KR20130107325A - Spell-check for a keyboard system with automatic correction - Google Patents

Abstract

User input is received that specifies a path that is continuously tracked across the keyboard represented on the touch-sensitive display. An input sequence is solved, comprising a key tracked by defined criteria and auxiliary keys proximate the tracked key. For one or more candidate entries of the defined vocabulary, a set-edit-distance metric is calculated between the input sequence and each of the candidate entries. Various rules are specified when imposing or not penalizing the calculation of the set-edit-distance metric. Candidate entries are ranked and displayed according to the calculated metric. The features described herein may be implemented as a device, programmed product, method, circuit, or a combination thereof.

Description

Spell check for keyboard systems with automatic correction {SPELL-CHECK FOR A KEYBOARD SYSTEM WITH AUTOMATIC CORRECTION}

Cross-reference of related patents

The present invention claims the priority of U.S. Patent No. 12 / 939,918, filed November 4, 2010, which is hereby incorporated by reference in its entirety.

The present invention relates to data input devices. More specifically, the present invention relates to a spell-check mechanism for keyboard systems with automatic correction capability.

Classic spell check ("edit distance") techniques for substituted / added / missing letters have a relatively long history. For example, K. Kukich's Techniques for Automatically Correcting Words, ACM Computing Surveys, Vol 24, No. 4 (December 1992); JL Peterson's Computer Programs for Detecting and Correcting Spelling Errors, The Communicators of the ACM, Volume 23, No. 12 (December 1980); And Spelling Corrections in J. Daciuk, International Construction of Finite-State Automate and Transducers, and their Use in the Natural Language Progressing (1988).

Classic spell checking techniques, however, can only handle a certain number of differences between the typed word and the exact word intended. Since the best modification candidate is assumed to have minimal changes, the spell checking algorithms are for example the movement of a typist's finger on the keyboard, or an urgent and incorrect tapping on a touch screen keyboard, and thus You are embarrassed by typing almost all the letters incorrectly.

In particular, in order to limit the amount of computational processing on low performance mobile devices, implementations of classical algorithms assume or impose constraints to reduce ambiguity and thus the number of candidate words considered. For example, they may depend on the first letters of the word to be corrected or may strictly limit the size of the vocabulary.

Another form of automatic error correction, useful for both keyboards and telephone keypads on touch sensitive surfaces, calculates the distance between each input location and near letters and compares the entire input order for possible words. Words whose letters are close to input locations, combined with the highest frequency and / or recency of word usage, are the best candidates for modification. This technique easily corrects both moved fingers and urgent tapping. It can also provide reasonable word completion even if the initial letters are not all entered correctly.

The following patents describe the use of a "Sloppy Type" engine for clarity and automatic correction obscure keys, soft keyboards and handwriting recognition systems: B. In the name of the invention "Keyboard System With Automatic Correction". US Patent 6,801,190 (October 5, 2004) filed by Alex Robinsom, Michael R. Longe; U.S. Patent 7,319,957 (2008) filed under the name "Handwriting And Voice Input With Automatic Correction" by U.S. Patent 7,088,345 (August 8, 2006) and U.S. Patent 7,277,088 (October 2, 2007), and Robinsons. January 15, and US Patent No. 11 / 043,525, filed January 25, 2005. See also US Pat. No. 5,748,512 (5 May 1998) to Garrett R. Vargas, filed under the name “Adjusting keyboard”.

In addition, the following patents include combinations of manual and speech input for text clarification: US Patent No. 11/143, filed June 1, 2006 by Longe et al. Under the name " Multimodal Disambiguation of Speech Recognition " / 408; And US patent application Ser. No. 11 / 350,234, filed Jul. 2, 2006, by Stephanick et al., Entitled "Method and Apparatus Utilizing Voice Input to Resolve Ambiguous Manually Entered Text Input."

The "sloppy type" technique described above uses distance based error correction on whole words. Assuming the length of the input sequence is the same as the length of the intended word and assuming that each input position is in the proper order, this helps to compensate for the increased ambiguity introduced by considering multiple adjacent letters for each input. . But in addition to small target errors, people also change keys when they type, double tap keys, miss a key entirely, or misspell words.

It may be desirable to provide a mechanism for handling all types of typing errors in such a way as to provide both corrective and acceptable execution.

User input is received that specifies a path that is taced continuously across the keyboard represented on a touch sensitive display. An input sequence is solved, comprising a key tracked by defined criteria and auxiliary keys proximate the tracked key. For one or more candidate entries of the defined vocabulary, a set-edit-distance metric is calculated between the input sequence and each of the candidate entries. When rules impose or do not impose a penalty on calculating a set-edit-distance metric, various rules are specified. Candidate entries are ranked and displayed according to the calculated metric.

The features described herein may be implemented as a device, programmed product, method, circuit, or a combination thereof.

1 is a flow diagram of a spell checking method for a keyboard with automatic correction according to one embodiment of the present invention.
2 is a hardware block diagram of an input system with spell checking and automatic correction according to one embodiment of the invention.
2A is a block diagram of a digital data processing machine according to one embodiment of the invention.
2B illustrates a preferred storage medium in accordance with one embodiment of the present invention.
2C is a background of a preferred logic circuit in accordance with one embodiment of the present invention.
3 is a table illustrating a standard edit-distance calculation between an input word and a target word using a matrix as a tool, according to one embodiment of the invention.
4 is a table illustrating set-edit-distance calculation for input on a 12-key mobile phone according to one embodiment of the invention.
5A and 5C are diagrams for describing a stem edit-distance and a stem set-edit-distance according to an embodiment of the present invention.
6 is a flow diagram illustrating steps for performing stem-edit-distance calculations and incremental filtering to identify a candidate word in accordance with one embodiment of the present invention.
7 is a matrix illustrating an example for the word "misspell" using standard edit-distance in accordance with one embodiment of the present invention.
8 is a matrix illustrating a method for finding standard edit-distance values based on a cell computed in accordance with one embodiment of the present invention.
9 is a matrix illustrating when words to be compared completely match according to an embodiment of the present invention.
10A-10B are a series of matrices illustrating when there is a mismatch between words being compared according to a date and time example of the present invention.
11 illustrates a rotated / transformed matrix space in accordance with one embodiment of the present invention.
12 illustrates a method for finding standard edit-distance values for the rotated matrix of FIG. 11 in accordance with an embodiment of the present invention.
13 is a table illustrating a combination of adjacent input settings for a language database search screen function according to an embodiment of the present invention.
14 is a length independent screening map for an input length 9 according to an embodiment of the present invention.
FIG. 15 is a length-dependent screening map for a target word of length 6 and an input length 9 according to an embodiment of the present invention.
FIG. 16 is a series of screens illustrating set-edit-distance spell correction with automatic correction in a region, according to one embodiment of the invention.
17 is a keyboard illustrating a trace according to an embodiment of the present invention.
18 shows a layout for a set-edit-distance matrix calculated for tracked input according to one embodiment of the invention.
19-23 illustrate a set edit distance matrix and various shadow matrices in accordance with one embodiment of the present invention.
24 is a flowchart illustrating a preferred operation procedure for resolving user input input through tracking technology according to an embodiment of the present invention.
25-26 are screen shots of a keyboard illustrating various approaches for determining auxiliary keys according to one embodiment of the present invention.
27-30 illustrate a set-edit-distance and various shadow matrices in accordance with one embodiment of the present invention.

Glossary of Terms ( Glossary )

For purposes of discussion herein, the following terms have the meanings associated with them.

Edit distance (also "standard" edit distance)-A sufficiently proven algorithm to compare two strings and determine the minimum number of changes needed to make one equal to the other.

The following abbreviations may be used here and in the drawings:

T-replaced (two consecutive letters are swapped)

I-inserted (Inserted, adds characters that were not in another string)

D-deleted (Deleted, missing additional characters from a string)

S-replaced (substituted, replacing letters with another letter at the same location)

X-the target cell to be counted

Improved editing distance, or set-editing-distance (or “fuzzy compare”) — the subject of this patent; Improved editing distance using a set of letters (each with an optional probability for each) to represent each input, rather than a single letter as at standard editing distance, and other optimizations.

Mode-operation status, for this example, one of two is specified, "exact" (as with the standard edit distance, just the exact-tab character / from each input event to match each candidate word). Value) or "regional""set-based" (using multiple letters / values per input); The mode can be user or system specific.

In-Area Input-A method (or event) that includes nearby / peripheral letters (with optional probabilities) in addition to the letters / keys that are actually tapped / pressed.

Set-based—use of multiple character values, rather than just one, to represent each input; Each set member can have a different relative probability; The set can also include variants with accent marks, for example, of the basic letters appearing on the keys.

"Classic comparison", "classic match", slope shape, or "correction in area"-full word matching using automatic correction taking into account the nearby letters above; In general, the number of inputs is equal to the number of letters in each candidate word (or word stem of the completed word).

Filter or Screen-A rule to simplify the entire comparison or search by identifying and removing words that do not eventually add to the selection list.

KDB-Keyboard Database; Information about the keyboard layout, the level of ambiguity surrounding each letter, and nearby letters for each letter.

LDB-Linguistic Database, that is, the main vocabulary for a language.

"Word tap frequency"-the contribution of the physical distance from the pressed keys to the likelihood that a word is a target word.

