DE19857070A1 - Determining orthographic text reproduction involves determining constituent hypothesis sequence by comparing digital data with vocabulary for representing text, rule-based processing - Google Patents

Abstract

The method involves providing a detection vocabulary containing words and word parts as word constituents to be detected; determining a sequence of constituent hypotheses by comparing the digital data with the detector vocabulary for representing the text with the determined constituent hypothesis sequence; rule-based processing of the constituent hypothesis sequence to group it into individual words to determine a word sequence hypothesis; determining a constituent of the word sequence hypothesis requiring correction; and providing at least on further constituent hypothesis for the constituents requiring correction. Independent claims are also included for an arrangement for determining an orthographic reproduction of a text and for a computer program product.

Description

The present invention relates to a method and an apparatus for Determination of an orthographic representation of a text, and in particular an application of such a method for an automatic system Speech recognition (ASE).

ASE systems usually work in such a way that spoken texts by means of of a microphone can be converted into electronic signals, which then can be stored and analyzed in order to convert these speech signals into one orthographic reproduction or representation of the spoken text determine. This is usually a comparison of the speech signal or from this determined data carried out with a dictionary in which a so-called recognizer vocabulary is stored. Depending on the degree of Similarity between the digital data and the content of the Recognizer vocabulary is then used to select terms from the vocabulary that should represent the spoken text.

The problems with such an ASE system are extraordinary diverse, essentially due to the complexity, diversity and high ambiguity of the spoken language. For example, the word "work" to be a verb or a noun, which, however, is based on the spoken text is practically undetectable for orthographic Playback has consequences. Another problem is that high number of composites of the German language, because for an ASE system it is extremely difficult to distinguish between two separate words or one Compositum are available.

Conventional systems for automatic speech recognition (ASE) work word-oriented. This is the smallest visible to the user  Unit in the recognition process the inflected or derived form of a lexeme, the latter comparable to a conventional dictionary entry. So this is it even the smallest unit accessible to the ASE user, the user must have one recognized form of a lexeme either as right or wrong discard. In particular, with conventional systems there is no possibility to take advantage of the fact that a lexeme is different Word forms can include, for example, the word forms wall, walls, Walls, - possibly wall - to the Lexem wall. Conventional ASE systems are limited to recognizing the specific word forms, they are not in the Able to deal with the fact in a flexible way that a recognized derived word form may be wrong, the corresponding one not derived lexeme but correctly recognized.

Another reason for the sometimes unsatisfactory performance conventional ASE systems consist in the principle limitation to one maximum number of recognizable word forms, hereinafter referred to as word forms - Vocabulary (= WfVok) called. Even if such vocabularies are ten or even Include hundreds of thousands of word forms, as they are never one Superset of a human speaker's vocabulary. Add to that the abundance of technical terms, neologisms (new creations) especially names decisive is the inflection (grammatical regular change of the word ending) and, as a special phenomenon of German, the composition (formation of words from partial words). The composition is a linguistically productive process, is subject to constant change. So z. B. not only through composites Linking (two or more) sub-words historically emerged, but this can currently be done by the language user at any time. Such ad hoc Composites are usually immediately decodable for other language users, not however, for a word-form-oriented ASE system, since they are naturally in one Recognizer vocabulary cannot occur.  

Because of the limited vocabulary and the described principally error-prone, since stochastically oriented recognition exists for the ASE system users the constant need for word correction. (A Detection rate of 98% must be rated as optimal.) This takes place at the Common ASE systems consistently word-form-oriented: for everyone Recognized word form will give the user a list of statistical if desired determined most likely alternatives, from which he presents the correct one Can choose shape. If this is not in the list (perhaps because it is not part of the vocabulary), it must be entered using the keyboard, even if it in a commonly available background vocabulary of the ASE system is included. Such a background vocabulary is provided in the event of a correction conventional ASE systems only the phonetic transcription, but only after manual entry of the orthographic form by the user.

The actual speech recognition conventionally takes place through probabilistic procedures, both as to the order of the sounds or smaller Units within a spoken word (i.e. comparing them determined sequence with ideally typical pronunciations of the word forms of the recognizer vocabulary), as well as the word sequences (i.e. the comparison of word candidate sequences with word sequences from text corpora analyzed offline). Because of the very high number Combinatorial bi- and trigrams of word forms can also be used in very large corpora these are not all observed. The result is Impossibility of irregular word sequences from exclude in advance, which means that basically one more or less extensive correction of a recognized text is necessary.

