DE102007014405B4 - Method for word recognition in character sequences - Google Patents

Abstract

A method for word recognition of an input sequence of characters in the detection of text or speech in a detection language, wherein one or more characters of the input sequence are ambiguous, wherein
a) each character of the input sequence is assigned at least one character of the detection language,
b) for a speech sample of strings of the detection language n-grams of a plurality of certain different length n-gram lengths and the frequencies of these n-grams are determined in the speech sample,
c) assigning the input sequence of the characters, taking into account their ambiguities, the plurality of corresponding unique character sequences of the detection language,
d) wherein each of the associated unique character sequences of the detection speech is decomposed into n-grams of the plurality of certain different length n-gram lengths and the product Π of the probabilities of the decomposed n-grams Wn / NW (n) from the frequency Wn of the individual n -Grams in the speech sample and the total number NW (n) of the respective n-grams in the speech sample for the corresponding lengths n is determined,
e) where unique character sequences associated therewith ...

Description

The The invention relates to a method for word recognition in sequences of N characters, of which one or more characters are ambiguous can.

The electronic recording of texts and speech is now routine become. However, it is not bug free, takes too long, requires too much memory or is not flexible enough in terms of the used devices or languages.

The electronic recording of texts and language is used, for example. For example, when entering text in keyboards. The most efficient and common is the normal computer keyboard, which provides a key or defines a key combination for each character to be entered or for each character to be entered. Other keyboards have fewer keys, such as mobile phone keyboards for sending text messages or PDAs for appointment input, special keyboards such as QWERTY keyboards, keyboards for the disabled or special purpose keyboards. When entering text into such keyboards, it is necessary to multiple assignments of the keys, so that the keys usually have to be pressed several times to activate the desired letter (multitap method). In the case of a conventional mobile phone z. For example, there is a common key for ABC Ä 2, and when the word "Baumhaus" is entered, for example, the key sequence 222886442887777 must be pressed (in numbers), but it will then return a unique word. Assuming a common German edition alphabet without case
O (2) = {a, ä, b, c}
O (3) = {d, e, f}
O (4) = {g, h, i}
O (5) = {j, k, l}
O (6) = {m, n, o, ö}
O (7) = {p, q, r, s, ß}
O (8) = {t, u, u, v}
O (9) = {w, x, y, z}
arise for the word "tree house" (number sequence 22864287) 4 · 4 · 4 · 4 · 3 · 4 · 4 · 5 = 61,440 possible letter combinations, if for each letter of the entered word each key is pressed only once. These letter combinations go from "aatmgatw" to "baumhaus" to "ccvöicvß". Among them, impossible combinations such as "ääüöiäüq" (also numbered 22864287) must be excluded and a meaningful list of possible words (hypotheses) offered, eg. For example, for the number sequence 343 "the", "marriage", "eid", the former possibility occurring most frequently. This is the task of word recognition.

at the voice input there are also ambiguities, for example in difficult sounds like "s" and "f", whose Formant frequencies mostly outside of 3.4 kHz (upper limit of the telephone transmission frequency). Also, there may be different spellings for the same phoneme give, d. H. when it is written, as is spoken, fall Assignment and selection decisions to (/ f / → f, v, ph, / a / → a, aa, Ah). There are already phoneme dictionaries for Use came.

Also for example, when reading electronic text files, such as when contain this typed text or language or it is digital Documents can act virtually similar problems detecting these files surrender until the words disbanded are.

The following description of the state of the art and the invention refers to the focus The text recognition when typing in keyboards for which the Invention first was provided. However, the invention is not limited to this field of application limited, but also on the above and other areas of the Text recognition and analysis or speech recognition applicable.

For word recognition, various character disambiguation methods have been used to correct the ambiguities. Some are based on the use of so-called n-grams, ie of contiguous character sequences with n characters, see also Wikipdia: N-gram, 24.03.2007. n-grams have been used in the analysis of large amounts of data on specific contexts (or phrases), for example, by the intelligence service, such as the search of emails on selected topics etc .. They are also used for sentence recognition due to predetermined word sequences, the n- Gramme in this Context. In the case of word recognition by means of n-grams, character sequences (also: strings) are compared with n-grams, which can have different lengths. A combination of n-grams of different lengths has proven to be useful, with the shorter n-grams providing alternative predictions and the longer n-grams providing greater uniqueness, but requiring a large memory such that n> 6 does not occur in practice. The different lengths of the n-grams take into account the frequencies of the individual letters, bigrams, trigrams and even short words. The disadvantage of using the n-gram method is that the documents are only very short.

Great attention has published the publication "Probalistic Character Disambiguation for Reduced Keyboards Using Small Text Samples "from J.L. Arnott and M.Y. Javed, AAC Augmentative and Alternative Communication, Vol. 8, pages 215 to 223 found. In the described disambiguation method n-grams are used for different lengths, but these are not really combined. It will ultimately only be one of several existing ones n-gram length looking for the largest n-gram length ever allows a decision and only this n-gram length used.