Argument

One embodiment of the present invention provides standard edit distance spell checking algorithms that work with probability-based automatic correction algorithms and data structures for obscure keypads and other predictive text input systems. Embodiments of the present invention also provide strategies for optimizing and placing different types of results.

1 is a flow diagram of a spell checking method for a keypad with automatic correction. FIG. 1 illustrates a user / input including an input sequence input by a user via a data entry device 105, which may be ambiguous. At least one dictionary 115 is also provided as a source of object meanings for the user's input. On each user input event 100 a user input sequence is provided to the system of the present invention. Each source 110, such as the dictionary 115 discussed above, is queried. As input to the system of the present invention, potentially on each user input event, all words 120 in each dictionary are provided in turn.

Upon receiving these inputs, the system executes the increased filtering and editing distance and in-region / probability calculations (140), discarding any words that do not meet the maximum threshold for similarity with the inputs. . The system then compares the results for the input sequence and dictionary inputs with the other best matches in the word selection list and discards the word if it is ranked too low on the list (140). If the list is full, the lowest ranked word in the list is dropped, and the word is inserted into the list based on the ranking (150). The list is then presented to the user.

2 is a hardware block diagram of an input system 200 with spell checking and automatic correction. Input device 202 and display 203 are coupled to processor 201 through appropriate interface circuitry. Optionally, speaker 204 is also coupled to the processor. The processor 201 receives input from the input device and manages all output to the display and speakers. The processor 201 is coupled to the memory 210. The memory includes a combination of temporary storage media, such as random access memory (RAM) and permanent storage media, such as read-only memory (ROM), floppy disk, hard disk, or CD-ROM. Memory 210 includes all software routines to control system operation. Preferably, the memory comprises an operating system 211, correction software 212 for calculating the editing distance and in particular performing a spell check, and associated vocabulary modules 213 discussed in detail below. Optionally, the memory may include one or more application programs 214, 215, 216. Examples of applications include word processors, software dictionaries, and foreign language translators. Speech synthesis software may also be provided as an application that allows the communication input system to have all modification capabilities to function as communication support.

Preferred Digital Data Processing Unit

Data processing entities such as a central processing unit (CPU) 201 may be implemented in various forms. Some embodiments may be a general purpose processor, digital signal processor (DSP), application specific semiconductor (ASIC), field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or excitation It includes any combination of those designed to perform the functions described in. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, eg, a combination of a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors with a digital signal processor core, or other such configuration.

As a more specific example, FIG. 2A illustrates a digital data processing apparatus 220. Device 220 includes a processor 222, such as a microprocessor, personal computer, workstation, controller, microcontroller, state machine, coupled to digital data storage 224. In the present embodiment, the storage device 224 includes not only non-volatile memory but also fast-access memory. Attribute-access storage 226 may be used to store programming instructions executed by processor 222, for example. Storage devices 226 and 228 may be implemented by various devices, such as those described in detail in conjunction with FIGS. 2B-2C. Many alternatives are possible. For example, one of the components 226, 228 may be removed and the memory 224, 226 and / or 228 may also be provided in the mounted processor 222, or external to the device 220. Can be provided.

Device 220 may also include connectors, lines, buses, buffers, electromagnetic links, networks, modems, converters, infrared ports, antennas, or other means for exchanging data with other hardware external to device 220. Like, input / output 221.

Storage media ( Storage Media )

Various instances of digital data storage may be used to provide storage, such as memory 210, to implement storage 24 and 228 (FIG. 2A), and the like. Depending on its application, such digital data storage can be used for various functions, such as for storing data or for storing machine-readable instructions. Such instructions may help to perform various processing functions on their own or may help install software programs on a computer, which software program may then be executable to execute other functions related to the present invention.

In any case, the storage medium can be implemented by almost any mechanism for digitally storing machine-readable signals. One example is an optical storage device, such as a CD-ROM, WORM, DVD, digital optical tape, disk storage 230 (FIG. 2B), or other optical storage. Another example is a direct access storage device, such as a conventional " hard drive ", conventional RAID (redundant array of inexpensive disks), or other direct access storage device (DASD). Another example of digital data storage includes electronic memory, such as ROM, EPROM, flash ROM, EEPROM, memory registers, battery back-up RAM, and the like.

A preferred storage medium is coupled to the processor such that the processor reads information from and writes information to the storage medium. In the alternative, the storage medium may be integral to the processor. In still other embodiments, the processor and the storage medium may reside in an ASIC or other integrated circuit.

Logic circuit

In contrast to a storage medium containing machine-executable instructions, such as described above, another embodiment uses logic circuitry to implement the processing features described herein. Depending on the specific requirements of the application in the fields of speed, cost, equipment cost, etc., this logic can be implemented by constructing a custom semiconductor with thousands of small integrated transistors. Such custom semiconductors may be implemented with CMOS, TTL, VLSI, or other suitable configuration. Other alternatives include digital signal processing chips (DSPs), discrete circuits (such as resistors, capacitors, diodes, inductors, and transistors), field programmable gate arrays, programmable logic arrays (PLAs), programmable logic devices (PLDs), and the like. It includes. 2C illustrates one embodiment of a logic circuit in the form of an integrated circuit 240.

In-zone modifications Combined  Edit distance

Edit-distance is the number of operations needed to replace one string with another. In essence, this is the number of edits that must be made manually with the pen to correct misspelled words. For example, to modify the input word "ressumt" to the target word "result", two edits must be made, "s" must be controlled, and "m" must be changed to "l". Therefore, "result" is edit distance 2 from "ressumt".

A general technique for determining the edit-distance between an input word and a target word uses a matrix as a tool (see FIG. 3). The approach compares the letters in the input word with the letters in the target word and gives the overall edit-distance between the words in the lower rightmost component of the matrix. The details of the calculations are complex, but generally the edit-distance (indicated by the number in the diagonal's components) increases each time the parts of the words begin to look different (and smaller values mean more similar). When working the matrix from the upper right to the lower left, the edit-distance does not increase if the letter in the target word is the same as the letter in the input word. If the letters in the target word are not the same, the edit-distance increases according to standard rules. As a result, the overall edit distance is the bottom rightmost element (bold outline).

Such a concept now extends to obscure input, where each input corresponds to a set of letters rather than single letters. An example of this is a text input system on a cell phone that allows a user to press keys corresponding to the letters that the user wants to enter, which solves the ambiguity inherent in the fact that the keys have multiple letters associated with them. The new term "set-edit-distance" refers to the extension of the edit-distance concept for ambiguous input. To illustrate the set-edit-distance, it is assumed that the user of the text entry system of the cellular phone presses the keys 7, 3, 7, 7, 7, 8, 6 and 8 while attempting to enter the word "result". This spelling correction on the ambiguous system finds the words with the smallest set-edit-distance for the input key sequence. The technique is similar to that of edit-distance, but instead of comparing the letters in the target word with the letters in the input sequence, the letters in the object are compared against a set of letters represented by the input key. If the target character is in the input set, the set-edit-distance does not increase. If the target character does not exist in the input set, the set-edit-distance increases according to standard rules. The matrix corresponding to the set-edit-distance is shown in FIG. 4, with the result at the bottom rightmost (bold outline).

Although the embodiment of FIG. 4 uses a key input on a mobile phone to illustrate the concept of set-edit-distance, this concept may also be a set of letters surrounding a key pressed on a QWERTY keyboard, or letters returned from a letter recognition engine. It applies to other obscure systems, such as a set of. Further, while the above embodiment assumes that the letters in the set are all equally likely, the system can be extended to incorporate the letter probabilities in the final set-edit-distance score.

In such an extended system, the input sequence may be represented as an array of one or more letter + probability pairs. The probability reflects the likelihood that the character identified by the system is the character intended by the user. US Patent 7,319,957 (January 15, 2008) by Robinson et al. In the name of the invention "Handwriting And Voice Input With Automatic Correction" and January 2005 by Robinson et al. In the name of the invention "Handwriting And Voice Input With Automatic Correction". As described in US Patent No. 11 / 043,525, filed 25, each of which is incorporated herein by reference. Probability may be based on one or more of the following:

스타일러스(stylus) 또는 손가락 탭 위치로부터 터치 스크린상에 표시되는 키보드 상의 각각의 인접한 글자까지의 데카르트 거리(Cartesian distance), 인접한 글자들의 빈도, 및/또는 각각의 글자 주위의 탭들의 분포;

Cartesian distance, the frequency of adjacent letters, and / or distribution of tabs around each letter from the stylus or finger tap position to each adjacent letter on the keyboard displayed on the touch screen;

알파벳의 인접 글자들의 지정된 파이 슬라이스(pie slice)들에 대한 조이스틱 틸트 방향 사이의 방사상 거리;

Radial distance between joystick tilt directions for designated pie slices of adjacent letters of the alphabet;

손으로 쓴 글자 및 가능한 글자 형태들/템플레이트들 사이의 유사성 정도, 예를 들면, "ink trail"은 글자 'c'(60% 확률)가 가장 유사한 것으로 보이나, 또한 'o'(20%), 'e'(10%), 'a'(10%)와 같은, 다른 글자들일 수 있다; 및

The degree of similarity between handwritten letters and possible letter types / templates, for example "ink trail", appears to be the most similar to the letter 'c' (60% probability), but also to 'o' (20%), other letters, such as 'e' (10%), 'a'(10%); And

글자/문자소(grapheme)가 음성 인식 전위(front-end)에 의해 처리되는 음소(phoneme) 또는 전체 단어 발음(utterance)에서 표현되는 확률

The probability that a letter / grapheme is represented in the phoneme or whole word utterance processed by the speech recognition front-end

Thus, the set-edit-distance is a standard edit distance applied to obscure sets, where penalties are assigned to each difference between the input lexical word and the target lexical word. Instead of the question "Is this letter different?", The question is "is this letter one of the possible candidates in the probability set?"

