errors).
• the Chinese typing model When a string or sequence of input text is clearly Chinese Pinyin text, the Chinese typing model returns a much higher probability than the English typing model.
• the language input architecture converts the input Pinyin text to the Hanzi text.
• a string or sequence of input text is clearly English (e.g., a surname, acronym (“IEEE”), company name (“Microsoft”), technology (“INTERNET”), etc.)
• the English typing model exhibits a much higher probability than the Chinese typing model.
• the architecture converts the input text to English text based on the English typing model.
• the Chinese and English typing models continue to compute probabilities until further context lends more information to disambiguate between Chinese and English.
• a string or sequence of input text is not like either Chinese or English, the Chinese typing model is less tolerant than the English typing model. Ns a result, the English typing model has a higher probability than the Chinese typing model.
• the Chinese typing model Upon receiving the initial string "woaidu", the Chinese typing model yields a higher probability than the English typing model and converts that portion of the input text to "INTERNET ".
• the architecture continues to find the subsequently typed portion “interne” ambiguous until letter "t” is typed.
• the English typing model returns a higher probability for "INTERNET” than the Chinese typing model and the language input architecture converts this portion of the input text to
• Fig. 8 illustrates a process 800 of converting a multilingual input text string entered with typographical errors into a multilingual output text string that is free of errors.
• the process is implemented by the language input architecture 700, and is described with additional reference to Fig. 7.
• the user interface 132 receives the multilingual input text string. It contains phonetic words (e.g., Pinyin) and words of at least one other language (e.g., English).
• the input text may also include typographical errors made by the user when entering the phonetic words and second language words.
• the UI 132 passes the multilingual input text string via the editor 204 to the search engine 134, which distributes the input text to the typing models 135(1)-135(N) and the sentence context model 216.
• Each of the typing models generates probable typing candidates based on the input text, as represented by steps 804(1)-804(N).
• the probable typing candidates that possess reasonable probabilities are returned to the search engine 134.
• the search engine 134 sends the typing candidates with typing probabilities to the language model 136.
• the language model combines the probable typing candidates with the previous text to provide sentence-based context and generates one or more conversion candidates of language text corresponding to the typing candidates by selecting a path through the typing candidates, as described above with respect to Fig. 3.
• the search engine 134 performs statistical analysis to select the conversion candidates that exhibit the highest probability of being intended by the user.
• the most probable conversion candidate for the text string is converted into the output text string.
• the output text string includes language text (e.g., Hanzi) and the second language (e.g., English), but omits the typing errors.
• the search engine 134 returns the error-free output text to the UI 132 via the editor 204.
• the converted language text is displayed at the UI 132 in the same in-line position on the screen that the user is continuing to enter phonetic text.
• Chinese language is the primary language
• English is the secondary language. It will be appreciated that the two languages can both be designated primary languages.
• more than two languages may form the mixed input text string.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• Computational Linguistics (AREA)
• Artificial Intelligence (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Health & Medical Sciences (AREA)
• General Health & Medical Sciences (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Probability & Statistics with Applications (AREA)
• Document Processing Apparatus (AREA)
• Machine Translation (AREA)

Priority Applications (3)

Applications Claiming Priority (4)

Publications (2)

Family

ID=26860054

Family Applications (1)

Country Status (5)

Cited By (4)

Families Citing this family (219)

Family Cites Families (83)

• 2000
  • 2000-06-28 US US09/606,660 patent/US6848080B1/en not_active Expired - Fee Related
  • 2000-10-13 CN CNB008152934A patent/CN1205572C/zh not_active Expired - Fee Related
  • 2000-10-13 JP JP2001536716A patent/JP5535417B2/ja not_active Expired - Fee Related
  • 2000-10-13 WO PCT/US2000/028486 patent/WO2001035250A2/en not_active Ceased
  • 2000-10-13 AU AU10868/01A patent/AU1086801A/en not_active Abandoned
• 2004
  • 2004-09-27 US US10/951,307 patent/US7424675B2/en not_active Expired - Fee Related
  • 2004-10-21 US US10/970,438 patent/US7302640B2/en not_active Expired - Fee Related

Also Published As

Similar Documents

Legal Events