If, for example, the word "tree house" is decomposed into n-grams of lengths 2 to 5, the following n-grams Vn (baumhaus) result:
n = 2 V2 (tree house) = {ba, au, um, mh, ha, au, us}
n = 3 V3 (tree house) = {construction, aum, umh, mha, hau, aus}
n = 4 V4 (tree house) = {tree, aumh, umha, mhau, house}
n = 5 V5 (tree house) = {baumh, aumha, umhau, mhaus}

The following frequencies are counted for some selected n-grams:
W2 (au) = 2
W3 (aum) = 1
W4 (tree) = 1
W5 (tree) = 1
Of course, n-grams that were not observed have a frequency of 0, e.g. B. W3 (lqü) = 0.

All known words of a language can now be composed of n-grams. Overall, there are z. B. in German (ignoring case) 30 letters, which limits the maximum number of possible n-grams: n = 2 900 n = 3 27,000 n = 4 810000 n = 5 24,300,000

at the special application mobile phone of intelligent text systems, without the use of text input according to the multitap method, There are only a few developments introduced in the market.

The US Pat. No. 7,129,932 B1 is concerned with keyboards comprising keys with multiple occupancy, for example for PDAs etc. Described is the finding of words and the word identification in sentences. It is intended to enter the word boundaries manually, see e.g. B. Claim 1. In a computer memory words are stored in a dictionary, also the frequency of occurrence of these words in a used language model is deposited. After entering a word boundary, the character set of a word just entered is compared with the words in the lexicon and the most likely word from the lexicon is proposed, ie the word most frequently found in the selected language model and at the same time derivable from the entered character string. During the word input, the still rudimentary word can be supplemented by the most likely word containing the existing word beginning. The described word and sentence prediction is a technique that has long been in use in speech recognition systems.

Very common is the T9 method of the company Tegic Communications, Inc. for the disambiguation of input character sequences, which in the US Pat. No. 6,307,549 B1 is described. An input character sequence is compared with stored vocabulary or word modules for correspondence. If there are several alternatives, these are presented in a selection list and an alternative selected by the user. For the vocabulary or word modules, a tree data structure can be provided. Each node of this structure represents a particular input character sequence, from which branch off several alternative sequences or possible ASCII character strings which are defined on the basis of a fixed key assignment. Where there is no path between nodes, there is no valid character sequence. For the extension of the Voka bulars is an extra dictionary provided. The disambiguation principle is therefore also referred to as dictionary-based disambiguation.

The US 2002/0183100 A1 describes a method for character selection when entering, for example, SMS messages. In the display, a character which statistically has the highest probability as a secondary character is respectively displayed as a sequence character depending on the preceding input, ie the already entered character sequence. The already entered string is fixed and will not be varied. The following character is calculated using a statistical database. Dictionaries are used for character selection, of which one contains word beginnings and words with up to three characters, the other words with four characters or more. Methods of this type are referred to as prefix-based disambiguation.

Also uses a statistical approach for the subsequent letter in the EP 0 924 594 A2 described method in which a two-dimensional table based on a preceding letter and a three-dimensional trigram table based on two previous letters is used.

Also in the disambiguation method according to WO 2004/003953 A1 (eZiText method of Zi Corporation of Canada, Inc.) the first two letters of a word are clearly entered and confirmed by the user. At the beginning of the word, the frequencies of bi- and trigrams are used. The prediction is based on a user dictionary, which preferably contains whole words and their frequency.

From the EP 1 710 668 A1 is a disambiguation method is known in which a memory with words and also n-gram objects and their frequencies is used. The n-gram objects may be words or parts of words and include mono-, bi- and trigrams.

Of the Invention is based on the object, a method for word recognition to create in character sequences that are suitable for use with character ambiguities and where word recognition is fast.

These Task is according to the invention in a Method solved with the features of claim 1. Advantageous developments the method according to the invention are the subject of the dependent claims.

at the method according to the invention for word recognition of an input sequence of characters during acquisition of text or language in a capturing language, with an or multiple characters of the input sequence are ambiguous, is every character the input sequence at least one character of the detection language assigned. To a language sample of strings of the acquisition language are n-grams of several certain different length n-gram lengths and the frequencies of these n-grams in the language sample determined. The input sequence the characters are taken into consideration of their ambiguities the several corresponding unique ones Assigned character sequences of the acquisition language. Each of the assigned unique character sequences of the detection language will be in n-grams of the several certain different length n-gram lengths decomposed and the product Π the Probabilities of the decomposed n-grams Wn / NW (n) is taken from the frequency Wn of the individual n-grams in the speech sample and the total number NW (n) of the respective n-grams in the speech sample for the corresponding lengths n determined. From these assigned unique character sequences those character sequences are removed where the product Π is zero. The remaining unique character sequences of the detection language are sorted according to the respectively determined product Π.

at the method according to the invention Thus, a display can recognize the selected n-grams and / or recognized words show, with the remaining unique character sequences of the Acquisition language, d. H. the possible ones words, be displayed in a list L.