It is therefore an object of the present invention to provide a method and a device for determining an orthographic reproduction of a To create text, in particular a correction of errors in the Orthographic rendering of the text in a particularly simple, flexible and versatile way is possible.  

A method and an apparatus according to the present invention are defined in the independent claims. The dependent claims define particular embodiments of the invention.

The object of the invention is essentially achieved by a Recognizer vocabulary is provided that consists of words and parts of words as Word constituents exist, further that based on this vocabulary Constituent hypothesis sequence is determined, which then using Rules is converted into a word sequence hypothesis. Eventually one will Constituents of this word sequence hypothesis in need of correction are determined and for this constituent in need of correction is at least one further constituent Hypothesis delivered.

In a preferred embodiment, the others are delivered Constituent hypotheses based on a phonetic-acoustic Similarity calculus. This is done, for example, in the vocabulary, whereby that too Background vocabulary can be included, searched for entries whose phonetic transcription similar to the constituent hypothesis. This will avoided the audio data for the delivery of further similar hypotheses the recognition session must remain stored, rather such Searches for further hypotheses can also take place "offline", d. H. based on the first recognition result as the underlying hypothesis under the application of algorithms that deliver phonetically similar results and only on the acoustic representations of the constituents by their phonetic Access transcription.

Advantageously, the delivery of further hypotheses can also under the use of inflection and / or derivation paradigms, that is possible further inflections of the first hypothesis. In this case, you can take advantage of the fact  make sure that the lexeme "house" may have been recognized correctly, however, incorrectly in the nominative instead of in the "home" dative.

In a further advantageous embodiment, the delivery further hypotheses include the segmentation of words one Word sequence hypothesis in their constituents or vice versa, so one incorrectly recognized compound to separate into its components or one incorrect word from its components that is not recognized as a compound to build.

It is particularly advantageous if different ways of delivery Another hypothesis is provided and the user - e.g. B. by a Menu control - can choose between the different options. Such a menu is advantageously structured hierarchically, so that on the first level z. B. between phonetic similarity and inflection / derivation can be selected while the menu item Flexion / Derivation e.g. B. is still divided into inflection, affigation, segmentation Segmentation stands for example for the separation or formation of composites.

Another way to deliver further hypotheses includes the Subdivision of constituent hypotheses into words and / or morphs and / or Syllables, whereby new hypotheses are provided for these subunits can. This enables, for example, focusing on individual errors recognized syllables when correcting while correctly recognized syllables (or words or Morphe) can be maintained without change. The delivery of further syllable or word or morph hypotheses can then turn up different types occur, for example based on a phonetic acoustic similarity calculation.

In a preferred embodiment, the determination of Words or constituents in need of correction by a rule-based  Evaluation of the n-grams of the constituent hypothesis sequence and / or the Word order hypothesis. This enables the identification of elements in need of correction done automatically without user intervention. In another advantageous In one embodiment, such a rule-based evaluation can be triggered by a low probability of detection, so that with a probable correctly recognized hypothesis sequence a generation of errors by the rule-based evaluation of the n-grams is prevented.

Word parts are preferably only in the recognizer vocabulary as Constituents added when given a certain minimum acoustic weight. This prevents "fragmentation" of the Recognizer vocabulary in too many constituents, which in turn Number of possible permutations increased and thus the requirements for Detection increases and thus their hit rate is reduced.

A background vocabulary is advantageous in addition to Recognizer vocabulary provided in which a variety of words with their Transcriptions is stored and used to find further hypotheses can be looked up. This is preferably this Background vocabulary - just like the recognizer vocabulary Constituent base built up in particular hypotheses for constituents to be able to deliver.

It is also advantageous if an unrecognized word or a Word form or constituent, which can be achieved by using one of the mechanisms for Delivery of further hypotheses was found in the background vocabulary (or the recognizer vocabulary) can be adopted. This will ultimately the available vocabulary is dynamically expandable and the Vocabulary limits in conventional ASE systems can significantly expanded. This newly found entry can then z. B. are constituents of a new compound.  

