Method for Automatically Detecting Meaning and Measuring the Univocality of Text

• US 20160357731A1
• Filed: 07/29/2014
• Published: 12/08/2016
• Est. Priority Date: 01/28/2014
• Status: Active Grant

• Current Assignee: Speech SenSz GmbH
• Original Assignee: Somol Zorzin GmbH
• Inventors: Zorzin, Luciano
• Granted Patent: US 10,303,769 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 40/232 Orthographic correction, e.g. spell checking or vowelisation
• CPC Class Codes: G06F 40/268 Morphological analysis
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/56 Natural language generation

Abstract

The invention relates to a method for automatically detecting meaning patterns in a text using a plurality of input words, in particular a text with at least one sentence, comprising a database system containing words of a language, a plurality of defined categories of meaning in order to describe the properties of the words, and meaning signals for all the words stored in the database, wherein a meaning signal is a clear numerical characterization of the meaning of the words using the categories of meaning.

Claims

1. A method for automatically detecting meaning-patterns in a text using a plurality of input words comprising a database system containing words of a language, a plurality of pre-defined categories of meaning in order to describe the properties of the words, and meaning-signals for all the words stored in the database, wherein a meaning-signal is an univocal numerical characterization of the meaning of the words using the categories of meaning, and wherein at least the following steps are carried out:

a) reading of the text with input words into a device for data entry, linked to a device for data processing,

b) comparison of all input words with the words in the database system that is connected directly and/or via remote data line to the system for data processing,

c) assignment of at least one meaning-signal to each of the input words, wherein in the case of homonyms two or more meaning-signals are assigned;

d) in the event that the assignment of the meaning-signals to the input words is univocal, the meaning-pattern identification is complete,

e) in the event that more than one meaning-signal could be assigned to an input word, the relevant meaning-signals are compared with one another in an exclusively context-controlled manner, wherein

f) on the basis of the combination of the meaning-signals of the input words among one another, it is determined whether a contradiction or a match is present in the meaning of the input word with respect to the context;

g) meaning-signal combinations that lead to contradictions are rejected, meaning-signal combinations for matches are automatically numerically evaluated in accordance with the degree of matching of their meaning-signals based on a pre-defined matching criterion and recorded,

h) automatic compilation of all input words resulting from the steps d) and g) are output as the meaning-pattern of the text.

2. The method as claimed in claim 1,

wherein,

in accordance with the pre-defined matching criterion, it is automatically decided whether the meaning-pattern for at least one input word of the text has more than one remaining meaning, so that no unique meaning-pattern and/or no unique meaning of the sentence exists in the context and a display of the non-uniqueness and its cause is provided and/or made available to a User Interaction Manager if required.

3. The method as claimed in claim 1,

wherein,

the text with the input words is a string of characters that originates from a written text and/or from any other source, including an acoustically recorded text using a speech recognition program, or photographed text, or OCR.

4. The method as claimed in claim 1,

wherein,

a signal for the degree of univocality of a text which can be further processed is generated if following step e) of the claim, the remaining number of meaning-signals for all input words of a text is known.

5. The method as claimed in claim 1,

wherein,

after a word meaning score “

SW” and

a sentence meaning score “

SS”

is calculated by a meaning modulator for all of the words of the text wherein the word meaning score is the number of entries of each word in the database system, coupled with the relevance of the meaning-pattern of each word in the context of the sentence;

a) if the meaning score “

SW”

for a word of the sentence is equal to 0 (zero), then the word is spelt incorrectly and the sentence receives the sentence score “

SS”

=0,

b) if the meaning score “

SW”

for a word of the sentence is greater than 1, then the analyzed sentence is incorrect, and/or not univocally formulated, because words with SW>

1 have more than 1 possible meaning in the sentence and its context, wherein the sentence score is then set to “

SS”

=“

SW”

,

c) if more than one word of the sentence have meaning score “

SW”

>

1, then the sentence score “

SS”

is set to the maximum value “

SW”

of the meaning scores of the words of the respective sentence,

d) if all the words of the sentence have a meaning score “

SW”