A significant advantage of the method according to the invention is that, regardless of the language and key assignment used, it solves assignment problems of character strings and sequences, resulting in meaningful word hypotheses. This is because no words but n-grams are used to perform word recognition in strings. The list of probable word hypotheses is recreated after each keystroke in entering a word, so that a continuous update of the hypotheses is achieved with the tap step. From this list, if it contains more than one word hypothesis, the user can choose his correct word, if he has already typed the word completely. The way in which the selection is realized is arbitrary. Is not the word complete yet typed, the user will continue to type new characters.

The Detection method according to the invention can be used in any languages, legal, technical areas etc. be applied by the respective vocabulary in the statistics is integrated. Also the assignment of letters or others Sign to the keys, d. H. the output alphabet or key assignments are freely selectable, without any changes or adjustments to the procedure become necessary. Already used Language samples can unchanged be, d. H. once a language sample can be created without effort on devices be transmitted with other key arrangements or assignments. The adaptation to any languages with their individual characters like the accent in the French, Hebrew, Cyrillic, Greek etc. characters can be easily inserted. The counting a complete language sample takes just a few minutes.

The Method according to the invention is capable of taking competing characters (letters due to Key press or phonemes due to voice input or digital data records) and to isolate the possible ambiguities possible words that a valid or meaningful word. It will be with each new typed or spoken letters the possible ones recognized single letters permuted, and with each added Letters can then again other ambiguities arise, which are resolved.

For word strings without spaces, when using the method according to the invention word ambiguities may exist in the word strings, with valid resolutions resulting when all resulting words are either valid whole words or valid words and at the same time have a valid word start or word end. This is illustrated by the following example, which uses the following labels:
(G) valid whole word
(W) valid word beginning, valid word end, valid words in the sense of pA, pE, pW (explained later)
(X) neither (G) nor (W), ie invalid word

Example: the weather is excellent

Dissolution tests:

• the weather is excellent (G) (W) (G) (W) → valid resolution
• the bet is outstanding (G) (W) (X) (W) → no valid resolution
• da swetter is outstanding (G) (X) (W) → no valid resolution
• that's what he wants (G) (X) (X) (G) (G) (G) (G) (W) → none valid resolution
• the weather is outstanding (G) (G) (G) (G) (G) (W) → valid resolution

at the application of the method according to the invention on the text input is usually uses a keyboard that includes keys that have multiple characters are assigned, and connected to the processor device is. When entering text of the N-character sequences accordingly applied a word recognition method, which operates according to the invention.

If the inventive method for the Voice input is used, a voice recorder used and in the voice input of the phonemes or phoneme sequences a conversion into N-character sequences, in particular of Text characters. The N-character sequences are word-coded applied, which works according to the invention.

The inventive method let yourself also advantageous when reading, for example, digitally available Apply text documents with character sequences. For this purpose, a reading unit for the Capturing the N-character sequences used and reading the N-character sequences a word recognition method applied. Through the invention, words in found the stored character sequences very quickly and reliably and be identified.

In an advantageous variant of the method according to the invention are as whole-word n-grams the words from the speech sample whose length corresponds to the n-gram length, and when the remaining words of the list L are displayed, first all the words are sorted by the whole-word probability pG = GN / NG, where GN is the whole-word n -Gram frequency and NG is the total number of all word n-grams of the speech sample. Thus, the whole-word n-grams are mostly short words that act like a lexicon for short words taking into account the frequency of occurrence, and a meaningful sorting of word hypotheses for short words of goodness (such as "the", "marriage", "oath") support. In the case of a language sample with the words "tree house", "hello", "you", "der" the bigram "you", trigram "the" and the 5-gram "hello" and as whole word n-grams result G2 (du) = 1, G3 (der) = 1, G5 (hello) = 1

Not observed n-grams have the frequency 0, z. B. G3 (lqü) = 0. In the speech sample, the total numbers NG (n) of all Whole word n-grams calculated. These result from the sum of all frequencies the whole word n-grams of the respective length.

at an advantageous embodiment the method according to the invention the n-grams, which are the beginning, are determined as word-beginning n-grams to form a word. The word-beginning probability pA = Π An / NA becomes where An is the beginning word n-gram frequency and NA is the total of all Word begin n-grams of the language sample is. When viewing the remaining possible Become words first all words sorted by the whole word probability pG = GN / NG displayed, where GN is the integer n-gram frequency and NG is the total number of all word n-grams of the speech sample and sorting according to pA · pW he follows.

In the case of the mentioned language sample with the words "tree house", "hello", "you", "der" the bigrams "ba", "ha", "you", de ", the trigrams" construction "," hal "," The ", the 4-Gramme" tree "," hall "and the 5-Gramme" baumh "as well as" hello ". Counted the following frequencies of the word initial n-grams result: A2 (ba) = 1 A2 (ha) = 1 A2 (du) = 1 A2 (en) = 1 A3 (construction) = 1 A3 (hal) = 1 A3 (the) = 1 A4 (tree) = 1 A4 (hall) = 1 A5 (tree) = 1 A5 (hello) = 1

Prefers Word-end n-grams are also used, where as word-end n-grams the n-grams are determined, which form the end of a word. The end-of-word probability pE = Π En / NE is determined, where En the word end n-gram frequency and NE is the total number of all word end n-grams of the speech sample is. When displaying the remaining possible words, all words are first sorted by the whole word probability pG = GN / NG and the word beginning n-gram probability pA = Π An / NA where GN is the integer n-gram frequency, NG is the total of all the word n-grams of the speech sample is, to the word-beginning n-gram frequency and NA is the total number of all word-beginning n-grams of the speech sample and the sorting takes place according to pA · pW · pE.