It is particularly advantageous if the various elements used for Correction of a hypothesis can serve, all are implemented and further a new hypothesis found by correction itself again as Serves as the basis for further correction. This allows the ultimately determined or generated correct orthographic rendering of the text become.

In a particularly preferred embodiment, the present invention in a method that the orthographic output of a given, according to the invention d. H. built on a constituent basis Language model (= the total of WfVok and word [form] follow-up statistics) working automatic speech recognition system (ASE) for continuous Dictation automatically processed so that regardless of which one is used Speech recognition engine based on linguistic word environment rules (1) irregular word sequences are corrected, (2) any not in the recognizer vocabulary, but it is probably in composites contained in an arbitrarily large background vocabulary are recognized and (3) enables the user to correct them if necessary that is - independent of existing audio data and independent of correctly or incorrectly recognized word boundaries - dynamic-approximate Offers word candidates or word part candidates that consist of an arbitrarily large, set of word forms independent of the recognizer vocabulary used or Subword forms, the dynamically variable background vocabulary, automatically to be selected.

The invention is described in detail below with reference to several Embodiments with reference to the accompanying drawings described. Show it:  

Fig. 1 shows the schematic structure of an ASE system or method according to the invention, which provides a Konstituentenhypothesenfolge, with an associated correction module;

Fig. 2 shows schematically the structure of the correction module or the course of the correction process.

An ASE system according to a first embodiment of the Invention lie on the side of the vocabulary to be recognized linguistically Based on units, here called constituents, the word forms or parts of word forms. These constituents are the primary elements of the Speech signal, they form the composites for example by assembly.

This inevitably means that the system must be able to to recognize the word limits of the spoken language without discreet dictation. While the user in ASE systems for discrete dictation, the spaces between the orthographic word forms through a pause in the flow of speech "Play" must exist in continuous systems Don't dictate like the one described here. With the possible here there is no non-stop speech in the resulting speech signal Detectable segmental correspondence for word boundaries. The Difference between compound and corresponding phrase, such as B. Tax return and tax return can be purely orthographic as a result Convention and thus regarded as irrelevant in terms of speech. The components of the Composites, the constituents, are the primary elements of the speech signal, on which are based on the word sequence statistics of the basic ASE.

There are preferably two prerequisites for this: First is to prepare the underlying language model so that it is based on the Constituents rather than traditional word forms. The other must previously the constituents of the word forms to be recognized meaningfully determined  become. A complete automatic disassembly of all composites is not Expedient, as senseless decompositions (e.g. consumers) are definitely excluded have to be what even when using morphologically based Grammars cannot be provided (see, for example, T. Pachunke et al., "Broad Coverage Automatic Morphological Segmentation of German Words ", Proceedings of the Fifteenth International Conference on Computational Linguistics, Nantes, France, July 1992, Vol. IV, pp. 1218-122). Would inadequate word forms disassembling would be a falsification of the body and thus the Word [form] follow statistics the result.

Constituents should be word-based, i.e. H. they have to be noisy are represented by at least one syllable and at least one parent Contain morph, which has a minor or main accent. Depending on the recognizer, each constituent should have a certain "acoustic weight" to come d. H. a certain expected amount of energy in the speech signal. In addition, the constituents should form syntagmas in the respective context can be interpreted. For example, the pre-wall disassembly is for The pretext is correct both orthographically and in terms of sound, but it leads to the partial words in front and wall to a falsified word [form] sequence statistics, because these do not underlie the pretext. Consider the common case of hypotactically constructed from basic word and part of determination Determinative composites of the type house wall or also rear wall, so is clearly that the surroundings of the basic word -wand be syntactically comparable with the wall, but fundamentally different from that of the word pretext.

The constituent boundary is preferably always morph and Syllable boundary; thus the word segments found are the smallest unit, up of both morphological and phonetic-syllable-oriented mechanisms can operate.  

After identifying 'decay-critical' composites, i.e. H. of those Composites for which a breakdown into their constituents seems sensible occur a corresponding vocabulary adjustment, d. H. a replacement of composites of Form ab through the word sequence 'a-b' in the corpora for the word [form] sequence statistics.

