Images

Classifications

• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
  • G10L15/18—Speech classification or search using natural language modelling
  • G10L15/183—Speech classification or search using natural language modelling using context dependencies, e.g. language models
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
  • G10L15/18—Speech classification or search using natural language modelling
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/20—Natural language analysis
  • G06F40/205—Parsing
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/20—Natural language analysis
  • G06F40/279—Recognition of textual entities
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/30—Semantic analysis
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F8/00—Arrangements for software engineering
  • G06F8/30—Creation or generation of source code
  • G06F8/31—Programming languages or programming paradigms
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
  • G10L15/18—Speech classification or search using natural language modelling
  • G10L15/1822—Parsing for meaning understanding
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
  • G10L15/18—Speech classification or search using natural language modelling
  • G10L15/1815—Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04M—TELEPHONIC COMMUNICATION
  • H04M3/00—Automatic or semi-automatic exchanges
  • H04M3/42—Systems providing special services or facilities to subscribers
  • H04M3/487—Arrangements for providing information services, e.g. recorded voice services or time announcements
  • H04M3/493—Interactive information services, e.g. directory enquiries ; Arrangements therefor, e.g. interactive voice response [IVR] systems or voice portals
  • H04M3/4938—Interactive information services, e.g. directory enquiries ; Arrangements therefor, e.g. interactive voice response [IVR] systems or voice portals comprising a voice browser which renders and interprets, e.g. VoiceXML

Definitions

• the technology disclosed relates to natural language understanding (NLU) and to analysis of meaning of text and spoken phrases.
• NLU natural language understanding
• the technology disclosed relates to new programming constructs and tools and processing patterns that implement those new programming constructs.
• FIG. 8 This figure is part of a simple calculator example written in the grammar definition language known as Grammatical Framework (“GF”).
• GF Grammatical Framework
• An abstract syntax 811 provides a foundation for a concrete syntax 812 , 813 , as explained in Aarne Ranta, Grammatical Framework: Programming with Multilingual Grammars, Chapter 8 (2011).
• Obscure functions including lön 814 , lin 815 , and oper 816 are part of GF's expression of parsing input.
• Many layers of special purpose programming and linguistic expertise are required of those who program vertical applications using NLU programming approaches such as GF.
• grammar-based NLU frameworks such as those by Nuance, are built around defining a fixed set of slots that represent the expected information supplied in utterances within a vertical application and determining how each phrase in the grammar results in filling those slots. See, e.g., Grammar Developer's Guide, Nuance Speech Recognition System Version 8.5, Chapter 4 (2003).
• Use of grammar slots is consistent with the W3C standard known as Voice XML. See, e.g., Scott McGlashan et al., Voice Extensible Markup Language (VoiceXML) 3.0, section 6.10 Field Module, Table 41 (8 th Working Draft December 2010). Version 2.0 of the VoiceXML recommendation has been implemented in BeVocal's Nuance Café, with grammar slots.
• a vertical application or vertical market application is software defined by requirements for a single or narrowly defined market. It contrasts with horizontal application.
• An example provided by Wikipedia of a vertical application is software that helps doctors manage patient records, insurance billing, etc. Software like this can be purchased off-the-shelf and used as-is, or the doctor can hire a consultant to modify the software to accommodate the needs of the doctor. The software is specifically designed to be used by any doctor's office, but would not be useful to other businesses.
• NLU natural language understanding
• FIG. 1 is a high-level block diagram of a method and system that can be used to implement the technology disclosed.
• FIG. 2 is a high-level sequence diagram of actions and associated components from application development through runtime.
• FIGS. 3 and 4 provide additional detail regarding the application parser number and the phrase interpreter.
• FIG. 5 is a block diagram of the programming language constructs interpret-block and interpret-statement.
• FIG. 6 is a block diagram with more detail of an implementation of the interpret-statement.
• FIG. 7 includes FIGS. 7 A- 7 E .
• FIG. 7 A shows a table with three weighted patterns.
• FIG. 7 B explains normalized weighting of pattern 123456.
• FIGS. 7 C- 7 E score three token phrases against the same pattern 123456.
• FIG. 8 is a prior art example of programming natural language understanding using the Grammatical Framework language to describe abstract and concrete descriptions of a calculator.
• FIGS. 9 A- 9 D are excerpts from the code of a NLU vertical application that understands prefix and infix requests to a calculator and performs the requested calculations; they illustrate interpret blocks and interpret statements.
• FIGS. 10 A- 10 C are excerpts from the code of a NLU application that handles dates.
• FIG. 11 is an example of action statements that sets a returned weight.
• FIG. 12 is a sample table block that represents song titles.
• FIGS. 1 - 12 A detailed description of implementations of the technology disclosed is provided with reference to the FIGS. 1 - 12 .
• speech recognition is typically applied first to produce a sequence of words or a set of hypotheses.
• this speech recognition is referred to as a combination of acoustic recognition and language, or linguistic, recognition.
• Speech recognition output is sent to the NLU system to extract the meaning.
• the weakness of such a system is that errors in the speech recognition output can cause the NLU system to misunderstand the intended meaning.