In the case of the mentioned language sample with the words "tree house", "hello", "you", "der", the bigrams "us", "lo", "you", "he", trigrams result as the word end n-grams "Off", "llo", "the", 4-gramme "house", "allo" and the 5-gramme "mhaus", "hello". Counted the following frequencies result: E2 (us) = 1 E2 (lo) = 1 E2 (du) = 1 E2 (er) = 1 E3 (off) = 1 E3 (llo) = 1 E3 (der) = 1 E4 (house) = 1 E4 (allo) = 1 E5 (mhaus) = 1 E5 (hello) = 1

Advantageous is in the inventive method for one N-character sequence with each additional character entered the words list L recreated, d. H. there is a continuous update of the Hypotheses. From the list L, if you have more than one word hypothesis contains the user can choose the correct word if he has the word already completely typed. How to make the selection from a proposal offer is, is arbitrary.

In the previous explanations, values of n = 2, n = 3, n = 4, n = 5 were used for the n-grams. These values are not rigidly fixed, but can be adapted according to the circumstances. When using the method according to the invention, values of n = 2, n = 3 are preferably used for the n-grams, for which the memory requirement is significantly lower than in the case of longer n-grams. It can ever after application also n-grams with n = 1 (ie single letters) are used.

Prefers be for The n-grams also use values of n = 4 and / or n = 5. The longer the maximum n-grams selected (i.e., greater maximum Values for n), the better the proposed word hypotheses. It will but also more extensive language samples necessary.

in the Memory can be a list of characters or character sequences and of these associated exchange characters, exchange character sequences or exchange n-grams. This is done in this way an adaptation to the habits of a user, certain characters or words (eg "sparrow" and not "rick"), certain Short forms (English: "dont" → "do not", French: "cest" → "c'est"), Special characters (eg smiley) to use to a special vocabulary etc. The short forms must but then also entered in their short form in the language sample with become.

Also may be appropriate be added to the memory in the memory, the n-grams of new words or special entries. The input of unknown words is not necessary. Sufficient is the update corresponding n-grams (word-begin n-grams, word-n-grams, word-end-n-grams, integer-n-grams). It does not make sense, all possible frequencies the n-grams on (·), Wn (·) S (·) and Gn (·) to save (for example there is for n = 5 over 24 million possible n-grams). This is not necessary either. Only a few of these n-grams occur in the language, i. H. the frequency Most n-grams are 0. Their storage can be omitted become.

Wortend-n-grams bring the statement that it is a valid complete one Word acts, and other features may be included in the sense of Speech data recognize a word as such. To the individual words too isolate, it is especially for Applications with reading files are also useful when word boundaries, in particular Word ends, in addition be entered to complete the word string in individual ones respectively words, z. B. "tree house" also in "tree house" to share.

The inventive method can also be equipped with a word prediction. This can run so that on the basis of an input N-character sequence one Word recognition for a character sequence with an assumed length of N + (1 to 1) characters carried out where l is the prediction length, d. H. the number of predicted entry steps is. After creation The list L is created from this one more list L 'containing all the words of the List L contains, these words are supplemented by characters or character sequences with the length 1 to l. Become from the list L ' all words removed, whose word probability is zero in the list L 'remained words will be sorted and the words the Lists L and L 'are displayed. This way can for not yet complete inputted word, a prediction will be made which Word or words the user has in mind when entering.

Useful in This word prediction method first searches all the words in the list L ' the whole word probability pG = GN / NG sorted, where GN the Whole-word n-gram frequency and NG is the total number of all word n-grams of the speech sample is. The sorting then takes place after the start word n-gram and the final word n-gram probability after pA * pW * pE.

In the above, the determination of the n-grams and their use have been explained. The following is a representation of the determination of the word probabilities.

The Language statistics included in the various n-gram groups is used to exclude on the one hand word hypotheses that most likely not one word of the current language, and on the other the remaining ones Hypotheses in an order according to their probable To bring correctness. In this case, w = w1w2w3 ... wN a word w of length N, composed of the letters w1w2w3 ... wN. It will the following occurrence probabilities are determined:

For a word w, the probabilities that w is a valid word in terms of n-grams of length n are calculated

From these n-gram-length distinguished word probabilities, an entire Word probability of the word w are calculated for the total of all trained n-gram lengths:

As soon as even a single one of the word-n-gram frequencies Wn (·) = 0 is, is also the word probability pW (w) = 0.

At the Example of the word "tree house" becomes this further explained.