Thus, one embodiment applies the following algorithm:

만일 매치를 야기하는 두 개의 확률 변환이 존재하면, 가장 낮은 편집 거리를 갖는 하나를 선택한다.

If there are two probability transforms that cause a match, the one with the lowest edit distance is selected.

만일 글자가 입력의 확률 세트 내에 존재하면, 또한 그러한 글자를 위한 영역-수정 확률 스코어를 계산한다..

If a letter is in the set of probabilities of the input, it also calculates a domain-corrected probability score for that letter.

맞춤법 수정 탭 빈도를 계산하기 위하여 단어 내의 모든 글자를 위한 모든 영역-수정 확률 스코어를 누적한다.

Accumulate all area-correction probability scores for all letters in a word to calculate the spelling correction tap frequency.

제로-세트-편집-거리 단어, 즉 어휘 단어 내에 동일한 단어 길이 및 각각의 단어가 입력 확률 세트들 내에 존재하는 단어들을 위하여, 탭 빈도만이 사용된다.

For zero-set-edit-distance words, ie words where the same word length in the lexical word and each word is in the input probability sets, only the tap frequency is used.

Multiple values are calculated or accumulated for matching and word list placement steps:

1. set-edit-distance;

2. of comparison, tap frequency;

3. stem edit-distance;

4. word frequency; And

5. Sources, eg dictionaries.

The tap frequency (TF) or stem of a word can be calculated as follows:

TF = probability of letter 1 * probability of letter 2 * ... (1)

This is similar to standard probability set auto-correction calculations, but the edit distance algorithm generates alternatives and then the largest calculated frequency among these alternatives is selected.

4 uses a comparison between the set-based input sequence and the entire target word. This concept can also be applied to compare sets of inputs relative to the start of the target word (stem). This allows the system to begin predicting spelling corrections before the user enters the entire input sequence. This is called stem-edit-distance. 5A-5B show partial input sequences. In these figures, the letters 'a' and 'c' may be members of the same set based on physical proximity on the touch screen qwerty keyboard, while 's' and 'g' are not. Since the letter 's' at the third position of the target word is in the set for the third input in FIG. 5A, the stem set-edit-distance between the input and the target word is zero. Since the third letter 's' is not in the same set for the third input in FIG. 5B, the stem set-edit-distance between the input and the target word is one.

The stem edit-distance is the edit distance value for the clearly entered or most probable letters, mainly the extract-tap value from each set of input probabilities, compared to the corresponding letters of the long target word. In this case, the most probable character from each input for the touch screen qwerty keyboard is the extract-tab character. Since the letter 's' at the third position of the target word is not the same as the extract-tap value for the third input in FIG. 5A, the stem edit-distance between the input and the target word is one. Similarly, the stem edit-distance between the words to be input in FIG. 5 is also one.

Sets for stem set-edit-distance may also be language specific. For example, variations in which the accents of letters in French are revealed may be members of the same set. FIG. 5C shows one embodiment where the variant of 'e' maps to the same key, resulting in zero stem set-edit-distance between the input and target words.

One embodiment of the present invention also provides a number of innovative strategies for adjusting the placement of words in a selection list to reflect the user's intent or input style. For example, results can be biased in one of two ways:

전 단어 우선권 - 예를 들면, 복잡하거나 또는 낮은 촉각 피드백을 갖는 좋지 못한 키보드, 및/또는 빠르거나 슬로피 타이피스트를 위하여, 결과들은 영역 내, 즉, 니어-미스(near-miss), 모든 입력의 수정 및 소수의 단어 완성들을 강조한다; 및

Full word priority-for example, for a bad keyboard with complex or low tactile feedback, and / or a fast or slippy typist, the results are in the area, ie near-miss, of all input. Emphasize corrections and few word completions; And

완성들의 촉진 - 뛰어난/정확한 키보드, 및/또는 완전히 신장시키기 위한 완성들을 원하는 느린, 세심한 타이피스트를 위하여, 결과들은 즉, 어느 정도는 정확한 탭, 입력 시퀀스 등을 기초로 하는 단어 완성들을 강조한다

Promotion of Completions-For a slow, meticulous typist who wants an excellent / accurate keyboard, and / or completions to fully elongate, the results emphasize words completion to some degree based on correct taps, input sequences, etc.

One embodiment of the present invention provides spell checking features that allow such systems to work with non-desktop devices, in particular with the "sloppy type" technique described above to be more useful for typing errors and all typists. The "sloppy type" system provides an enhanced text input system that uses word-level clarity to automatically correct inaccuracies in user keystroke entries. Specifically, the “sloppy type” system includes (a) a touch sensitive surface that includes an auto-correction keyboard area that includes a plurality of letters of the alphabet, each of the plurality of letters having known coordinates within the auto-correction keyboard area. Corresponding to a location having a position, each time a user contacts a user input device within an auto-correction keyboard area, a location associated with the user contact is determined and the determined contact location is added to the current input sequence of contact locations; (b) having a memory in contact with a plurality of objects, each object being a word or a string of one or more letters forming part of a word, each object also associated with a frequency of use; (c) an output device having a text display area; And (d) a processor coupled to the user input device, memory, and output device, the processor comprising: (i) contact locations and auto-correction keyboard area for each determined contact location in the input sequence of contacts; A distance value calculation component for calculating a set of distance values between one or a plurality of letters within and corresponding known coordinate positions; (Ii) for each generated input sequence, identify one or a plurality of letters in memory and, for each of the one or a plurality of identified candidate objects, based on the calculated distance values and the frequency of use associated with the object. A word evaluation component that evaluates each identified candidate object by calculating a matching metric and ranks the evaluated candidate objects based on the calculated matching metric values; And (iii) identify (a) one or a plurality of candidate objects according to their evaluated ranking, (b) present the identified objects to the user, and output the text to the text display area on the output device. And an optional component for allowing selection of one of the objects shown.

Optimizations

In theory, given a sufficiently large edit distance score, every word in the vocabulary can be considered a modification. However, database processing must occur in real time each time a user types, and especially for mobile devices, there are limitations on available processing power and working memory. Therefore, it is important to optimize all parts of the combined edit distance algorithms and to eliminate processing steps when possible. For example, the highest level criterion for discarding possible word matches is to allow only one edit / modification for every three actual inputs, up to three edits for any word being compared.

Other performance improvements may include, for example, without limitation:

편집 거리 계산들을 최소화하기 위한 전략들, 예들 들면, 비교가 전부 거부되도록 허용할 수 있는 셀들을 계산하는 첫 번째 통과(first pass).

Strategies for minimizing edit distance calculations, e.g., a first pass that calculates cells that can allow the comparison to be rejected entirely.

시스템은 사용자가 또 다른 글자를 입력할 때와 같이, 이전 통과의 결과들로부터 시작하거나; 또는 임시로 예를 들면, 사용자가 엔트리를 멈출 때까지, 단축되거나, 부분적인, 또는 흐릿한 선택 리스트를 나타내는, 이전 단어들을 줄인다,

The system starts from the results of the previous pass, such as when the user enters another letter; Or temporarily reduce the previous words, for example, indicating a shortened, partial, or blurred selection list until the user stops the entry,

예를 들면, 덜 엄격한 대부분에서, 필터링의 레벨들은 편집 거리 매트릭스 계산들이 완료되기 전에, 동안에, 또는 이후에 적용되는데, 예를 들면:

For example, in most less stringent levels of filtering are applied before, during, or after edit distance matrix calculations are completed, for example:

정확한 첫 번째 단어, 그렇지 않으면 고려로부터 대상 단어를 철회;

Withdraw the target word from the exact first word, otherwise consider;

확률 세트 내의 첫 번째 단어 니어-미스, 영역 내;

First word near-miss in the probability set, in the region;

어휘 단어의 첫 번째 글자는 반드시 첫 번째 두 입력들 중의 하나와 매치하여야 하는데, 예를 들면, 하나의 첨가, 하나의 드롭(drop), 또는 하나의 치환된 쌍을 허용한다;

The first letter of a lexical word must match one of the first two inputs, for example allowing one addition, one drop, or one substituted pair;

어휘 단어의 첫 번째 글자는 반드시 첫 번째 두 입력들 중의 하나의 세트 내에 존재하여야 한다;

The first letter of the lexical word must be in one set of the first two inputs;

다른 필터링 개념들 및 변형들이 적용될 수 있다; 및

Other filtering concepts and variations can be applied; And

비 필터링

Non-filtering

The word frequency can be approximated based on Zipf's Law, which explains that given some corpus of natural language pronunciation, the frequency of a word is inversely proportional to its rank in the frequency table. Thus, the most frequent word occurs approximately twice as often as the second, and the second most often occurs twice as often as the fourth. In one embodiment, an approximation is used, rather than the value stored for each word in the lexical database:

F _n = F ₁ / n (the frequency of the Nth word is the frequency of the first word divided by the word position) (2)

Other adjustable configuration parameters may include:

니어-미스 섹션 당 단어 완성들의 수;

Number of word completions per near-miss section;

맞춤법 수정의 수; 및

Number of spelling corrections; And

맞춤법 수정 모드, 표준 편집 거리 또는 세트-편집-거리(글자 확률들을 갖거나 혹은 갖지 않는).