• the technology disclosed includes an integrated approach that decodes both speech and meaning concurrently. If at any time interval, a partial hypothesis is not a prefix of a parsable phrase, it can be eliminated or pruned from a set of token recognition hypotheses for the speech recognition. For those hypotheses that are parsable prefixes, the partial score of the parse can be accumulated, influencing the more likely phrases from a meaning parser's perspective to have a higher score for speech recognition. This approach increases accuracy by eliminating meaningless phrases and by promoting more likely phrases. This approach also improves latency. Response time decreases due to the fact that both steps are executed concurrently as opposed to sequentially.
• the technology disclosed applies the run time meaning parser to new programming constructs that extend a programming language, such as a general-purpose programming language, using the interpret blocks and interpret statements.
• the interpret blocks process token lists (or phrases), produce scores for pattern matches and incomplete pattern matches, and return meanings associated with completed pattern matches.
• the interpret statements identify meaningful patterns that can be constructed from one or more words in a natural language, extended pattern tables, custom statistical language models and other interpret blocks.
• Interpret blocks also link interpret statements with action statements written in the general-purpose programming language which can be invoked by a completed pattern match to perform actions programmed using the power of the general-purpose programming language.
• the first example involves recognition of street addresses.
• This example illustrates defining patterns to process a token string that ambiguously expresses a street address.
• ambiguously expressed we mean that the same token string may match multiple patterns and may have multiple meanings, as in the following example.
• the patterns that the token string may match are expressed using interpret blocks and statements, extended pattern tables and action statements.
• the goal of this example is to be able to detect and parse street addresses and validate them by performing some logic.
• the user may or may not include the word “north” in the query, and if included, may put it in the wrong place as in “first street north” instead of “north first street”.
• interpret block that includes an interpret statement with an extended pattern table that, when the matches are returned, uses an auxiliary table to look up valid street addresses for particular streets:
• n_street US_STREET_ADDRESS_TABLE( ) ⁇ as ⁇
• the optional STREET_NUMBER( ) block captures a number in the address and returns its value as a street address number.
• This block can process input tokens that are as simple as a digit sequence, which, once defined can be easily used, or it can be extended to support more complex cases such as street numbers with dashes, fractions or letters of the alphabet.
• the modularity of this block supports incremental extension from simple to more complex cases.
• the next entity is a more detailed example of an extended pattern table US_STREET_ADDRESS_TABLE( ) that can represent all the variations of all the street addresses that the developer intends to support.
• This table is sometimes called an extended pattern table in contrast to bigram and trigram sequences in conventional statistical language models.
• An extended pattern table is used to parse the meaning of a token sequence, rather than predict the likelihood that a speaker or typist would utter or enter the token sequence.
• the table returns two values: an “id”, a unique number which points to a data structure that contains or refers to information about each street, such as a list of valid street number and the full name of the street as a string.
• the street name alone is provided here as an example of multiple return values.
• Other information such as the street's city, state and geo coordinates can also be part of the data structure that the id points to.
• N first street has the direction “north” before the street name.
• “north” can come after the street name, with a weight of 1/5, which again is normalized by the interpret pattern to have a probability of 1/(1+1/5), since sum of weights need to add up to 1.
• the “street” suffix can be skipped completely, but with a weight of 1/51. If a suffix is provided, it can be the correct suffix “street” with the weight of 1/(1.1).
• a modified regular expression style syntax can used as seen with weights prefixing speech recognition results (or text inputs) as quoted strings combined with symbols such as pipe (“
• weightings are automatically normalized and so the table expression to recognize any one of a, b, or c: (10 “a”
• this table includes “N First Street”, “1st street” and “21st avenue”, which all match the query example above. Assuming no other row of the table matches that query example, then the table entry that is pointed to by “N First Street”, returns 2 rows of the table for when “one twenty” is matched to the street number, and 1 row of the table when “one” is matched to the street number.
• the action statements in the programming language inside the interpret block can then evaluate various conditions, including the weight of the matched row, the range of street numbers in that row, the current location of the user and the distance from that location to the location of the matched row, to pick at least one result. See below discussing FIG. 11 , where action statements that adjust the weighting based on the distance between the user the matched location are discussed.
• n_number which is set to a street number, if one is present
• n_street which points to an entry in an extended pattern table of US streets.
• n_street is the head of a linked list of specific streets, each with a full street name and other properties.
• the code below determines the street with the best weight among the candidate streets. When performing this comparison, it takes the street number into account by calling an auxiliary function to check if a particular number is valid on a particular street.
• the program above can be extended to include additional logic such as the user location (see FIG. 11 discussion), skipping the city and choosing the right target based on population and location, and so on.
• the matching pattern can be made to return a result with fewer token matches by making the. “san jose”. segment optional, using square brackets as in: . [“san jose”].
• the table analysis will not return a completed phrase recognition if the token list omits the city name.
• FIG. 1 illustrates a block diagram of an example environment 100 in which the meaning parser and new programming constructs can be used for custom natural language understanding
• the environment 100 includes at least one client computing device 155 , 156 that includes a processor and at least one application running on the processor 157 .
• the environment also includes a communications network 135 that allows for communication between various components of the environment.
• developers prepare and submit application code 140 .
• Application code 140 can be stored for later processing or submitted directly to the application parser 130 .