For the determination of bigrams and assuming that there are a total of NW (2) = 100 bigrams, we obtain: p2W (Baumhaus) = W2 (ba) / NW (2) * W2 (au) / NW (2) * W2 (um) / NW (2) * W2 (mh) / NW (2) * W2 (ha) / NW (2) * W2 (au) / NW (2) * W2 (us) / NW (2) = 1/100 * 2/100 * 1/100 * 1/100 * 2/100 * 2/100 * 1 / 100

For trigrams follows: p3W (tree) = W3 (construction) / NW (3) * W3 (aum) / NW (3) * W3 (umh) / NW (3) * W3 (mha) / NW (3) * W3 (hau) / NW (3) · W3 (off) / NW (3)

From these results ultimately results pW (Baumhaus) = ... · p2W (Baumhaus) · p3W (Baumhaus) · ...

A great help in assessing whether there is a word w are the word-beginning probabilities. For if there are no words in the language sample with the word-beginning n-grams of the word w, this will most likely not be a valid word of the language. Accordingly, the probability that a word w exists in the sense of the beginning of the word is calculated

As soon as even a single one of the word-beginning n-gram frequencies An (·) = 0 is also the word beginning probability pA (w) = 0.

An example will explain this in more detail. Suppose that you want to calculate the word beginning probability for n-grams of length 2 to 5 and the word w = tree. Given the frequencies from the list of top-of-the-word bigrams illustrated above, which in this example always has the value 1. Some exemplary total numbers of word beginning n-grams are NA (2) = 12, NA (3) = 10, NA (4) = 13, NA (5) = 11. This yields pA (tree) = A2 (ba) / NA (2) * A3 (construction) / NA (3) * A4 (tree) / NA (4) * A5 (tree) / NA (5) = 1/12 * 1 / 10 · 1/13 * 1 / 11th

Another great help in assessing if there is a word w is, as mentioned, the word end probabilities. If there are no words in the language sample that end in the same string of letters as the word w, then this is probably not a word of the language. The word end probabilities can be calculated directly from the word end n-grams:

As soon as even only one of the word end n gram frequencies En (·) = 0 is also the word end probability pE (w) = 0.

One Unknown word ending does not necessarily indicate a nonsensical word hypothesis but may as well be an indication that a word is still there not completely entered.

This explains the following example: Let's say you want the word end probability for n-grams of length 2 to 5 and the word w = Baumhaus be calculated. Given the frequencies from the list of word end bigrams illustrated above, which are always 1 in this example, and some example total numbers of word end n grams are NE (2) = 22, NE (3) = 20, NE (4) = 23, NE (5) = 21.

This results pE (tree house) = E2 (us) / NE (2) · E3 (off) / NE (3) · E4 (house) / NE (4) · E5 (mhaus) / NE (5) = 1/22 · 1 / 20 · 1/23 · 1/21

The following applies to the whole-word probabilities: If the word w to be evaluated is so short that whole-word n-grams of the same length have been determined from the speech sample, the occurrence probability of w can be simply specified as pG (w) = GN (w1 ... wN) / NG (N)

An example will explain this in more detail. Suppose that the whole-word probability is calculated for the words w = der, w = du and w = lqü. Given the frequencies from the above-mentioned speech sample (always 1 or 0) and some exemplary total numbers of whole-word n-grams, let NG (2) = 33, NG (3) = 30. This yields pG (der) = G3 (der) / NG (3) = 1/30 pG (du) = G2 (der) / NG (2) = 1/33 pG (lqü) = G3 (der) / NG (3) = 0/30

It will now be based on the example of "Baumhaus" the concrete Process closer described to the variety of possible character combinations for a Word (here as already stated above: 61.440) a meaningful list to create word alternatives.

Out the list L of possible words become all words whose probabilities pW (w) = 0 or pA (w) = 0 are removed. These words are almost certainly not a correct word dar. For performance reasons is it makes sense already in building the first word list L this criterion apply and at all only words to be included in the list for the pW (w) ≠ 0 and pA (w) ≠ 0 applies.

• (a) Zuerst stehen alle Wörter mit pG(w) ≠ 0, absteigend sortiert nach pG(w). Damit bekommen Wörter Priorität, für die Ganzwort-n-Gramme vorhanden sind.
• (b) Danach kommen alle Wörter mit pE(w) ≠ 0, absteigend sortiert nach pA(w)·pW(w)·pE(w). Damit bekommen Wörter, die ein komplettes Wort repräsentieren, Priorität vor solchen, die (bisher) nur teilweise eingegeben sind.
• (c) Es folgen alle restlichen Wörter, absteigend sortiert nach pA(w)·pW(w). In Vergleich zu der vorhergehenden Wortgruppe aus (b) werden Wörter hinten angestellt, die erst teilweise eingegeben sind. Sinnvoll wäre, dass die Eingabe von 2286428 als Hypothese zwar das Teil-Wort ”baumhau” findet, aber eventuell bessere Hypothesen bevorzugt, die nach (b) bereits ein gesamtes Wort repräsentieren.

From the few remaining word hypotheses a sorted list is generated, whereby the order of the individual words results from the following three criteria:

• (a) First all words with pG (w) ≠ 0 are listed, sorted in descending order according to pG (w). This gives priority to words for which whole-word n-grams exist.
• (b) Then all words come with pE (w) ≠ 0, sorted in descending order according to pA (w) · pW (w) · pE (w). This gives words that represent a complete word priority over those that are (so far) only partially entered.
• (c) All remaining words follow, sorted in descending order by pA (w) · pW (w). Compared to the previous word group from (b), words are placed behind, which are only partly entered. It makes sense that the entry of 2286428 as a hypothesis, although the sub-word "tree" finds, but may prefer better hypotheses, which after (b) already represent an entire word.