Spelling correction mode, standard edit distance, or set-edit-distance (with or without letter probabilities).

Run spell correction

Spelling corrections on large word lists is central processor intensive and even more so when memory is limited. Thus, in order to achieve acceptable performance, the entire system must be optimized based on the selected spelling correction features. The resulting system can thus be quite rigid from the feature point of view. Without specific optimization, performance can be one or two grades worse.

Most spell correction execution depends on:

허용된 편집들, 모드들, 및 필터들과 같은, 맞춤법 수정 특성들

Spelling correction properties, such as allowed edits, modes, and filters

"퍼지 비교" 기능(단어가 입력과 일치하는지를 결정하는)

"Fuzzy compare" function (determines whether a word matches input)

낮은 레벨 언어 데이터베이스 검색 기능

Low level language database search function

언어 데이터베이스 포맷(구조 및 행동)

Language database format (structure and behavior)

언어 데이터베이스 내의 단어들의 수 및 그것들의 길이 분포

The number of words in the language database and their length distribution

언어 데이터베이스에 대하여 키보드 데이터베이스는 얼마나 애매한가.

How obscure is the keyboard database?

Each of these elements is described in more detail in the next section.

Spelling Corrections

Allowed edits

The number of edits allowed is a very important performance factor. The more edits, the more ambiguity there is in the comparison, so more words match and get into the selection list for prioritization. If the comparison is too lenient, the effect is that too many unwanted words get into the list.

In a preferred embodiment, the number of edits allowed is related to the input length and one edit is given for every third input up to three. The parameters of one input per three inputs are assumed throughout the following embodiments.

Modes  And filters

Modes and filters are used to control performance as well as the result set. Two examples of modes are exact input and area. For example, on a touch screen soft keyboard, the user can not only accurately tap on the desired letter but also represent an almost accurate area of the letter. In the correct input mode, only the correct tab character from each user input is considered. In the in-zone mode, some or all of the nearby letters displayed by each user input are considered.

Spelling corrections to the correct input reduce ambiguity and make candidates appear more input (even if the input is incorrect). This is effective for a keyboard database that features accurate tap values, such as a touch screen soft keyboard. 12 key systems (for standard telephone keypads) cannot have any useful correct tap values; Each key press can be represented by a digit of a key instead of one of the letters, and one cannot feel that one letter on each key will be more than the other, which is the intended letter.

Unfortunately for 12 key systems, the keyboard databases act as a mode layout in the general area, that is, each input produces at least three letters per set, and more letters are produced when the accented words are included, but correct It does not have the correct tap value that can be used for input mode and filtering.

The filter is a screening function that ends another consideration of candidate words if it does not meet the minimum criteria achieved. For example, ONE / TWO filters are mostly for performance improvement, allowing the first letter in a word to be closely associated with the first or second input and rejecting any candidate words that do not match.

"Fuzzy compare" function

The fuzzy comparison function allows for a specific difference, edit distance between the words being entered and compared. The concept is to calculate the editing distance and to pass or reject words based on their values.

Accurate calculation of the editing distance is an expensive practice. The solution is to locate the screening mechanism before the actual calculation. To the extent denied "down" it is acceptable, but if possible a "up" denial should be avoided. Words that pass screening because of "deny down" are removed after the actual distance calculation.

Fast screening is important to maintain acceptable performance on each key press. Potentially, large amounts of words can come in for screening and normally only a few are delivered. Thus, for good performance all of the screenings are also very efficient. What has been done after screening is a less important method of implementation, but there is still a significant amount of data coming in, especially for the particular input combinations where thousands of words enter into the select list insertion function.

In one or more embodiments, the spelling correction works with the probability set comparison logic of the automatic corrections in the area. There are words that are accepted by set comparisons that are not accepted based on the spelling correction calculation. This is the case for input within a region when spell correction is set in the correct input mode or when using the correct filters. Word completion is also simpler for classical comparisons but costs the edits in spelling correction.

In a preferred embodiment, the fuzzy comparison steps are as follows:

1. Screens for words that are too short

2. Screens for set based matches

3. Stem Edit-Distance Calculation

4. Screen for ONE / TWO

5. Screen for set-edit distance

6. Screen for position-locked letters

7. Set-Edit--Calculate Distance and Frequency

8. Stem Edit-Distance Calculation

These steps are shown as the flow diagram of FIG. 6, representing one implementation of the calculations 130 of FIG. 1.

Screening for classical comparisons, processing of word completions, etc., are placed in step 2 before further spell correction calculations. This excludes all "classic" complexity from subsequent code. This also means that if spell correction is turned off, all other calculations can be skipped.

The algorithm is drawn as comparing two words against each other. In most embodiments, this is generalized so that one word corresponds to input codes. In the sample matrices in the figures referenced below, the input sequence is shown vertically. Thus, rather than each single input word position being a single letter, as with standard editing distances, it is actually a set of letters that are either obscure or correspond to input in the region. The comparison produces a match if any character in the set matches.

1. Screens for words that are too short

If the word is too short for spelling correction, that is, less than the input length minus the available editing distance, it can be rejected immediately.

2. Screens for set based matches

This is an iteration over an input sequence, which proves to match the corresponding position in the word being compared, i.e. each letter in the candidate word must be present in each input set.

If there is a non-match and the word is too long for spell correction, that is, if it is longer than the input length + available editing distance, it can be rejected immediately.

3. Stem Edit-Distance Calculation

This is an iteration over all the signs in the input sequence and is executed only when there is a set based match. Every difference from the correct tap value increases the stem distance, for example the candidate word "tomorrow" may have a stem distance of 0 for accurate tap input of "tom" and a stem distance of 1 for "tpm". have. The word tap frequency is also calculated during the iteration.

If it is a valid classic match, then the "fuzzy comparison" of candidate words is completed at this point. Candidate words are inserted into the selection list.

4. Screen for ONE / TWO

This is a quick check to see if the first letter in a word matches the first ONE or TWO input codes. If not, the word is rejected.

5. Screen for set-edit distance

Conceptually this is a very simple task because the improved editing distance follows the conventional definition of using insertions, deletions, and substitutions and substitutions (the latter are mainly included for text entry modification). It is very difficult to implement this in an efficient way.

The conventional method of calculating the edit distance uses a matrix. One embodiment is shown in FIG. 7. All edges (gray figures) are predefined and always the same. The remainder is calculated across the columns first, from left to right and from top to bottom. Each individual position is calculated by taking the minimum values corresponding to the insertions, deletions, substitutions, and substitutions. Substitution and substitution values are conditioned on whether a match exists for such locations. The resulting edit distance is known as the lower right corner, in this case "2".

To find the values based on the cell to be calculated, ie the cell indicated by "X" in FIG. 8: The cost to take a replacement ('D') cell is 0 or 1 depending on whether a match exists. The substitution ('T') cell can only be taken when both letters match, ie both current and previous letters and the cost is one. Insertion ('I') and deletion ('D') are also one cost each. Thus, the cost of a cell is the cost already calculated for that cell as well as the additional cost just mentioned.

This is a computationally very expensive method, especially for calculating long word distances. In one embodiment, the maximum allowable editing distance is set so that up to 1% of words pass such a limit. If the allowed distance is too high, the entire word list can make it into the selection list and the entire concept of spelling correction is lost. Thus, initially the correct distance is not important, rather it is important whether the result is below or above the rejection limit. More effort may be spent calculating the exact distance, frequency, etc. for the few words that pass this test.

The purpose of the screening step is to verify as soon as possible that the resulting distance is above the rejection limit.

As shown in FIG. 9, except for the length, consider when the words being compared match. It is not possible for any of the cells to have a low value. A comparison of six and nine words in length results in an edit distance of three, as expected.

This initial matrix can be used when calculating any two words. In practice only the values in the cells selected for comparison need to be updated according to the method. The goal is to push the lower right cell above its rejection limit. In order to do so, some of the cells that depend on obtaining this value actually have a high value, and therefore must be proved recursive.

For this embodiment, the length difference is 3 and the first letter does not match (change the first 'x' to 'y' in FIG. 10A), the rejection can be proved by counting four cells; The remainder of the relevant cell updates is included. The iteration in FIG. 10B shows the effect on the recalculated cells (bold outlines) and other dependent cells in each iteration.

As a result, towards the center diagonal and the diagonal with the resulting value get increased values. This occurs whenever the last cell, which gives the lowest value in another cell, is increased as a result of the complete comparison mismatch.

The illustrated matrices only describe what happens when there is a word length difference. If the length difference is zero, the central diagonal becomes the main matrix and a value large enough to support, ie affect the calculation, must come from both sides of the resulting diagonal to prove rejection.