• the application code may reference custom statistical language models 110 (SLMs) and extended pattern tables 115 . These SLMs and tables may be submitted with the application code 140 or may previously have been submitted by or made available to developers.
• SLMs statistical language models 110
• the network 135 includes the Internet.
• the network 135 also can utilize dedicated or private communication links that are not necessarily part of the Internet.
• the network 135 uses standard communication technologies, protocols, and/or inter-process communication technologies.
• the client computing devices 155 , the application parser engine 130 and the phrase interpreter engine 117 each include memory for storage of data and software applications, a processor for accessing data in executing applications, and components that facilitate communication over the network 135 .
• the computing devices 155 execute applications, such as web browsers (e.g., a web browser application 157 executing on the computing device 156 ), to allow developers to prepare and submit application code 140 and allow users to submit phrases to be interpreted by the phrase interpreter engine 117 .
• the computing devices 155 , 156 may be for example a workstation, desktop computer, laptop, a tablet computer, mobile phone, or any other type of computing device.
• the application parser engine 130 receives applications and parses them, producing a parse tree or an event stream. It produces application data structures 120 from the parsed application code 140 .
• An application data structure 120 may represent a single application 140 . Alternatively, multiple applications 140 may be used to prepare a single application data structure 120 .
• the application data structure can, for instance, be a tree, a state machine, or a network of valid tokens.
• the application data structure can be compiled or in an interpretable structure.
• the application data structure 120 can include nodes that reference the custom SLMs 110 and the extended pattern tables 115 . This data may be stored collectively or in copies on multiple computers and/or storage devices.
• the acoustic-language recognizer 128 can be a conventional acoustic or speech recognition component that outputs tokens. It can operate in one or more stages. In this application, an acoustic-language recognizer or processor 128 can potentially only include a single stage, acoustic recognition-only processing without application of separate linguistic analysis.
• the technology disclosed can be applied to coupling preliminary processing to meaning-specific patterns, even when the tokens from preliminary processing are phonemes or other tokens that are not full words. For instance, an acoustic stage can process input sound samples to produce phonemes. These phonemes can be passed to a language or linguistic stage that considers and scores sequences of phonemes. Language recognizers sometimes use diphone or triphone analysis to recognize likely sequences of phonemes. Language recognizers sometimes use statistical language models to recognize statistically likely sequences of words.
• Phrase interpreter engine 117 includes an acoustic-language recognizer 128 and the meaning parser 118 .
• the phrase interpreter engine 117 like the application parser engine 130 , is implemented using at least one hardware component.
• the engines are implemented in hardware, firmware, or software running on hardware.
• Software that is combined with hardware to carry out the actions of a phrase interpreter engine 117 can be stored on computer readable media such a rotating or non-rotating memory.
• the non-rotating memory can be volatile or non-volatile.
• computer readable media does not include a transitory electromagnetic signal that is not stored in a memory; computer readable media stores program instructions for execution.
• FIG. 2 in overview, includes an editor 210 used to generate at least one electronic record 211 that includes code with the programming constructs disclosed.
• the electronic record 211 is transmitted to a parser 221 , which may build a parse tree or emit a series of events.
• the parser output is used by an interpreter or compiler 231 to create executable pseudocode or object code.
• the runtime system 241 uses the pseudocode or object code to recognize natural language.
• FIG. 2 is a high-level sequence diagram of actions and associated components from application development through runtime.
• the components illustrated in FIG. 2 operate on computing devices that include a processor and memory coupled to the processor. While components are indicated by blocks, systems that implement the technology disclosed may include subdividing the indicated blocks into more components or combining multiple components into fewer blocks.
• computer-aided software engineering tools such as an integrated development environment (IDE)
• IDE integrated development environment
• Smart editors check syntax using edit time parsing to recognize keywords and the structures implied.
• An integrated development environment also may invoke an interpreter or compiler.
• Some IDEs also include runtime support for debugging a program. Debugging tools recognize the programming constructs disclosed. They allow a programmer to set breakpoints and monitor program execution. Accordingly, there is a range of environments that can implement the technology disclosed, from authoring through launched applications.
• a smart program editor 210 can recognize the programming constructs and keywords corresponding to the technology disclosed. It can color code or otherwise highlight the keywords. It may check the syntax of the programming constructs disclosed as a programmer types. It can create stub code for structures to be completed by the programmer. While smart editors have been used for many years, the programming constructs disclosed herein are new to natural language understanding programming, and offer new opportunities for developer support.
• Regular editors also can be used to author electronic records including program code.
• a regular editor When a regular editor is used, a pretty printer that recognizes the disclosed programming constructs can be used to format the code to make it more readable. Code formatting is often supported in CASE tools, IDEs and smart editors. Still, there are standalone code pretty printing products, not shown in FIG. 2 , that can apply the technology disclosed.
• a parser 221 receives one of more records 211 and converts them to machine recognized format.
• One machine recognized format is a parse tree.
• Another machine recognized format is a stream of events.
• An interpreter or compiler 231 uses the output of the parser.
• the interpreter typically uses the parser output directly in combination with a runtime 241 to execute or debug a program.
• An interpreter may persist an intermediate format for execution, debugging or optimized compilation.
• the intermediate format may be programming language independent.