The Invention will be described below with reference to embodiments and the drawing further explained. In the drawing show:

1 a processor device for carrying out the method according to the invention when entering text into a keyboard,

2 a processor device for carrying out the method according to the invention in voice recording,

3 a flow chart of the method according to the invention for word recognition,

4 a flow chart to supplement the n-grams in memory and

5 a flow chart for the prediction of words in already entered subwords.

1 shows a processor device including peripherals, with which the inventive method can be used in the text input. A keyboard 10 with buttons 11 , an ad 13 and a memory 15 are to a processorein direction 12 connected. The keys 11 the keyboard 10 are assigned several characters, so that the character input is not immediately uniquely identifiable strings, words, etc. arise. The memory 15 contains n-grams and frequency values associated with these n-grams, denoted by the reference numeral 16 are symbolized. The screen display 14 the ad 13 illustrates the remaining words determined using the stored n-grams and their frequency values as possible words, here the three alternative words "the", "marriage", "eid".

In 2 a processor device is shown with peripherals for the word recognition in the voice recording. A voice recorder like a microphone 20 , an ad 22 and a memory 24 are with a processor device 21 connected. In the speech input, uniquely unambiguously identifiable phonemes or derived graphemes, N-character sequences of strings or words, etc. are not immediately generated. Basically, the approach is analogous to that in text input. The memory 24 contains n-grams and frequency values associated with these n-grams, denoted by the reference numeral 25 are symbolized. The screen display 23 the ad 22 illustrates the remaining words determined using the stored n-grams and their frequency values as possible words, here the three alternative words "the", "marriage", "eid".

3 shows that the method for word recognition is essentially characterized by the following method steps. In step 101 the procedure is the current state of the input, z. A sequence of N keystrokes. From this input is in step 102 the list L of all possible word hypotheses based on the existing input ambiguities generated by permutation of all possible combinations. In the process step 103 For example, the whole-word probabilities pG, the word probabilities pW, the end-of-word probabilities pE, and the beginning-of-word probabilities pA are computed for each word hypothesis of the list L. Based on these probabilities will be in the process step 104 removes from the list L all word hypotheses whose word probabilities pW or word-beginning probabilities pA are zero and which therefore do not represent a valid word with great certainty. If a word prediction is to be performed, the query branches 112 for the production of in 5 illustrated and explained in more detail below creation of the prediction list L '. If valid whole words exist in the list L, characterized by hypotheses with non-zero integer probabilities, the query branches 105 in the process step 106 , the all valid whole words, sorted in descending order of their whole word probabilities, on the display 14 represents. step 107 removes all in the process step 106 displayed hypotheses from the list L, thus avoiding the multiple output of the same hypothesis. If valid complete words exist in the list L, characterized by hypotheses with non-zero word end probabilities, the query branches 108 in the process step 109 , the complete valid words, descending sorted by the product of their word-end-probabilities, word-beginning-probabilities and word-probabilities, to the previous edition on the display 14 appends. step 110 removes all in the process step 109 displayed hypotheses from the list L, thus avoiding the multiple output of the same hypothesis. All remaining hypotheses of the list L are in the process step 111 , descending sorted by the product of their word-beginning probabilities and word probabilities, to the previous output on the screen 14 added. If a word prediction is to be performed, the query branches 112 for the output of the list L ', added to the previous edition on the screen 14 , If the user accepts one of the displayed words, the query branches 115 to process step 116 which provides the selected word to any application and deletes the current character sequence so that upon the next input, the method of FIG 3 in step 101 begins with an empty character sequence, ie a new word.

The appropriate embodiment of the supplement of the memory 15 reproached n-grams by new words (unknown or accepted) is represented by the following process steps 4 , described. In the process step 201 is the basis of the complement determination of all n-grams Vn (w) of the memory 15 w to be integrated. If the word w has a length covered by the whole-word n-grams, the query branches 202 in the process step 203 which updates the frequency of the word-word-associated whole-word n-gram. In the process step 204 the word w is stored in the database of the word-in-n-grams in memory 15 integrated by updating the frequencies of the word-beginning n-grams of all n-grams Vn (w) representing valid word-beginning n-grams of the word w. In the process step 205 the word w is written to the database of the word n-grams in memory 15 integrated by updating the frequencies of the word n-grams of all n-grams Vn (w). In the process step 206 the word w is written into the database of the word end n-grams in memory 15 integrated by updating the frequencies of the word-end n-grams of all n-grams Vn (w) representing valid word-end n-grams of the word w.