It becomes a lexicon with the principles described above uses constituent-segmented word forms, which are used for Vocabulary generation covers necessary corpora. Because of the let constituents used form almost any number of other composites, from of which a certain proportion are actually used words results for the Users actually increase the vocabulary compared to one word form-oriented ASE system. This integrated into the ASE system Word forms lexicon is designed to be expandable, both with composites New formations from already saved constituents, as well as new ones Constituents. These are included in the constituent vocabulary available. For the user, these are two different recording processes not recognizable as different, it operates via a corresponding front end only the dynamic word form lexicon, which is internal on a constituent basis is constructed.

Based on this environment, the ASE system initially delivers one first hypothesis for a text to be recognized. This hypothesis is Usually in need of correction, the correction is then made by a complex from control apparatus and program modules, which continues the first hypothesis processed.

The actual complex of control apparatus and program modules, which postprocesses the ASE's continent output, hereinafter referred to as "postprocessing module", is connected downstream of the basic recognition as shown in FIG. 1:

The individual components of the "postprocessing module" are shown schematically in FIG. 2:

The input of the "postprocessing module" is a sequence of n recognized Word constituents including those from the ASE engine Probability values provided next hypotheses. In any one Repeatable sequence of levels is this input (in the following only "Constituents") modified and with information from external sources enriched. A potential but not necessary source is the Word correction by the user.

In a first (analysis) stage, syllable segmentation takes place Constituents. The result is a weighted analysis of the phonetic Constituent structure. The next step is a morphological analysis of the Constituents added, on the one hand the actual composition (Composition of constituents to composites) and the other for later steps information about further inflected or afflicted or derives derived forms as well as the part of speech.

In a syntactic level, word order rules are replaced by partial Parsing applied to the n-gram, where u. a. on the previously won Part of speech information is accessed. These rules can be triggered by "Low probability zones", i. H. they are only processed when the the sequence of constituents determined by the ASE engine is unlikely or theirs Probability is below a certain threshold. In a integrative-phonetic level are optimized by means of an engine-dependent Similarity calculus in turn formed hypotheses about the constituents.

The correction module provides the user with the results of the word, syllable and morph segmentation, so that he can carry out the following actions via suitable graphical interfaces:

• - Any combination or separation of constituents
• - Case sensitivity
• - Flexion / Derivation / Prefecture

In the interaction of these operations with a suitable one Word morphing can be performed by the user to present the hypotheses: he selects a word form or constituent hypothesis, which is then passed on to the The place of what was originally recognized moves and maintains depending on it Selection of newly generated, specific hypotheses from which, in turn, the selects the most suitable. He can repeat this process until the desired one Word is reached. This is a keyboard-free, for example, also by voice Correction is easy to use.

For those skilled in the art, there are more easily Modifications of the described embodiment. For example, the individual correction mechanisms in different orders become. Furthermore, the method described cannot only be used in an ASE System are applied, but also for example in a system for automatic spell checking and correction.

The described embodiment can be different Implement the wise, for example, the modules described by means of programs running on a computer and thus purely by software can be realized, but also a hybrid realization is partly by means of Some of the software is easily imaginable for the specialist using hardware.

For application to a speech recognition system are implicit of course, different hardware components, such as a microphone and corresponding components for converting speech into digital signals necessary, but the provision and implementation of these funds prepare the  Professional no problems. It is further understood that the invention in addition to its Realization as a method also by means of a corresponding device can be realized by a data medium that contains program code, which a computer for performing the method according to the invention prompted.

Claims (21)