=1 then the sentence is univocal and receives the sentence score “

SS”

=1,

e) if words have a meaning score “

SW”

=−

2, then they allow both upper and lower case spelling, wherein the sentence score “

SS”

then receives the value “

SS”

=−

2, until the correct upper or lower case spelling of the words with “

SW”

=−

2, in this sentence, is finally calculated using further iterative steps,

f) if the text originates from speech input and if words have a meaning score “

SW”

not equal to 1 and belong to a homophone group—

identified from the data processing system—

then they receive the meaning score “

SW”

=−

3, and the sentence score “

SS”

retains the value −

3 until the correct homophone of the group in this sentence and its context is finally calculated using further iterative steps,

g) if words of the sentence have meaning score “

SW”

>

1, then with words of an arbitrary number “

v”

of preceding or of “

n”

following sentences of the text it is checked whether words are included here which due to the modulation of their meaning-signals lead to “

SW”

=1 in the input sentence, wherein for normal speech applications and easily understandable texts, usually “

v”

=1 and “

n”

=0.

6. The method as claimed in claim 5,

wherein,

with words with SW=0, a storable error message is generated, which in particular indicates spelling errors of all the words of the text and in particular the calculated possibilities for eliminating the error, and is stored sequentially in an error-message-storage and is available to a User Interaction Manager when required.

7. The method as claimed in claim 4,

wherein,

with in words with “

SW”

=−

2, a storable error message is launched, which in particular indicates the presence of case errors in the spelling of all the words of the sentence, naming the word position in the sentence, the cause of the error and displaying possibilities for eliminating the error calculated from the storage of the database system, and is stored sequentially in the error-message-storage and is available to a User Interaction Manager when required.

8. The method as claimed in claim 1,

wherein,

in the event that no words have SW=0, a meaning modulator updates on a rolling basis the main theme—

as the most frequent, valid constraint reference from in the form of its meaning-signal—

of the current paragraph in the form of the meaning-signals of the constraint references and is made hierarchically retrievable and available to a User Interaction Manager when required.

9. The method as claimed in claim 1,

wherein,

in the case of sentences with SS>

1 an autotranslation message is generated, which lists the still existing #SW meaning possibilities of each word and in each case retrieves the most common synonyms of each word from the database system using its valid meaning-signals and stores them sequentially in the autotranslation storage and makes them available to a User Interaction Manager when required.

10. The method as claimed in claim 1,

wherein

it is part of a computer-implemented translation device for the translation of texts, in particular sentences of a natural language into a target language, by using “

meaning-checking”

, wherein a sentence with score SS=1 is automatically acquired, or the text is processed until at least one sentence with sentence score=1 exists, and/or there are no unprocessed sentences left with SS unequal to 1.

11. The method as claimed in claim 10,

wherein

the text is translated into the selected target language, taking into account the pre-defined, univocal meaning-signals of all words and all additional information that are available in the storages and Interaction Manager.

12. The method as claimed in claim 10,

further comprising

an application of language-pair-specific rules of the database system, which by adjustment of the order of the words in the input sentence in relation to their morphology and inflection, and of the order of the sentence constituents, determines main clauses, dependent clauses, inserted dependent clauses, subjects, predicates, objects, text parts between hyphens, text parts between two brackets (open/closed), and places the sentence in memory in the target language in an order that is at least as semantically, morphologically, grammatically and syntactically as correct in the target language as in the input sentence, taking into account all sentence-related entries in memories.

13. The method as claimed in claim 1,

wherein

the resulting words of the translation are displayed and/or acoustically reproduced, or represented on an output medium so that they are perceptible by other sensory organs.

14. The method as claimed in claim 1,

wherein

in the presence of words with homophones in a sentence and appropriate specification, a review of the degree of meaning-signal correspondence of the present word and all its other homophonous spellings from database system in relation to the context is performed automatically, whereupon an automatic replacement by the homophone with the highest meaning modulation in the sentence takes place and/or an error message is output via error-message-storage and Interaction Manager if there is insufficient computational differentiation among the meaning-signals of the words of an identical homophone group in the context.