• a compiler typically uses the output of the parser to compile object code or pseudocode.
• Object code can be optimized for particular platform.
• Pseudocode can be machine independent and can be run on a virtual machine, instead of directly on a physical machine.
• a preprocessor also may use output from the parser to expand the disclosed programming constructs into code in the programming language before its interpretation or compilation.
• a preprocessor may be integrated with a parser.
• a variety of devices 235 are indicated which may be targets for NLU development. These devices may include handheld devices, such as smart phones and tablets, mobile devices such as laptops and workstations or PC.
• NLU components can be deployed to servers 237 coupled in communication with other devices 235 .
• FIGS. 3 and 4 provide additional detail regarding the application parser number 130 and the phrase interpreter 117 .
• reference numbers are reused, such as reference 120 for the application data structure, they refer to the same component as previously described.
• the application code 140 , application parser 130 , application data structure 120 , extended pattern tables 115 , and custom statistical language models 110 are the same components as previously described.
• the application parser 130 parses application code 140 . In some implementations, it recognizes interpret blocks and interpret statements, as explained in the context of the street address example above and in the context of FIG. 9 below. Other implementations may perform these actions in different orders and/or perform different or additional actions than illustrated in FIG. 3 .
• the application parser recognizes patterns 331 in interpret statements and extended pattern tables.
• the extended pattern tables 115 may be stored separately from the application code 140 and reused in a variety of applications.
• the application parser 130 handles integration of tables 333 referred to in application code 140 with the data tables themselves, which may be stored separately 115 .
• the application parser 130 handles integration of custom statistical language models 335 referred to in the application code 140 with custom SLMs, which may be stored separately 110 .
• the custom SLMs 110 may be stored separately from the application code 140 and reused in a variety of applications.
• the application parser 130 Upon parsing the application code 140 and integrating it with the tables 115 and the custom SLMs 110 , the application parser 130 produces one or more application data structures 120 .
• the phrase interpreter 117 includes an acoustic-language recognizer 128 and meaning parser 118 , both of which are used to interpret spoken phrases.
• a meaning parser 118 can operate without an acoustic-language recognizer 128 .
• the meaning parser 118 is useful by itself and can operate in text-based environments that do not receive spoken input.
• the acoustic language recognizer 128 When processing spoken input, the acoustic language recognizer 128 produces token phrases and alternative token recognition hypotheses 437 . At each time interval during interpretation of spoken input, hundreds or even thousands of token recognition hypotheses 437 can be generated. To expand on this point, an interval such as every 10 or 100 milliseconds could be selected to test and generate token recognition hypotheses. At each interval, a thousand, three thousand, or an even larger number of hypotheses could be generated. In some implementations, enumeration of hypotheses could exhaust all combinatorial possibilities for hypotheses.
• the conventional acoustic-language recognizer scores the alternative token recognition hypotheses that it produces and selects a set of token recognition hypotheses for further processing.
• An acoustic recognizer applies an acoustic recognition score to alternative token sequences, selects the set of token sequences and sends them to a language or linguistic recognizer.
• the language recognizer applies a language model, such as a statistical language model, and scores the token sequence hypotheses. It returns the scores to the acoustic recognition stage, which combines the acoustic and language recognition scores.
• the acoustic-language recognizer 128 sends token recognition hypotheses 437 to the meaning parser 118 as token phrases, sequences, or lists 438 .
• the meaning parser 118 processes the tokens and returns values 438 .
• Complete and incomplete parses of tokens can be scored by the meaning parser 118 to return meaning recognition scores. Unrecognizable token phrases can be flagged as such in the returned values.
• Completed parses of token phrases that satisfy a recognized pattern can further return data and/or references to data structures that express the meaning of the token phrase.
• a token phrase being processed can be all or part of the token recognition hypothesis 437 .
• meaning parser 118 includes a token processor 455 , table handler 465 , an SLM handler 475 , and a scorer 485 . Some implementations may have different and/or additional modules than those shown in FIG. 4 . Moreover, the functionalities can be distributed among the modules in a different manner than described or illustrated.
• the token processor 455 receives tokens 438 in the hypotheses 437 . It processes these tokens against the application data structure 120 . As tables and statistical language models are encountered or invoked, the table handler 465 and SLM handler 475 are invoked.
• the table handler 465 handles extended patterns expressed as rows in the tables 115 . Additional details of these patterns and the processing of rows are described below in the context of FIG. 7 .
• the SLM handler 475 handles custom statistical language models 110 .
• Mixing the indication of custom SLMs into an extended pattern, whether in an interpret statement for the role of the table creates a context for invoking the SLM.
• This context favors custom SLMs over general SLMs. For instance, the subject line of an email will use different language constructs and different phrases in the body of an email. Accordingly, different custom SLMs would be used in patterns for subject lines and message text.
• a scorer 485 accumulates and normalizes a meaning recognition score during processing of a token phrase.
• the scorer 485 can generate scores for both partial and completed pattern recognition. Scoring of token sequences against patterns is the subject of FIGS. 7 B- 7 D , below.
• Meaning parser 118 further executes action statements contained within interpret statements. These action statements are discussed above in the context of the introductory example and below in the context of FIG. 9 .