Diagonal lines in the calculations make it more difficult to optimize data access patterns (accessing real memory corresponding to locations). Operation in the rotated / transformed matrix space is another optimization; Reference is made to FIG. 11. Cells in the center diagonal (bold outline) become a single column. Default values for the corner cells, if referenced, are added with new "9" s (shown in gray) to provide a sufficiently large value that immediately exceeds the maximum possible edit-distance. In this deformed space, cell calculation relationships change as shown in FIG.

6. Screen for position-locked letters

Since it is not executed on the spelling correction candidate, there is a need to identify input signs with locked positions, i.e., positions that do not allow the movement or change of a value. It is just a repetition of input symbols with locked positions, checking that they match. If not, the word is rejected.

7. Set-Edit--Calculate Distance and Frequency

The algorithm for screening for edit distance can be modified to calculate other things such as edit distance and word frequency. However, it should not be merged into screening code. Such code must be kept separate and optimized for pure screening. Different versions apply to more thorough, screening words, because they must evaluate different cells and choose the best option for low distance and high frequency. They must also handle things, such as possible locked sign values (values, not positions).

If the set-edit-distance value exceeds a certain limit, the candidate is rejected.

8. Stem Edit-Distance Calculation

It is also a variation of the screening algorithm for two reasons:

First, the stem distance can be very different because it is always based on the exact match. Thus, the value may be higher than the intended maximum for distance. Distance values higher than the maximum are not completely accurate because of algorithmic optimizations but are still good enough.

Secondly, the stem distance is also different in that it may not take into account the total length of the candidate words. In order to be compatible with non-spelling correction words, the stem distance calculation will stop at the length of the input. Some additional checking around the end cell is needed to get the minimum value following insertions and deletions.

Low level language database search function

Fuzzy comparison functions can be made very efficient in screening and calculation, but are not sufficient alone for good performance, especially on embedded platforms. Depending on the input, almost every word in the vocabulary can be potential spelling correction candidates. This typically occurs when you enter the ninth or tenth input in most languages when one edit is allowed per three inputs.

All words having an input length of 9 to lengths 6-12 are potential spelling correction candidates and all words longer than 12 are potential completion candidates. For example, at input length 9, more than 70% of the Finnish vocabulary can be considered for comparison based on spelling correction and another 20% based on word completion. This creates a significant efficiency problem because spelling modifications require the most computational effort. The following strategies seek to increase the efficiency of the database search process by integrating one or more of the screening functions described previously.

Search strategy for non-spelling corrections

Incorporated herein by reference. Unruh; Erland, Kay; Preferred implementation of a lexical database, as described in US Patent No. 11 / 379,354, filed April 19, 2006, filed by David Jon under the name "Efficient Storage and Search Of Word Lists and Other Text." The example is designed and optimized without spelling correction. The overall input length is mapped directly to interval streams and the sparest streams are first visited to help with fast jumping in the word list. Once a match exists, complete characters can be selected from streams that do not match the input.

With this strategy, words that are too short are automatically skipped because they do not have letters that match the corresponding input.

Search strategy for spelling correction

With spelling correction, words in the language database fall into three categories, depending on the length of the input:

너무 짧은 단어들

Too short words

완성들이 될 수 있는 긴 단어들

Long words that can be completions

맞춤법 수정을 위하여 적용가능한 단어들(입력 길이와의 특정 길이 차이)

Words applicable for spelling correction (specific length difference from input length)

Each of these categories is described in the following sections.

Too short words

These can easily be skipped by checking the interval stream corresponding to the last letter in the shortest allowed word. For example, if the minimum length is 6, the sixth interval stream must not be empty (with termination zero); If empty, it is possible to jump directly to the end of the interval.

Long words

Another stream can be used to check for long words, just as a special interval stream can be used to check for words that are too short. For example, if the maximum length is 12, the thirteenth stream determines whether the word is long or not.

Long words can be handled exactly the same way as if spell correction was turned off. Streams mapped to the input can be used for jumping and the completion part is obtained from the rest of the streams.

Spelling correction words

Unlike the previous two categories, which can be searched efficiently, all words belonging to this category must be sent by default to calculate the editing distance. This is a performance that is not feasible, although screening is required at the language database search level. As long as it provides a performance gain, this screening can be rejected significantly below.

A complex factor is that spell correction modes and filters can operate in the correct mode but the input is still set based, so non-spell correction candidates may be set propensity matches but spell correction cannot use set propensity information. As a result, any screening process must adhere to the set propensity comparison logic as well.

An aspect of a language database navy blue screening function for a preferred embodiment is shown in FIG. 13. With set propensity comparison logic, the target word does not match the input sequence because the 4GHI key does not contain a 'd' in its set. However, set-edit-distance comparison logic allows any input to be inserted, deleted or replaced. Thus, the set each represented by the input extends to a combination of sets containing adjacent keys. The number of adjacent keys included depends on constraint parameters such as the number of edits allowed.

As described in the following paragraphs, many screening functions from the fuzzy comparison function can be applied and integrated into the database search process.

Filter ONE / TWO

Filter ONE / TWO can be used for jumping. If the interval stream zero (the first letter in the word) does not match the corresponding input (first or second input, depending on the filter), a jump may occur.

If the filter settings do not match the set based comparison logic, they must be accompanied by a falling stream. The resulting jump is limited to the short of two (closest end in one of the two streams). This filter applies only to spell correction candidates.

Input based screening

Although the available edits match words, it looks more different than the input, and there are still restrictions on what can be matched. The limited number of edits available means that only a limited number can be applied if inserted and deleted, so there is a limit on how far the letters in a word are from the input related stream and still count as a match.

Screening can be applied independent of filters, but filters can be made part of the screening in an efficient manner. Screening must be very fast, so the complexity must be kept low.

To reject a word, one more miss is needed than the available number of edits. For example, for edit distance 3, four omissions must be found. If there are nine inputs and the word being compared has a length of 6, the length is compared to 9 because positions 7, 8, and 9 have 0 as the termination code, which cannot always be compared with any combination of inputs. . If the word is longer than the input, the length of the word is compared.

Length-Independent Screening

One solution for screening when word length is not predetermined is to set a second, fabricated, input that can be used for screening matching. It is produced in such a way that all positions are a combination of the surrounding original positions.

For input length 9, the combination map is similar to that shown in FIG. All 'lxx' columns are positions in the input. Each row is a position within the word being compared. For example, the fourth letter in a word can match any of the first seven inputs and cannot be calculated as the edit used. The twelfth character matches the ninth input, which is much more restrictive.

If any letter in a word does not match a combination it counts as missing and thus invokes potential editing. Words with sufficient omission can be discarded by this screening.

If the word is shorter than the input, the difference can be subtracted from the available edits and the comparison only requires checking of the available positions. Thus, if the length difference is equal to the number of edits available, then only one position should be made to not reject the word.

The same restrictions apply here as applied for the filter. If there is correct / in-region significance, rejection must be achieved by failure of the set based interval stream.

The longest jump possible is for the closest end of the falling interval stream, whether combination or set based.

Since there is a requirement for a falling set based stream to exist to make a jump, there is no need to limit the jump with respect to changes in the word length category.

Length-dependent screening

In a preferred embodiment of length dependent screening, the calculation of the length of the words being compared can define the combinations applicable for that length. For example, for length 6 and input length 9, the combination map is similar to FIG.

This features more limited combinations, but adds the cost of finding the word length for selecting combinations. It also limits the possible jump lengths up to the chunks of words having the same length, as soon as the length is changed, the combinations are also changed. Thus, this is also a requirement to minimize the number of word length changes across the language database.

Apart from having length dependent patterns, the description of independent screening applies here as well.

Choice List Placement Strategies and Algorithms

The combined algorithms result in a sequence that probably contains a word that the user has already typed when the input sequence is completed, or one that the user has started typing if the input sequence represents a stem of words or phrases. List of word choices for selection.

The word list sort order may be based on factors in probability, edit distance, word extinction / frequency (as stored in each database), word length, and / or stem edit distance in the region. The word list layout may also depend on which of the two or more different list profiles or strategies is used. For example:

Pre-word priority

1. The whole word always comes before word completion;

2. Source dictionaries, eg, user defined, up-to-date, plug-in, macro substitutions related to the main vocabulary, context;

3. Edit distance, eg, small value before large value;

4. Stem edit distance, eg smallest first; And only if the editing distance is greater than zero and equal for two word selections;

5. frequency, for example, the largest first; Tap frequency x word frequency

The order of evaluation is as above, for example criterion 3 is considered only when criterion 2 is the same for the compared items. For this reason, for example, spelling corrections on custom user words may appear before modifications in the area for standard lexical words.

Word completions

1. Stem edit distance;

2. Word Completion

3. source;

4. Edit distance

5. Frequency.

Since the stem edit distance is the first standard and the completion is the second, the list is effectively segmented as follows:

The whole word with zero errors, the exact tapping sequence is the same as the word

Completion (s) with 0 error stem (s)

All word (s) with one near miss

Completion (s) with one near-miss stem

...

The system may allow a specified default strategy. It can also automatically apply placement based on recognized patterns of word selection, in addition to frequency / extension information recorded in the source database. For example, the system can sense most of those times when the first letter selects word completion that exactly matches the input, thus shifting the word list placement propensity towards the "promote completion" profile.