15. The method as claimed in claim 1,

wherein

for processing and/or reconstructing garbled texts from automatic speech recognition of a natural language in the presence of background noise and/or text with typing errors, OCR, and subject to the condition that for at least one word SS=0, the possibilities for reformulating the sentence are automatically and systematically determined, by correctly spelling incorrect words, firstly with the priority on words that are similar to homophones of the relevant word, or that correspond to omissions of letters, spaces or typical typing errors when operating a keyboard, including upper/lower case, and accenting.

16. The method as claimed in claim 15,

wherein

the meaning-signals of correctable words are used to investigate whether sentences with a sentence score SS=1 are produced which the user then receives as prioritized output, and/or the procedure is terminated if no usable hits can be identified after a user-specified time, wherein the input sentence is then tagged with the information of the words that were analyzed for correction, and if only sentences with a score unequal to 1 exist, those having the fewest words with SW=0 are prioritized for the tagging, wherein the overall result obtained is made available to a User Interaction Manager via error-message-storage and autotranslation storage.

17. The method as claimed in claim 1,

wherein

for a search engine for searching in databases, the textual content of which are tagged by “

meaning-checking” and

can be queried automatically based on the automatic tagging.

18. The method as claimed in claim 17,

wherein

an automatic database updating takes place in accordance with the meaning-signals of all of its words before the search process.

19. The method as claimed in claim 1,

wherein,

an automatic inclusion of all same-language synonyms and all foreign-language synonyms in all their valid inflections is included in the search (same meaning-signal as the search term).

20. The method as claimed in claim 1,

wherein,

when using multiple search words, a combination of the meaning-signal hits as claimed in the association logic of the search words is carried out.

21. The method as claimed in claim 1,

wherein,

it performs a computer-implemented evaluation of the relevance of statements in the form of text in natural language to a topic specified in writing, by, in the case of an automatically acquired sentence with sentence score SS=1, the meaning-signals of the words of the sentence with pre-defined combinations or patterns of meaning-signals being automatically compared with tagged words of the comparison topic.

22. The method as claimed in claim 21,

wherein,

the overlap of the meaning-signals of the topic specification and the input sentence with pre-defined meaning modulation patterns is ranked, taking into account the existence of meaning-signals of logical operators and/or disjunctors and/or other sentential connectors within the sentence structure of the input sentence.

23. The method as claimed in claim 1,

further comprising

a computer-implemented conduct of automatic dialogs by computers and/or “

responding computers”

with users, so that the spoken input of a user is acquired as text by the responding computer and processed with “

meaning-checking”

.

24. The method as claimed in claim 23,

wherein,

a breakdown of the input text into individual sentences is carried out by the responding computer, and an automatic evaluation is made as to which of these sentences are statement sentences, question sentences, or exclamation sentences.

25. The method as claimed in claim 1,

wherein

the meaning-signals of the statement and question sentences of the user are compared based on their matching/correspondence with a tagged database of the statement sentences, response sentences and standard question sentences of a machine-readable text ontology of the responding/dialog-participating computer, which exists in the same natural language—

but not necessarily—

as the natural language in which the user interacts, wherein at least one of the following steps is carried out;

(a) in the case of matching values of the meaning-signals of the input sentences of the user above a certain level, with the computer ontology of the responding computer, the response and statement sentences rated the highest in the matching/correspondence value are identified from the computer ontology being used,

(b) the responding computer generates a structured, automatic response for the user by confirmation of the highest ranking sentences of the user in relation to the computer ontology by the responding computer via a speech output system in accordance with the state of art and/or other sensorially detectable transmission procedure,

(c) offering the highest ranking response sentence of the computer ontology of the responding computer via a speech output system in accordance with the prior art and/or other sensorially detectable transmission procedure which only allows the user to make controlled answers on request,