• FIG. 5 is a block diagram of the programming language constructs interpret-block 511 and interpret-statement 515 .
• a program using these constructs is stored in one or more electronic records.
• the program includes one or more interpret-blocks 511 , 531 .
• the blocks include one or more variables 521 , 541 to be returned from the block and one or more interpret-statements 515 , 525 , 535 , 545 .
• Variables returned during execution of a block can be accessed by containing blocks.
• the variables are declared public, the values returned also can be accessed by subsequently invoked blocks that are not containing blocks.
• FIG. 6 is a block diagram with more detail of an implementation of the interpret-statement 615 .
• the interpret-statement includes a pattern 615 and an action 625 triggered by matching the pattern.
• the pattern is the modified regular expression of words in a target natural language and additional interpret-blocks, as previously described.
• the words in the natural language are terminal symbols and the additional interpret-blocks are non-terminal symbols. While a regular expression is a convenient and well-understood pattern formulation, other patterns also can be used. Patterns in the interpret-statements are used to match text or utterances. Multiple patterns in interpret-statements can match parts of a single input text or utterance.
• a parser may flag and regroup words or word patterns that will match multiple interpret-statements, and take advantage of this to optimize the application data structure 120 .
• the action statements include programming instructions in the extended programming language, such as a general-purpose programming language. These action statements include assigning values to the variables of the block-statement, which represent understanding of parts of the input text or utterance.
• the action statements may also modify the weight (score) of the parse, and eliminate parses using the special excludethis( ) statement.
• the excludethis( ) statement is a special statement that effectively sets the weight of a parse to 0. Since weights in the example are accumulated through multiplication, a weight of 0 should remove a partial parse from the list.
• FIG. 11 An example of weight modification and excludethis( ) is provided in FIG. 11 .
• the user can ask for nearby location, which is defined in another block called LOCATION( ). If the physical location of the returned location is not within 100 miles of the user's current location, the action statement calls the excludethis( ) statement which eliminates this parse.
• LOCATION( ) the excludethis( ) statement which eliminates this parse.
• the same approach could be extended to the location set by the user for the origin of the search. Otherwise the weight of the parse is adjusted by the value of the distance of the user from the location. This effectively gives preference to locations that are closer to the user.
• other constructs may be provided to provide similar exclusion functionality.
• FIG. 7 includes FIGS. 7 A- 7 E . It begins with a table showing three weighted patterns in the table and corresponding return references, or ids, 123456, 123457, 123458 and street names. While these extended patterns are illustrated as part of the table, similar patterns can be used to define the pattern of an interpret statement with the added capability to reference other interpret blocks, tables, and SLMs.
• FIG. 7 B explains normalized weighting of pattern 123456.
• FIGS. 7 C- 7 E score three token phrases against the same pattern 123456.
• the pattern for 123456 is expanded across three rows 723 , 725 , 727 to better show the nesting and weighting. Above each row are digits indicating a depth of nesting. These digits are above parentheses and square brackets. Immediately below each row is a summary of weighting juxtaposed with the rest of the pattern, e.g. “next row” and “above” designations. In an integrated development environment, nesting pairs might be indicated by matching colors of parentheses or brackets. Below each row are indications of weights assigned to segments of the pattern.
• Weights are indicated in two ways. Inside curly braces, alternatives are separated by the symbol
• a second type of weight is indicated by fractions without curly brackets, tracking optional pattern elements that appear in square brackets. For instance, in line 723 , the pattern element “north” is indicated as optional within square brackets.
• the weight 10 precedes the token. This indicates that it is 10 times as likely that the token “north” will precede “first” in a reference that means “N. 1st Avenue”, as it is likely to be omitted. When the token “north” appears in the token phrase, this term in the pattern is given a weight of 10/11. When the token is omitted, this term in the pattern is given a weight of 1/11.
• FIGS. 7 C- 7 E are scoring examples.
• the three rows in pattern 123456 are reproduced as rows 733 , 735 and 737 of FIG. 7 C ; rows 743 , 745 and 747 of FIG. 7 D ; and rows 753 , 755 and 757 of FIG. 7 E .
• Nesting levels appear above parentheses and brackets in the pattern. Resulting weights appear in curly braces below the pattern.
• the first half of the disjunction between lines 733 and 735 is matched. Accordingly, weights are assigned to line 733 and not to line 735 . As between the two alternatives, the total weight available is 5/6 for line 733 and 1/6 for line 735 . In curly braces, this is expressed as 1/(1+1/5). In the scored token string “north first street san jose”, the token “north” appears, so weight of 10/11 is assigned. The token “first” is mandatory for a pattern match and it appears with an implicit weight of one. One of the tokens “street”, “avenue”, or “road” appears, the first of which is 50 times as likely as an alternative, so a weight of 50/51 is assigned.
• the token string does not include the optional word “in”, which is weighted as unlikely to be used.
• the omission of this token effectively has a weight of 10/11.
• the token string also omits the state, which is equally likely to appear be omitted, so a weight of 1/2 is applied to the omission.
• the partial scores illustrated in this figure can be combined by multiplying them together.
• Each of the partial scores is between zero and one.
• the product of the scores also will be between zero and one.
• an average of the scores could be calculated, which also would be between zero and one.