FIG. 16 shows a sample user interface during operation of the present invention, in this case representing a set-editing-distance spelling correction with automatic correction in the area. In this embodiment on the mobile device, the candidate words each appear across the bottom of the screen on the user input. The string on the left, shown in italics, is the exact sequence of tab letters, each key pressed on its QWERTY thumb keyboard for this device. Arrows indicate default (highest ranking) word selection. The second screen shows the three word completions provided after the keys "b" and "o" are pressed. The third screen is closer to the input sequence "bok" if the letter "w" is inserted in the middle (standard edit-distance of 1) and "i" is adjacent to "k" on the keyboard (using autocorrection in the area). "Bowl" is indicated as a candidate to match. The fifth screen shows "going" as the default word choice because "g" and "i" are adjacent to the inputs of "b" and "k"respectively; The second word selection is represented as "being", with "e" replaced by "o" (edit-distance of 1). The correction parameters penalize auto correction differences in the area less than the edit-distance differences.

Other Features and Applications

Auto-replacement, eg, macros: Although word completion can be applied to expanded text, both in region and spelling corrections can be applied to shortcut keys. Thus, if the input sequence approximately matches the shortcut and stem of the expanded text, the ranking of the macro can be increased. Macros can be predefined or user-definable.

For advertising purposes, keyboard flagging benefits from automatic substitutions and / or spelling corrections. For example, if a word in a mobile message is a slang or misspelled text, embodiments of the present invention may find a keyword of a useful sponsor.

One embodiment of the present invention provides a full message buffer, i.e., batch mode, whether its text was originally entered ambiguously or explicitly through multi-tabs or received as a message or file from another device. mode).

Spelled word selection can be the basis for further input, word completion, etc., if the input method allows for automatic expansion of word selection, including rules built around punctuation. In one embodiment, the cascading menu submits a list of word completions for the selected word or stem.

Embodiments of the invention can also be applied to ambiguous entries for search and discovery. For example, if the user's input sequence does not closely match by the mobile device's content or server-based search engines, one or more modified interpretations of the spelling that cause the match may be provided.

While the above embodiments illustrate the use of embodiments of the invention with Latin based languages, other embodiments may address specific needs of other alphabets or characters.

Application to track input

Introduction( introduction )

As an improvement to the above description, hardware, software, firmware, circuitry, and other features will be configured here using a "tracking" technique. With the tracking technique, the user tracks a single, continuous path through (or nearly through) the desired letters in the input words, and the prediction technique of the system 200 indicates which words are entered and displayed on the screen 203. Calculate If there are multiple words that can be predicted from the tracked path, the system 200 provides a list of choices.

Tracking technology is also described in the publications such as the following patents filed by Nuance Corporation or its subsidiaries: (1) United States, registered February 13, 2007, under the name "FAST TYPING SYSTEM AND METHOD" by Levi. Patent 7,175,438; (2) U.S. Patent No. 7,251,367, filed Jul. 31, 2007 by Zhai under the name "SYSTEM AMD METHOD FOR RECOGNIZING WORD PATTERNS BASED ON A VIRTUAL KEYBOARD LAYOUT"; (3) US Patent No. 7,487,461, registered on Feb. 3, 2009, entitled “SYSTEM AND METHOD FOR ISSUING COMMANDS BASED ON PEN MOTIONS ON A GRAPHICAL KEYBOARD” by Zhai et al .; (4) US Pat. No. 7,706,616, filed March 27, 2010, entitled “SYSTEM AND METHOD FOR RECOGNIZING WORD PATTERNS IN A VERY LARGE VOCABULARY BASED ON A VIRTUAL KEYBOARD LAYOUT” by Kristensson et al. (5) US Patent No. 2008/0270896, published on October 30, 2008 by Kristensson, entitled “SYSTEM AND METHOD FOR PREVIEW AND SELECTION OF WORDS”; (6) United States Patent 2007/0094024, published on March 26, 2007 by Kristensson et al. Entitled " SYSTEM AND METHOD FOR IMPROVING TEXT INPUT ON A SHORTHAND-ON-KEYBOARD INTERFACE. &Quot; Each of the above patents is incorporated herein by reference in its entirety.

17 is a partial screen shot of soft keyboard 1700. The traced pattern 1706 represents the entry of the user with the word "pretty", which in turn includes tracing through the letters PRETTY. The tracked pattern can also be referred to as the tracked path of tracking, without any intended limitation. In operation, the keyboard 1700 displays some or all of the traces 1706 representing the feedback. Tracking 1706 follows the user's finger or stylus. In this embodiment, the tracking disappears or fades when the user raises the finger or stylus. In one embodiment, the length of the portion of the tracked 1706 display may vary depending on the speed of movement of the user's finger or stylus. In another embodiment, the keyboard 1700 does not display the trace 1706. In the upper area 1702, the keyboard 1700 displays some potential words that match the user's tracking 1706, including PRETTY, PERRY, PETTY, and PREY.

In order to use a tracking technique with spelling correction techniques as described above, there are various modifications and additions in accordance with the embodiments summarized below.

operation sequence

24 illustrates an operation sequence for resolving a tracking input of a user. With continued reference to the embodiment of FIG. 2, the display 203 used in this embodiment includes a touch sensitive display, such that the input device 202 includes components that may appear to be integrated with the display 203. do. Of course, there may be additional input devices within the device 200, but the associated input device for the purpose of this sequence 2400 is a composite touch sensitive display. For purposes of this embodiment, the lexical module 213 includes a lexical database with at least multiple entries. The vocabulary module 213 may be referred to as a dictionary or a vocabulary.

With continued reference to the embodiment of FIG. 2, operations 2400 are executed by the central processing unit 201 in this embodiment. In operation 2402, the CPU 201 displays a soft keyboard on the touch-sensitive display 203. The soft keyboard includes one or more letters. For purposes of explanation, a QWERTY keyboard is discussed, as illustrated by keyboard 1700.

In step 2404, device 200 receives a trace of the user through the touch-sensitive display surface. The tracking includes a single continuous tracking that touches multiple displayed keys. The touched keys include a start key at which tracking begins and an end key at which tracking stops. For ease of explanation, this embodiment uses tracking of "pretty" as shown in FIG. In one embodiment, central processing unit 201 stores coordinates or other machine-readable representations of tracked paths.

Based on the tracking received in step 2404, step 2406 achieves an input sequence. The input sequence includes contacted keys as described above. The input sequence further includes various other 'secondary' keys that are not actually touched by the trace, but are close to the trace.

Step 2406 may use one or more different criteria to identify secondary keys. For example, as shown in FIG. 25, when a trace contacts a key, all keys within the defined radius of the touched key are part of the input sequence. As shown in FIG. 26, under different approaches, when a trace contacts a given key, a defined rectangular 'touch-area' at the center on the given key is considered as an auxiliary key, and thus part of the input sequence. Depending on the area of the touch area, this criterion may be useful for defining the auxiliary keys more restrictively. One embodiment of the operative touch area is twice the size of the representable soft key. In the embodiment of Fig. 26, the auxiliary keys of the contacted "G" key include neighbors T, Y, F, H, C, and B.

Another embodiment of the criteria for identifying auxiliary keys is an ellipse along the trajectory of the trace. This gives additional consideration to horizontal undertravel and overtravel errors, minimizing vertical errors in tracking. This approach thus has a preference for keys along the direction of the tracking movement.

Optionally, step 2406 can extend the input sequence to include variations of the keys in the input sequence. For example, accents, umlauts, typographical changes, and variations corresponding to foreign languages and alphabets may be considered. 23 shows a list of input sequences corresponding to trace 1706 and the probabilities of respective keys of the input sequence.

Step 2406 includes (1) a "primary" including a start key and its aids, (2) an end key and its aids, and (3) all keys for which a predetermined minimum direction change occurs and the aids of such a key. Or define a set of keys. The keys touched between these primary keys (and their assistants) are called "intervening" keys. Keys in the input sequence that are not "primary" keys are referred to as "secondary" or "optional" keys. Thus, secondary keys include intermediate keys and their assistants.

In the example of FIG. 17, the trace contacted P, O, I, U, Y, T, R, E, R, T, and Y. Here, step 2406 defines the following keys: (1) primary keys comprising a starting key, such as P with auxiliary key O, (2) intermediate keys O, I without aids due to the speed or direction of tracking. , Secondary keys comprising U, Y, T, R, (3) primary keys comprising direction-change-key E with auxiliary keys W and R, (4) no aids due to the speed or direction of tracking Secondary keys comprising intermediate keys R and T, and (5) primary keys comprising an end key Y and its assistants T and U.

After step 2406, an input sequence is defined. In step 2408, central processing unit 201 compares the input sequence from step 2406 with some or all entries from vocabulary 213, one at a time. Vocabulary entries that are considered at any time are referred to as "current" entries.

The comparison of all entries in the vocabulary 213 may be too consuming or exhaustive of processing resources, so various techniques may be used to limit the comparison of only specific vocabulary entries. For example, step 2408 may define a comparison to entries in vocabulary 213 with the highest frequency of use. Optionally, the determination in step 2408 defining such a comparison can be made in real time, for example, when there is a given workload on the central processing unit 201, or the processing of step 2408 is at a defined level. When it arrives, it continues for a defined period of time, or performs a given number of matrix operations.