5 describes the method steps for generating a word prediction list L ', referenced in 3 , Process step 114 , From a list L is in the process step 301 produces a new list L 'containing for each hypothesis of the list L all concatenations of this hypothesis with all permutations of the known output alphabet in the lengths of 1 to 1 characters. From the list L 'are in the process step 302 removes all hypotheses that have either a word probability of zero or a word-first probability of zero or a word-end probability of zero. The remaining hypotheses of list L 'are in the process step 303 sorted so that all hypotheses that represent a valid whole word and are sorted in descending order of the whole word probability pG follow the further hypotheses and are sorted in descending order of the product of their end-of-word probabilities, beginning of word probabilities, and word probabilities. The edition 304 of the prediction method is thus the sorted list L '.

in the The following is the process of the invention using concrete embodiments in German and Latin using a telephone keypad further explained.

For the first example, an extensive voice sample of 688,000 words has been used, containing many compound words and also books from the trivial literature. The n-gram occupancy is as follows: Table 1 Number of n-grams in the linguistic coarse n = 2 max. 900 n = 3 max. 27,000 n = 4 max. 810000 n = 5 max. 24,300,000 Word n-grams 779 (= 87%) 9,253 (= 34%) 47,853 (= 6%) 121,139 (= 0.5%) Word initial n-grams 464 (= 52%) 2,965 (= 11%) 9,303 (= 1%) 20,113 (= 0.1%) Wortend-n-grams 469 (= 52%) 2,636 (= 10%) 8,714 (= 1%) 20,535 (= 0.1%) Full Wortn-grams 125 (= 14%) 526 (= 2%) 1,614 (= 0.2%) 3,670 (= 0.01%)

At From this table one recognizes well that just by the n-grams with n = 4 and n = 5 represents a lot of knowledge is, as the bulk (95% -99.99%) all four- and five-letter combinations do not occur in the language sample.

The The word recognition procedure now reads "finished", "prom", "beer box", "foreign trade turnover", "football game" for the words "is", "the" "tree house" their digit sequences 478 327 22864287 72466 337844. 2272458772255, 2437527836, 287364263357867289, 387225577435 are recognized. To the Comparison are the results when using the T9 technology the company Tegic stated.

Table 2 shows the result, with the searched or typed word shown in bold. Behind it is the respective word hypothesis list. Table 2 input Suggestions according to the invention Suggestions after T9 478 is, eats, gst, gru, grt, grü is, eats, hsv 327 that, that, bar, fas, eas, där, far, fär, fäs that, that, faq, barrel, dcs, fas, far, fäs, fär, eas 22864287 tree house (no suggestion, unknown word) 72466 already, beautiful, schno, schmo, schmö, schnö already, well, yes, schmö 337844 ready, destig, firm, festhi ready, firm 2272458772255 prom, prom abschlussackl 2437527836 beer, beer, beer, beer, beer, beer, beer, beer, beer, beer AIDSlarten (correct word unknown!) input Suggestions according to the invention Suggestions after T9 287364263357867289 foreign trade turnover, foreign trade turnover, foreign trade turnover, foreign trade turnover (no suggestion, unknown word) 387225577435 football game football game

The Differences between the two word recognition methods lie essentially not when editing simple, common words. Lots compound words, which develop when using the method according to the invention can, with the conventional one T9 method can not be found. The calculation time is Creation of the o. G. Word hypothesis list in the unmeasurable area.

The following example refers to Latin on a 6-key keyboard with the following key mapping:

The letters j, k, w, z do not appear in Latin and are therefore omitted. The example shows how easily text input can be adapted to new languages as well as other keyboards. This results in the following output alphabet:
O (0) = {a, b, c, d}
O (1) = {e, f, g, h}
O (2) = {i, l, m, n}
O (3) = {o, p, q, r}
O (4) = {s, t, u}
O (5) = {v, x, y}

The language statistics were determined from a language sample containing only a series of longer Latin original texts (eg Caesar "Commentary Libri VII de Bello Gallico", "Commentary Libri III de Bello Civili", "Libri Incertorum Auctorum") with a total of 128,000 words duration. After counting, the following are documented by the possible n-grams: Table 3 Number of n-grams in the linguistic coarse n = 2 max. 484 n = 3 max. 10648 n = 4 max. 234256 n = 5 max. 5153632 Word n-grams 356 (= 74%) 2,696 (= 25%) 10,130 (= 4%) 19,274 (= 0.4%) Word initial n-grams 205 (= 42%) 1,177 (= 11%) 3,049 (= 1%) 5,117 (= 0.1%) Wortend-n-grams 152 (= 31%) 931 (= 9%) 3,160 (= 1%) 6,465 (= 0.1%) Whole-word n-grams 54 (= 11%) 163 (= 2%) 540 (= 0.2%) 1280 (= 0.02%)

Out Table 3 again shows that it is precisely the n-grams with n = 4 and n = 5 a lot of information is supplied, which is the language sample itself does not contain.