• multiplicative and additive approaches can be used, for instance, multiplying together weights in a pattern segment (between two periods) and taking the average of weights across the pattern segments.
• a weight of 1/(1+1/10) is assigned.
• the token string does not include the optional word “in”, which is weighted as unlikely to be used.
• the omission of this token effectively has a weight of 10/11.
• the token string also omits the state, which is equally likely to appear be omitted, so a weight of 1/2 is applied to the omission.
• FIG. 7 E is similar to FIG. 7 C , but omitting “north” from the token list.
• the first half of the disjunction between lines 733 and 735 is matched because “north” is mandatory in the second line 735 .
• weights are assigned to line 733 and not to line 735 .
• the total weight available is 5/6 for line 733 and 1/6 for line 735 . In curly braces, this is expressed as 1/(1+1/5).
• the token “north” is omitted, so weight of 1/11 is assigned.
• the token “first” is mandatory for a pattern match and it appears, with an implicit weight of one.
• One of the tokens “street”, “avenue”, or “road” appears, the first of which is 50 times as likely as an alternative, so a weight of 50/51 is assigned. Since “street” is being matched in this case instead of the alternatives “avenue” or “road”, a weight of 1/(1+1/10) is assigned.
• the token string does not include the optional word “in”, with a weight of 10/11.
• the token string also omits the state, with a weight of 1/2 is applied to the omission.
• a sequence of input tokens can have more than one parse (or partial parse), with each parse having its own weight.
• the weights for all of the partial parses can be combined to represent the total weight (or probability) of a sequence of tokens.
• the weight of a sequence of tokens is useful and can be used, for example, as the score of the sequence of tokens in a speech recognition engine where multiple sequences of tokens are being considered.
• the total weight is the sum of all the weights, which makes statistical sense as weights are considered to be probabilities (the probability of a sequence of tokens is the sum of the probabilities of all its possible parses).
• the maximum weight can be used.
• the average of the weights can be calculated.
• FIG. 12 is a portion of a table block that represents song titles. This table may have thousands or even millions of rows but only 3 rows are shown in this example.
• the weight for each record can represent a measure of popularity, and if normalized to add up to one, can be considered the probability of the record.
• the system is capable of automatically normalizing weights to 1, as explained in the example of FIG. 7 , which saves a developer the burden of verifying correct sums of weights. If the user asks for “I just called to say I love you by stevie wonder”, then the popularity measure of 0.01 is used (in addition to the appropriate adjustment of 1/2 to skip the optional “by”). However, if the user asks for “I just called to say I love you”, then there are multiple matches in the table, and the weights can be added because the probability of the user asking for this song title should be the sum of the probabilities of all songs with this title.
• an interpret block is used with a table and two custom statistical language models to compose an email message.
• the table is a simple table of contacts.
• the SLMs are for subject and message.
• CONTACT( ) is a table pattern representing a list of valid contacts in a database that the user can send emails to.
• SUBJECT( ) represents a statistical language model
• MESSAGE also represents a statistical language model, and the developer can choose to use a different language model for each. This makes sense as the statistical properties of an email subject could be different from that of the body of an email.
• the system Given the query, “email ernie subject lunch meeting message i am running late”, the system matches “ernie” as the recipient (assuming Ernie is in the CONTACT( ) database), the subject becomes “lunch meeting” and the body of the email becomes “i am running late”. The system scores the subject line using the statistical language model represented by SUBJECT( ), and the body of the email using the statistical language model represented by MESSAGE( ). The system can do that because as it is parsing the query, it knows the state of its various parses. Now assume the following query:
• the code in FIG. 9 is an example of one implementation of the technology disclosed. This example applies natural language understanding to requests to a calculator to perform calculations.
• This code written as an extension of C++, is unmistakably more elegant than the prior art FIG. 8 from GF.
• the prior art GF code is written with multiple levels of abstraction in a special purpose programming language. Significant specialized expertise is required to even read the prior art code in FIG. 8 .
• one of ordinary skill in the art could well read the code in FIG. 9 and understand it with a general orientation, instead of a manual that is dozens or hundreds of pages long.
• the technology disclosed can be used to extend a programming language by the addition of a small number of constructs.
• the technology can serve as a core around which a programming language is designed.
• the technology disclosed is used to extend a mainstream, well-understood programming language such as C++.
• it may extend other programming languages such as Objective C, C#, Javascript, Java, Python, Ruby, or Perl. Extending a general purpose programming language takes advantage of an existing large pool of programmers who already are familiar with the language.
• the technology disclosed also could be incorporated into a special purpose programming language with benefits to those who understand or choose to learn the special purpose language.
• Interpret-statements such as 913 , 914 , 928 include both a pattern and an action to be performed (or action statements) when the pattern matches an input text or utterance.
• the patterns in this example are modified regular expressions of terminal and non-terminal symbols, with optional weights. For instance, in interpret-statement 928 of FIG. 5 b , the natural language word to match is “minus”, which is called a terminal symbol by those working in NLU.
• the actions triggered by various pattern matches are expressed in a general purpose programming language.
• the patterns can include non-terminals that invoke other blocks, pattern tables, and SLMs.
• FIG. 9 A-D are four pages excerpted from a longer working NLU vertical application for a calculator that can be expressed in nine pages of clear and readable code.