In the illustrated embodiment, step 2408 performs a comparison between the input sequence and the current lexical entry by calculating the set-edit-distance as described above. The result is a metric indicating the degree of similarity between the input sequence and the current lexical entry. The set-edit-distance is calculated in a similar manner to FIG. 4, where each matrix column takes into account multiple letters that can be intended by the user's action. However, in the present embodiment applicable to the tracking input, different matrix columns represent different keys grouped together with: a start key and one column for its assistants, each of which indicates a significant change in direction of the tracking One column for the key (and assistants), one column for each group of intermediate keys and their assistants, and one column for the end key and its assistants.

In sequence 2400, the calculation of the set-edit-distance is also distinguished from the non-tracking embodiments of the set-edit-distance by the application of a number of rules 2420. Rule 2421 defines a number of groups of primary keys. Each group appears on a different column in matrix 1800. At least one key from each group must be considered within the set-edit-distance or whether a penalty exists. In other words, a penalty is estimated for each group not represented in the current lexical entry. These groups include (1) both the start key and its assistants, as shown by 1802 in matrix 1800, and (2) both the end key and its assistants, as shown by 1808, and (3) As illustrated by 1805, includes all keys for which a predetermined minimum change in direction occurs and aids to such keys. Thus, if the current vocabulary misses both the start key and their aids, there is a penalty, and if the current vocabulary entry misses both the end key and its aids, another penalty exists.

Rule 2422 relates to secondary keys. For these keys, rule 2422 specifies that no penalty exists if the current lexical entry omits any or all of these keys. In set-edit-distance calculations, this rule allows free deletion for each of the secondary keys. Secondary keys are shown by columns 1804 and 1806 in matrix 1800.

Rule 2323 presents a rule regarding repeated keys 2423. For example, rule 2423 specifies that there is no penalty if the current lexical entry uses any key in the input sequence two or more times in succession. This allows double-leeters, because it is difficult for users to represent repeating letters using tracking. Thus, in set-edit-distance calculation, this rule allows the addition of free. Beyond tracking technology, these rules can be implemented to address various non-tracking user inputs, such as keyboards, 12-key keypads, and the like. For example, in the case of a 12-key keypad, pressing the "3" key once can yield the word "FED" because the "3" key represents the letters "3EFD".

Rule 2425 relates to punctuation, numbers, and symbols. This rule implies that set-edit-distance calculation 2408 does not incur any penalty if the current vocabulary entry includes a defined group (such as non-alphabetic characters such as punctuation, numbers, signs, and greeting / Cyrillic). Indicates no addition. These are examples of letters that are more difficult to type. As a variation of this embodiment, rule 2425 specifies that there is no penalty for using the letters of a defined group unless such a group is individually shown at the time the trace is made. For example, if the soft keyboard does not include a display of a palette of punctuation characters at the time the trace is made, there is no penalty if the current lexical entry contains one or more of these punctuation characters without an input sequence. Do not. Thus, in set-edit-distance calculation, this rule allows the addition of free.

In addition to rule 2423, rule 2425 has a wider range of techniques than tracking techniques. For example, these rules can be used to broaden the interpretation of the essentially obscure user input in which keys are submitted through a keypad that simultaneously represents multiple letters. Thus, in the case of rule 2425, the device may provide a telephone keypad user with words that contain numbers or punctuation or specific symbols that the user did not enter. In the case of rule 2423, the device may select words having different letters of the same key that are repeated (such as "OO" instead of "O") or repeated (such as "FED" instead of "F"). It can be considered as kW. One preferred patent relating to the resolution of ambiguous 12-key input is US Pat. No. 5,818,437, published October 6, 1998, by the name of the invention "REDUCED KEYBOARD DISAMBIGUATING COMPUTER" by Grover et al. The entire contents of these patents are incorporated herein by reference.

The matrix of FIG. 20 shows the calculation of the set-edit-distance for the candidate word "pretty". In this embodiment, the calculated set-edit-distance is zero.

One possible change to the calculation of the set-edit-distance of step 2408 is the consideration of probabilities based on keyboard geometry. Primarily, the calculated set-edit-distance metric for a given candidate word may be further modified according to the probability that the user intended the letters of the candidate word. For example, if the approach of FIG. 25 was used to identify supplemental keys, each supplemental key may be associated with a probability calculated according to equation 1:

Probability = 1-distance / radius [Equation 1]

The distance here is equal to the distance between the key and the auxiliary key centered on the circle, and the radius is the circle radius.

In another embodiment of FIG. 26, each auxiliary key is associated with a probability calculated according to equation 2:

Probability = Overlap / Touch-Area [Equation 2]

Here, the overlap is an area of the target auxiliary key crossing the touch area, and the touch area is an area within the touch area.

As an improved or alternative approach, the key strike for a given key may be based at least in part on the speed of the tracked path past the given key. In other words, the key strike probability may be inversely proportional to the tracking speed. Thus, if the user's finger of the stylus moves faster through a given key, the key strike probability of that key is less.

Regardless of which approach is used to determine these probabilities, the probabilities of all keys in the candidate word are considered according to one of the various approaches. In one simplified embodiment, the probabilities of all keys in the candidate word are multiplied together, and the set-edit-distance is divided by this number. Thus, candidate words with more auxiliary keys cause an inflated set-edit-distance.

In another embodiment, the probabilities are input into the probability shadow matrix, as illustrated by FIG. 22. This is used as a secondary consideration, for example, to resolve the tie between two candidate words having the same set-edit-distance. In this embodiment, the probability calculated by the shadow matrix of FIG. 22 is 12,642,870. The probabilities input into the matrix of FIG. 22 arise from all key strike probabilities implemented as described above, as in the embodiments of FIGS. 25-26. For example, such probability calculation can be performed using tap frequency, as discussed in detail above.

In this regard, step 2408 may calculate various other shadow matrices for use in resolving relationships between candidate words having the same set-edit-distance. In one embodiment, these shadow matrices include a matrix that calculates the number of free additions needed to reach the candidate word. 19 shows an embodiment of such a shadow matrix. The matrix of FIG. 19 indicates three free additions. The shadow matrix also includes a matrix for calculating the stem-edit-distance for candidate words. 21 is an embodiment of such a shadow matrix. In this embodiment, the stem-edit-distance from FIG. 21 is zero.

In one embodiment, step 2408 calculates the set-edit-distance as illustrated by FIG. 20 and omits the consideration of one or all shadow matrices if the set-edit-distance matrix exceeds a defined limit size. Can be streamlined. This helps step 2408 complete more quickly.

After step 2408, step 2410 takes the lexical inputs for which the set-edit-distance is calculated and ranks them according to their set-edit-distance matrix. Step 2410 provides a visual input of these highest ranked candidate entries, as on display 203, according to a given standard. For example, the standard may specify the top 10 entries, the top 20 entries, the correct number of entries on the display screen, or other criteria.

Although the routine 2400 describes various "free" additions and deletions, one embodiment of the sequence still tracks the use of these free features. This data can be used for a variety of purposes, such as to provide an improved confidence measure that primarily breaks the relationship between words that do not use any free additions or deletions versus words that use such.

Specific embodiments

To further illustrate the sequence 2400, the following embodiments are given. In step 2402, the system 200 displays the soft keyboard 1700 shown in FIG. In step 2404 the system 200 is tracked by the user, shown by reference numeral 1706 in FIG. In step 2406, the system defines a user's input sequence. Here, the tracking is in direct contact with the keys P, O, I, U, T, R, E, R, T, and Y. The entire input sequence, including these keys and their assistants, includes:

(1) PO: primary keys containing auxiliary O as well as contacted start key P because it satisfies the rectangular probability approach and satisfies the preference for keys along the direction of the tracking movement.

(2) OIUYTR: secondary, intermediary keys O, I, U, Y, T, and R without aids based on the speed or direction of the trace.

(3) EWR: W and R, as well as the direction change keys E contacted, because they are aids for E.

(4) RT: Secondary, intermediary keys R and T without aids based on the speed or direction of the trace.

(5) YTU: T and U with contacted exit key Y because they are auxiliary to Y.

Step 2406 then calculates the set-edit-distance for the input sequence of step 2406. In one embodiment, step 2406 is performed (at least in part) using the matrix 1800 of FIG. 18. This matrix uses concepts similar to the matrix of FIG. 4 but with some new twists due to the use of tracking. These have been described above and will be explained further below.

The words 1810 are representative of some of the input entries compared to the current input sequences at step 2406. As described, the comparison can be made with some or all of the lexical entries. Column 1802 corresponds to PO, column 1804 corresponds to OIUYTR, column 1805 corresponds to EWR, column 1806 corresponds to RT, and column 1808 corresponds to YTU. The asterisks in columns 1804 and 1806 indicate that the keys of these columns are secondary keys.

In the present embodiment, rule 2420 is applied as follows. Rule 2321 allows the start key P or its secondary O to be considered in column 1802 and the end key Y or its secondary T or U to column 1808 and the direction change key E or its secondary W or R to column 1805. It is essential to be. Column 1804 of the secondary keys OIUYTR and column 1806 of the secondary keys RT are the secondary keys, and under rule 2422, they can be ignored without loss to the set-edit-distance. Rule 2423 specifies that any of the keys 1802, 1804, 1805, 1806, and 1808 can be repeated many times without penalty. Rule 2425 does not apply to this particular embodiment because none of the candidate entries 1810 contain punctuation or special marks in this embodiment.