In the following should
Multa legas facito, perlectis negie multa.
Qualis artifex pereo!
[Cato Maior: "See that you read a lot and when you read it, much of it is ignored." And "What great artist is going away with me!"]
on the basis of their inputs 24240 21104 100243, 313210424 2112111 24240. 340224 0342115 31313! be recognized. The complete sorted list of hypotheses is given per word, the correct hypothesis in bold: Table 4 input Suggestions according to the invention 24240 multa 21104 legas, legate, negat, nefas 100243 facito, habito, gadito, gadisq 313210424 perlectis, perlectit, perlectiu 2112111 neglege 24240 multa 340224 qualis, qualit, publiu, rudimu 0342115 artifex 31313 pereo, reper, peper, pereq, perfr, perfo, pergr

Claims (16)

A method for word recognition of an input sequence of characters in the detection of text or speech in a detection language, wherein one or more characters of the input sequence are ambiguous, wherein a) each character of the input sequence is assigned at least one character of the detection language, b) to a speech sample of strings the detection language n-grams of a plurality of certain different length n-gram lengths and the frequencies of these n-grams are determined in the speech sample, c) the input sequence of the characters, taking into account their ambiguities, the plurality of corresponding unique character sequences of the detection language are assigned d) wherein each of the associated unique character sequences of the detection speech is decomposed into n-grams of the plurality of certain different-length n-gram lengths, and the product Π of the probabilities of the decomposed n-grams Wn / NW (n) from the frequency Wn of the individual n-grams in the sp and the total number NW (n) of the respective n-grams in the speech sample is determined for the corresponding lengths n, e) where those character sequences removed from these unique character sequences are removed in which the product Π is zero, and f) the remaining unique character sequences of the detection language are sorted according to the respectively determined product Π. Method according to claim 1, characterized in that that it for the text input is applied to a keyboard, the Keyboard includes keys that are assigned to multiple characters, and the keyboard is connected to the processor device, and at the text input of the N character sequences a word recognition method is applied. Method according to claim 1, characterized in that that it for the voice input is used, and wherein a voice recorder is used, and in the speech input of the phonemes or phoneme sequences converting into N-sign sequences, in particular of text characters, done, and on the N-character sequences Word recognition method is applied. Method according to claim 1, characterized in that that it for the reading of character sequences is used wherein a reading unit is used, and in reading the N-character sequences Word recognition method is applied. Method according to one of claims 1 to 4, characterized that as whole-word n-grams the words be determined from the language sample, the length of the n-gram length corresponds to, and in the display of the remaining words the List L first, all words sorted by the whole word probability pG = GN / NG displayed where GN is the integer n-gram frequency and NG is the total number of all the word n-grams of the speech sample is. Method according to one of claims 1 to 4 and 5, characterized characterized in that the n-grams are determined as word-beginning n-grams, that form the beginning of a word, the word beginning probability pA = Π An / NA where An is the word-beginning n-gram frequency and NA is the total number of all word-beginning n-grams of the speech sample is, and when displaying the remaining possible words, all words are sorted first are displayed after the whole-word probability pG = GN / NG, where GN is the integer n-gram frequency and NG is the total number of all word n-grams of the speech sample is and sorting according to pA · pW where pW = Π pnW the word probability is and pnW = Π Wn / NW, Wn is the word n-gram frequency and NW is the total number of all word n grams. Method according to one of claims 1 to 4 and 5 or 6, characterized characterized in that the n-grams are determined as word-end n-grams, that form the end of a word, the word end probability pE = Π En / NE where En is the word end n-gram frequency and NE is the total number all the word end n-grams of the speech sample is, and at the display the remaining possible words first all words sorted by the whole word probability pG = GN / NG and the Word beginning n-gram probability pA = Π An / NA are displayed, where GN the whole-word n-gram frequency, NG is the total number of all word n-grams of the speech sample, At the word beginning n-gram frequency and NA is the total number of all word-beginning n-grams of the speech sample is, with the sorting after pA · pW · pE. Method according to one of claims 1 to 4 and one of claims 5 to 7, characterized in that from the considered character sequence of N characters, a list L of the remaining unique character sequences the capturing language is created and using an N-character sequence with each additional character entered the word list L is recreated. Method according to one of claims 1 to 8, characterized that for the n-grams values of n = 2, n = 3 are used. Method according to claim 9, characterized in that that for the n-grams values of n = 4 and / or n = 5 are used. Method according to claim 9 or 10, characterized in that for the n-grams the value n = 1 is used. Method according to one of claims 1 to 11, characterized that in memory a list of characters or character sequences and from these associated exchange characters, exchange character sequences or exchange n-grams is stored. Method according to one of claims 1 to 12, characterized that the n-grams in memory are supplemented to recognize of new words or special entries. Method according to one of claims 1 to 13, characterized that word boundaries, in particular word ends, are entered. Method according to one of claims 1 to 14, characterized that based on an input N-character sequence, a word recognition for one Character sequence with an assumed length of N + (1 to 1) characters carried out becomes, wherein from the considered character sequence of N characters one List L of the remaining unique character sequences of the detection language is created, after creating the list L from this one further list L 'created will, all the words contains list L, these words are supplemented by characters or character sequences with the length 1 to l, out the list L 'all words be removed, their word probability Zero is, the words remaining in list L 'are sorted and the words lists L and L ' become. Method according to claim 15, characterized in that that in the list L 'first all words sorted according to the whole word probability pG = GN / NG, where GN is the integer n-gram frequency and NG is the total number of all word n-grams of the speech sample is, and the sorting then after the start word n-gram and the End-word n-gram probability after pA * pW * pE.