• the excerpts from this nine-page program extend a general purpose programming language with just two programming constructs and use the general purpose language to implement actions triggered by matched patterns.
• FIG. 10 the code samples provided apply the programming constructs disclosed.
• FIGS. 10 A- 10 C are four pages excerpted from code from a longer working NLU application that handles dates. Interpretation of natural language that expresses a date is more challenging than implementing a calculator, as described below, because dates can be expressed in so many ways.
• recurring dates are excluded 1013 ; today, tomorrow and the next day are handled 1014 , 1015 , 1016 ; a day of week and date of month are combined 1021 ; a day of week and ordinal number are combined 1026 ; a part of day, such as morning, is combined with an ordinal number 1031 ; a legal document date style is handled 1033 ; and a day of week is combined with “this” or “next” week 1036 .
• interpret blocks 1011 can cause several interpret blocks (.ter files) to be included in the code before the DATE( ) interpret block 1012 .
• Several interpret-statements 1014 , 1015 , 1016 , 1021 , 1026 , 1031 , 1033 , 1036 appear in block 1012 .
• the exclude-statement 1013 stops execution of the block without processing subsequent interpret-statements and without assigning values to the block variables 1012 , when the input includes phrases that indicate recurring dates, as this block does not handle recurring dates.
• exclude statements are a type of interpret statement.
• Interpret-statements 1014 , 1015 , 1016 handle the words “today”, “this”, “tomorrow” and “day after tomorrow”.
• Optional words that do not change the selected date include approximate times of day, such as “morning”, “evening”, “afternoon” etc. Some of these optional words may trigger assignment of a positive value to the variable “pm_hint”.
• Parts of the pattern in interpret-statement 1014 for instance, cause values to be assigned to n1 or n2. If values have been assigned to n1 or n2, the last line in 1014 resets pm_hint from 0 to 1.
• several qualitative block variables 1012 such as month_index_implicit, year_index_implicit, week_delay and pm_hint are given values.
• Interpret-statements 1021 , 1026 , 1031 , 1033 and 1036 include weights 1022 , 1027 , 1032 , 1034 , 1037 assigned when optional approximate times of day are part of the input. Four of the five interpret-statements use the same weight. The fifth 1033 , assigns a much smaller weight 1034 . These weights can be used by a runtime system to select among multiple interpret statements that may be triggered by a particular input pattern. In some implementations, the parser automatically normalizes weights returned by interpret-statements and parts of interpret-statements. Normalizing weights to total one another chosen value simplifies human understanding of output and debugging output when the example code runs. Normalizing weights to sum to one effectively produces a Bayesian probability distribution that can be used at runtime.
• the DATE( ) example in FIG. 10 illustrates the power of extending a general purpose programming language through the juxtaposition of patterns with actions.
• the general purpose programming language is used in this example and others to convey complex logic that could be difficult to express in a special purpose programming language.
• the immediate juxtaposition of patterns and actions in the interpret-statements makes it easy to see how language is being understood and what patterns are being matched. This contrasts with other approaches that rely on separate abstract syntax and concrete syntax or that express how to proceed depending on “slots” that have been filled.
• the action statements can inter-relate complex computations including mismatches across “slots”, e.g. invalid addresses.
• the main programming constructs in FIG. 9 are enumerations, interpret-blocks and interpret-statements.
• Enumerations 911 are programmed in the underlying general-purpose programming language, in this case in C++.
• a number of interpret-blocks are identified 912 , 917 , 921 , 926 , 941 . Many others are unlabeled.
• the interpret-block ARITH_QUERY( ) 912 for instance, includes three variable arguments response, *formula and valueStr. It also includes two interpret statements 913 , 914 .
• the interpret-block is distinguished by its name; interpret-statements are distinguished by their patterns. In other implementations, the interpret-statements also could be named.
• Interpret-statements 913 , 914 in ARITH_QUERY( ) 912 both have patterns that are non-terminal symbols, invoking other named interpret-blocks.
• Interpret statement 913 has slightly more complicated pattern, which concatenates results from two named interpret-blocks.
• the interpret-block ARITH_WHAT( ) has no variables and therefore does not return any values. This is because the sole interpret-statement 916 in the block is triggered by literals such as “tell me” and “what is”, which contribute to recognizing a natural language query, but which do not contribute to performing the desired calculation. From this example, one sees that interpret-blocks and interpret-statements can, in some cases, be configured as trivial filters that ignore parts of the input.
• the interpret-blocks ARITH_INFIX_TAIL( ) 926 and ARITH_INFIX( ) 521 combine to interpret operators in the input and execute the requested operations.
• One block includes multiple interpret statements 927 , 928 , 929 that each address a different operator, such as “plus” 927 or “minus” 928 . Enumerated operators are assigned to the variable “op”, which is one of the variables returned by ARITH_INFIX_TAIL( ).
• the other block ARITH_INFIX( ) 921 uses the value and formula returned by invoking ARITH_INFIX_TAIL( ) 926 as part of pattern in the single interpret-statement 922 contained in the block.