(d) sending of a link and/or sensorially detectable information by the responding computer—

as claimed in certain rules of the ontology and appropriate to the meaning of the user'"'"'s questions—

which the user receives, in order to retrieve and/or read more detailed information on his questions and then to be able to put more targeted questions to the responding computer that the user himself might otherwise only have found in the computer ontology which is readable for him after some search effort of his own,

(e) in the case of matching values of the meaning-signals below a certain matching level, a standard dialog based on its previous questions is called up in the responding computer, to which the user can only answer “

Yes”

or “

No”

, and/or by uttering controlled pre-defined, in particular spoken, alphanumeric, audible, sensible or visually perceptible options, and/or that an automatic detection is carried out in the responding computer of the moment from which the intervention of a human being is needed, by automatic evaluation of the redundancy of the dialog or of content-based patterns such as anger or impatience, of meaning-signal patterns in the verbal responses of the user during the dialog and/or visually perceivable responses of the user via a camera in the immediate environment of his data input device.

26. The method as claimed in claim 1,

further comprising

a computer-implemented, enhanced spell-checking, by using “

meaning-checking”

, wherein in particular an automatic execution takes place but without the sentence itself being tagged with the meaning-signals after having reached a sentence score>

0, equivalent to the fact that the text is only checked for spelling errors and corrected interactively by the user, but without necessarily any tagging of the sentence with e.g. semantic or logical additional information taking place.

27. The method as claimed in claim 1,

further comprising

a computer-implemented word recognition during typing of words on keyboards by using “

meaning-checking” and

automatic completion of the words with words from the database system which best match the syntax and context existing at this point in time.

28. A computer-implemented method for the semantic encryption of sentences of a natural language, using “

meaning-checking”

as claimed in claim 1,

wherein,

“

m”

words in each sentence are replaced in a grammatically/semantically well-formed manner, and/or “

n”

words are added in a grammatically/semantically well-formed manner, which have suitable meaning-signals compared to their immediate, contextual environment, which indicate that by insertion, negation, relativization or omission and/or by use of antonyms thereof from the database of the database system the sentence meaning can be changed significantly, but without the sentence score being changed, equivalent to the fact that, after the automatic modification, the text contains no additional semantically/factually less meaningful sentences than the original from which it is produced, with “

m”

>

=1 or “

n”

>

=0, and wherein at least one of the following steps is carried out;

a) all alphanumeric chains which are proper names and/or dates and/or pure numbers which have their own meaning-signals, or to which automatically matching meaning-signals can be automatically assigned, and/or single words marked in advance are each replaced by coded, anonymized keywords, to which shortened meaning-signals, appropriate to the degree of anonymization, are automatically added,

b) the user'"'"'s starting sentences are stored on the user'"'"'s system taking account of the original order, and a log file is stored of all changes that were created as sentence variants or anonymizations, wherein each change and derivable content of the change and the position in the respective sentence of the text are recorded,

c) the user is assisted with “

meaning-checking”

, to identify from other retrievable text databases on the system he is using than the current text itself, sentences that are semantically—

but not logically—

similar to sentences from the input text to be encrypted, and that have a sentence score SS=1,

d) the number of sentences of the original text is increased to at least 7 if over the input text plus sentence variants there are less than 7 sentences to be encrypted,

e) a text is created which contains the user'"'"'s starting sentences, plus “

m”

appended sentences which are automatically created variants of his,

f) a stochastic scrambling of the sequence of the existing sentences is carried out and the explicit modification of the sequence before and after the scrambling is appended to a log file,

g) if the unchanged, but scrambled text and the generated log files are available, the original text which the user originally entered, can be flawlessly reconstructed to match the original,

h) potential system queries of the encrypted text are tagged on the individual words and sentences in such a way that, after reconstruction of the original text autotranslation queries, error messages and semantic information of the sentences can automatically cancel each other out, so that context-related information items which due to the scrambling are initially no longer in context, are reconstructed automatically in the original text, and without user interaction if this was not required in the unscrambled text.