1. A computer-running method for determining an orthographic reproduction of a text represented by a sequence of digital data, which comprises:
Providing a recognizer vocabulary which includes words and parts of words as constituent words to be recognized;
Determining a sequence of constituent hypotheses by comparing the digital data with the recognizer vocabulary to represent the text by means of the determined constituent hypothesis sequence;
Rule-based processing of the constituent hypothesis sequence for grouping the constituent hypothesis sequence into individual words in order to determine a word sequence hypothesis;
Determining a constituent of the word sequence hypothesis in need of correction; and
Provide at least one further constituent hypothesis for the constituents that are in need of correction. 2. The method according to claim 1, characterized in that determining the constituent hypothesis in need of correction comprises:
Determining a word of the word sequence hypothesis in need of correction;
Segmentation of the word in need of correction into its constituents;
Determination of one of the constituents resulting from the segmentation as constituents in need of correction. 3. The method according to claim 1 or 2, characterized in that the delivery of at least one further constituent hypothesis based on a phonetic-acoustic similarity calculus as phonetically-acoustically similar Constituent hypothesis is supplied.   4. The method according to claim 3, characterized in that the phonetic-acoustic similarity calculus based on phonetic acoustic representations of the constituents without recourse to any Audio data of a speech recognition device is performed. 5. The method according to any one of the preceding claims, characterized in that it further comprises:
Application of inflection and / or derivation paradigms for the delivery of at least one further inflection and / or derivation hypothesis for the constituent hypothesis that needs to be corrected. 6. The method according to any one of the preceding claims, further comprising:
Segmentation of at least the constituent hypothesis in need of correction into words and / or morphs and / or syllables;
Determination of a word or morph of the constituent hypothesis in need of correction;
Delivery of at least one further word and / or morph and / or syllable hypothesis for the word and / or morph and / or the syllable in need of correction of the constituent hypothesis in need of correction. 7. The method according to any one of the preceding claims, characterized in that it further comprises:
Rule-based evaluation of the n-grams of the constituent hypothesis sequence and / or the word sequence hypothesis;
Determination of words and / or constituents of the constituent hypothesis sequence and / or the word sequence hypothesis in need of correction based on the evaluation;
Delivery of at least one further constituent hypothesis and / or word hypothesis for the determined constituent requiring correction and / or the determined word requiring correction. 8. The method according to any one of the preceding claims, characterized characterized in that several more hypotheses each for the hypothesis in need of correction to be delivered. 9. The method according to any one of the preceding claims, characterized by an interface for determining a constituent in need of correction Hypothesis and / or word hypothesis by the user. 10. The method of claim 9, wherein the interface further comprises:
a choice of the following corrective actions:
Assembling or separating constituents;
Influencing case sensitivity;
Delivery of further hypotheses through inflection and / or derivation and / or preformation. 11. The method according to any one of the preceding claims, characterized characterized in that Parts of the word are then included as constituents in the recognizer vocabulary be given a minimum level of acoustic weight as expected Energy quantum comes in the speech signal. 12. The method according to any one of the preceding claims, characterized by:
Provision of a background vocabulary which can be accessed by means of a phonetic-acoustic similarity calculus for determining further constituent hypotheses or for determining word hypotheses. 13. The method according to any one of the preceding claims, characterized characterized in that the constituents of the recognizer vocabulary are components of composites.   14. The method according to any one of the preceding claims, characterized in that the constituents of the recognizer vocabulary
be represented at least by a syllable that bears a word accent and / or
contain at least one stem morph and / or
can be interpreted as syntagma-forming in the respective context. 15. The method according to any one of the preceding claims, characterized in that it further comprises:
Use of word form sequence statistics to determine corrective hypotheses of words and / or constituents, whereby in the word form sequence statistics the composites formed from constituents are replaced by the respective constituents. 16. The method according to any one of the preceding claims, characterized in that it further comprises:
Inclusion of new constituents and / or new composites in the recognizer vocabulary and / or a background vocabulary. 17. The method according to any one of the preceding claims, characterized characterized in that the sequence of digital data the audio data of a device for Speech recognition or the text data of a written text. 18. The method according to any one of the preceding claims, characterized by Implementation of the method according to one of the preceding claims based on a constituent hypothesis sequence, which is replaced by a Constituent hypothesis of a constituent hypothesis sequence by a new one Hypothesis arises.   19. Device for determining an orthographic reproduction of a text represented by a sequence of digital data, which comprises:
a recognizer vocabulary device which comprises words and parts of words as word constituents to be recognized;
a device for determining a sequence of constituent hypotheses by comparing the digital data with the recognizer vocabulary for representing the text by the determined constituent hypothesis sequence;
means for rule-based processing of the constituent hypothesis sequence for grouping the constituent hypothesis sequence into individual words in order to determine a word sequence hypothesis;
means for determining a constituent of the word sequence hypothesis in need of correction; and
a device for supplying at least one further constituent hypothesis for the constituent in need of correction. 20. The apparatus of claim 19, further comprising:
a device for performing a method according to one of claims 2 to 18. 21. Computer program product which comprises:
a computer readable or interpretable code which causes the computer to carry out a method according to any one of claims 1 to 18.