FIG. 28 shows a matrix for calculating the set-edit-distance for the lexical entry 'potter' of FIG. 18. In this embodiment, the trailing word 'pretty' beats 'potter' because the set-edit-distance for 'pretty' is zero.

If a relationship existed, step 2408 may consider one or more shadow matrices for the word "potter" to resolve the relationship. In this regard, the matrix of FIG. 27 shows a shadow matrix that calculates the number of free additions needed to reach three individuals, the candidate entry "potter" in this embodiment. FIG. 29 shows the shadow matrix for the stem-edit-distance, and FIG. 30 shows the shadow matrix for the key strike probability.

Other embodiments

While the foregoing descriptions illustrate a number of illustrated embodiments, it will be apparent to those skilled in the art that various changes and modifications may be made without departing from the scope of the invention as defined by the appended claims. Accordingly, the disclosed embodiments represent the subject matter broadly contemplated by the present invention, and the scope of the present invention includes other embodiments that may be apparent to those skilled in the art, and the scope of the present invention is therefore appended to the appended claims. Limited only by them.

105: data input device
110: source
115: dictionary
200: input system
201: processor
202: input device
203: display
204: Speaker
210: memory
211: operating system
212: Fix Software
213: Vocabulary Module
Application: 214, 215, 216
220: digital data processing device
221: input / output
222 processor
224: Digital Data Storage
226, 228, 230: memory
1700: Soft Keyboard
1706: Tracking

Claims (23)

A display comprising a touch sensitive display surface;
Digital data storage including a lexical database comprising multiple lexical entries; And
Coupled to the display and the storage device,
Directing the display to indicate an arrangement of keys, each representing one or more letters,
Receiving, via the display surface, a user input comprising a single continuous trace of contacting the multiple keys in sequence, including a start key, an end key, and any intermediate key between the start key and the end key. step,
Defining an input sequence comprising any auxiliary keys including the touched keys and also including keys indicative of the proximity defined in the touched keys,
Comparing the candidate entries from the vocabulary to the vocabulary, including operations of calculating a set-edit-distance metric to calculate a matching metric between the input sequence and a candidate entry for each candidate entry. Equipped,
The set-edit-distance calculation imposes a penalty for excluding all keys from the first group including the start key and auxiliary keys for the start key from the candidate word,
The set-edit-distance calculation imposes a penalty for excluding all keys from the second group including the end key and auxiliary keys for the end key from the candidate word,
The set-edit-distance calculation imposes a penalty for excluding all keys from the third group, including all contacted keys and their aids, from which the defined minimum change in direction of the tracking occurs, at least. Impose,
The set-edit-distance calculation imposes no penalty on excluding any key of the input sequence outside all the first, second and third groups from the candidate word,
The set-edit-distance calculation adds no penalty for using any key in the input sequence two or more times in succession in the candidate word, and
Ranking the candidate words according to a criterion comprising at least the calculated set-edit-distance metric, and providing an output of at least some of the ranked candidate words. A text input device comprising a processor programmed to.
2. The text input of claim 1, wherein the set-edit-distance calculation adds no penalty for candidate words that include a defined group of letters including any of punctuation, numbers, or signs. Device.
The method of claim 1, wherein the operations comprise: calculating a keyboard-geometry-based probability in conjunction with each calculated set-edit-distance metric;
And using said probability calculated to break the relationship between similarly ranked candidate words.
4. The text input device of claim 3, wherein for each of the intermediate keys, the keyboard-geometry-based probability decreases with the speed at which the tracking avoids the intermediate keys.
4. The text input device of claim 3, wherein the keyboard-geometry-based probability of a given key changes with the radius of the key from a point on the tracking.
4. The method of claim 3, wherein the keyboard-geometry-based probability of a given key varies according to the area of the given key that intersects a defined rectangular touch area centered on the contacted key closest to the given key. Text input device.
2. The text input of claim 1, wherein said ranking operation is performed to rank candidate words according to criteria comprising said set-edit distance metric and language-model-based probabilities. Device.
2. The method of claim 1, wherein the operations further comprise tracking the additions and deletions of free and using the tracked free additions and deletions to break the relationship between similarly ranked candidate words. Characterized by a text input device.
The method of claim 1, wherein the operations further comprise calculating a stem-editing-distance and using the calculated stem-editing-distance to break a relationship between similarly ranked candidate words. Characterized by a text input device.
Guiding the display to indicate an arrangement of keys, each representing one or more letters;
Receiving, via the display surface, a user input comprising a single continuous trace of contacting the multiple keys in sequence, including a start key, an end key, and any intermediate key between the start key and the end key. step;
Defining an input sequence comprising any auxiliary keys that include the touched keys and also include keys indicative of the proximity defined in the touched keys;
Comparing the candidate entries from the vocabulary to the vocabulary, comprising operations of calculating a set-edit-distance metric to calculate a matching metric between the input sequence and a candidate entry for each candidate entry. Equipped,
The set-edit-distance calculation imposes a penalty for excluding all keys from the first group including the start key and auxiliary keys for the start key from the candidate word,
The set-edit-distance calculation imposes a penalty for excluding all keys from the second group including the end key and auxiliary keys for the end key from the candidate word,
The set-edit-distance calculation imposes a penalty for excluding all keys from the third group, including all contacted keys and their aids, from which the defined minimum change in direction of the tracking occurs, at least. Impose,
The set-edit-distance calculation imposes no penalty on excluding any key of the input sequence outside all the first, second and third groups from the candidate word,
The set-edit-distance calculation adds no penalty for using any key in the input sequence two or more times in succession in the candidate word; And
Ranking the candidate words according to a criterion including at least the calculated set-edit-distance metric, and providing an output of at least a portion of the ranked candidate words.
At least one of the operations is executed by a processor.
11. The method of claim 10, wherein the set-edit-distance calculation adds no penalty for candidate words that include a defined group of letters including any of punctuation, numbers, or symbols.
11. The method of claim 10, wherein the operations comprise: calculating a keyboard-geometry-based probability in conjunction with each calculated set-edit-distance metric;
Using the probability calculated to break the relationship between similarly ranked candidate words.
13. The method of claim 12, wherein for each of the relay keys, the keyboard-geometry-based probability decreases with the speed at which the tracking avoids the relay key.
13. The method of claim 12, wherein the keyboard-geometry-based probability of a given key changes with the radius of the key from a point on the tracking.
13. The method of claim 12, wherein the keyboard-geometry-based probability of a given key varies according to the area of the given key that intersects a defined rectangular touch area centered on the contacted key closest to the given key. How to.
11. The method of claim 10, wherein said ranking operation is performed to rank candidate words according to a criterion comprising said set-edit distance metric and a language-model-based probability.
11. The method of claim 10, wherein the operations further comprise tracking the additions and deletions of free and using the tracked free additions and deletions to break the relationship between similarly ranked candidate words. How to feature.
11. The method of claim 10, wherein the operations further comprise calculating a stem-edit-distance and using the calculated stem-edit-distance to break the relationship between similarly ranked candidate words. How to feature.
An article of manufacture comprising at least one medium of digital data storage comprising a non-transitory storage of a program executable by a processor to carry out the operations of claim 10.
Receiving a user input specifying a path that is continuously tracked across the keyboard represented on the touch-sensitive display;
Resolving an input sequence of keys tracked by a defined criterion and auxiliary keys proximate the tracked keys;
For each of one or more candidate entries of a defined vocabulary, calculating a set-edit-distance metric between the input sequence and the candidate entry,
The set-edit-distance calculation imposes a penalty for excluding all keys from the first group including the start key of the route and auxiliary keys for the start key from the candidate entry,
The set-edit-distance calculation imposes a penalty for excluding all keys from the second group including the end key of the route and auxiliary keys for the end key from the candidate entry,
The set-edit-distance calculation imposes a penalty for excluding all keys from the third group, including all contacted keys and their aids, from which the specified minimum change in direction of the path occurs, at least. Impose,
The set-edit-distance calculation imposes no penalty on excluding any key of the input sequence outside all the first, second and third groups from the candidate entry,
The set-edit-distance calculation adds no penalty for the candidate entry using any key in the input sequence two or more times in succession; And
Ranking the candidate words according to the calculated set-edit-distance metrics, and providing an output of some or all of the ranked candidate words,
At least one of the operations is executed by a processor.
An apparatus comprising a touch sensitive display coupled to a processor programmed to perform the operations of claim 20.
An article of manufacture comprising at least one medium of digital data storage comprising a non-transitory storage of a program executable by a processor to carry out the operations of claim 20.
display;
A user input device comprising a keypad of multiple keys, or a touch sensitive display surface;
Digital data storage including a lexical database comprising multiple lexical entries; And
Coupled to the display, the user input, and the storage device,
A sequence of keys comprising, via the user input device, one of the keys adjacent to and in contact with successive traces entered through a touch-sensitive display surface, or one of the pressed keypad keys in which one or more jointly represent multiple letters. Receiving user input,
Identifying different candidate words, each representing a combination of letters potentially represented by the sequence of keys,
Comparing the candidate words with inputs of the vocabulary to score the compared vocabulary entries according to the reproducibility of the user input; a processor programmed to perform operations comprising:
The operation of the comparing step may include punctuation, symbols, numbers, repetition of any letter represented together on one key of the user input, together on one key of the user input, which are not included in the input sequence. And prevent any vocabulary entry including any of the use of multiple letters represented from being penalized.

Images