• This interpret statement includes a series of “if” and “else if” statements with the same effect as a case statement, testing the value of the variable “op” returned by ARITH_INFIX_TAIL( ) p26. For a “minus” operator 928 in the input stream, one of the else if clauses 923 matches the value of “op” and the value returned by ARITH_INFIX_TAIL( ) 926 is subtracted from the value returned by ARITH_PREFIX( ) 932 .
• FIG. 9 D Another example in FIG. 9 D is the interpret-block ARITH_PREFIX_UNARY( ) 941 , which handles unary operators.
• the first two interpret-statements 942 , 943 take into account parentheses around unitary operators.
• Additional interpret-statements 944 , 945 interpret various operators. Some of these interpret statements, which operate on patterns that begin with “the” could be reformulated to make “the” an optional part of a regular expression.
• FIG. 9 illustrates elegance of the interpret-block and interpret-statement programming constructs, implanted as an eminently readable extension of a general-purpose programming language.
• interpret-statements can be invoked from a single interpret block to combine a variety of applications with a common entry point. An example follows:
• each of the above commands could have been independently developed by different individuals and entities and tied together by the common entry point by yet another developer.
• each domain or vertical can be created or maintained by experts in that domain. For example, weather service providers can wok on the weather vertical, while navigation is created and maintained by other experts and so on.
• a method includes an automated method of accurately recognizing a speaker's meaning.
• This method includes producing and scoring partial transcriptions of a spoken phrase at intervals as the phrase is spoken using at least one acoustic-language recognition stage and a meaning parser stage.
• the acoustic-language stage generates a multiplicity of token recognition hypotheses that build on prior partial transcriptions and selects a set of the token recognition hypotheses using at least prior scoring of the prior partial transcriptions at earlier intervals and current acoustic-language scoring of the token recognition hypotheses at a present interval.
• the meaning parser stage concurrently processes particular token recognition hypotheses in the set of the token recognition hypotheses; determines whether a particular token recognition hypothesis has a parsable meaning; rejects unparsable hypotheses; and scores and returns at least one parsable meaning score for a particular token recognition hypothesis that has a parsable meaning.
• the acoustic-language recognition stage further stores for use at subsequent intervals combined scores of the token recognition hypotheses for current partial transcriptions, the combined scores incorporating at least the acoustic-language recognition scores and the parsable meaning scores.
• the acoustic-language recognition stage can prune the prior partial transcriptions used to build the token recognition hypotheses using the meaning parser stage rejections of the unparsable hypotheses.
• the method includes selecting at least one completed transcription of the spoken phrase that has been scored as recognizable by the acoustic-language recognition stage and as parsable by the meaning parser stage.
• the meaning parser stage can further take the actions of processing the token recognition hypothesis that includes at least one ambiguously expressed element and at least one dependent element that correlates with the ambiguously expressed element against an interpretation pattern and of processing and scoring the ambiguously expressed element against multiple rows of an extended pattern table that is invoked while processing the interpretation pattern, at least some of the rows declaratively expressing weighted alternative expressions of a particular ambiguously expressed element.
• Handling ambiguously expressed elements further include applying logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table using return values from the scored rows in combination with against at least information from a group consisting of (1) the dependent element from the token recognition hypothesis, (2) an optional element in the token recognition hypothesis, and (3) supplemental information not included in the token recognition hypothesis.
• One implementation can further include applying logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table, including comparing valid dependent values in an auxiliary table with the dependent element from the token recognition hypothesis.
• One implementation can further include applying logic expressed in a general programming language to process supplemental information not included in the token recognition hypothesis and rescore at least some of the scored rows.
• One implementation can further include applying logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table against optional elements in the token recognition hypothesis.
• One implementation can further include the meaning parser stage scoring the token recognition hypotheses against an interpretation pattern that includes at least one predicate condition and at least one statistical language model applied when the predicate condition is satisfied.
• a particular token recognition hypothesis includes an ambiguously expressed element and a dependent element that correlates with the ambiguously expressed element; the meaning parser stage further takes the actions of scoring the token recognition hypothesis against a plurality of interpretation patterns built from a model pattern, the model pattern implementing at least: (1) a table pattern that includes rows in an extended pattern table, at least some of the rows declaratively expressing weighted alternative expressions of a particular ambiguously expressed element; and (2) a statistical pattern that includes a predicate condition and a custom statistical language model applied when the predicate condition is satisfied.
• applying the table pattern includes scoring the token recognition hypothesis against multiple rows of the extended pattern table.
• applying the statistical pattern includes scoring the token recognition hypothesis against the custom statistical language model.
• Processing results of scoring the tech in recognition hypothesis against rows of the table can further include applying logic expressed in a general programming language to process and rescore at least some scored rows from the extended pattern table using at least information from a group consisting of: (1) the dependent element from the token recognition hypothesis; (2) an optional element in the token recognition hypothesis; and (3) supplemental information not included in the token recognition hypothesis.
• This processing also can include applying the logic expressed in the general programming language to process and rescore at least some of the scored rows from the extended pattern table against valid dependent values in an auxiliary table and the dependent element from the token recognition hypothesis.
• It can include applying the logic expressed in the general programming language to process and rescore at least some of the scored rows from the dependent value table against the optional element in the token recognition hypothesis.
• It can include applying the logic expressed in the general programming language to process the supplemental information not included in the token recognition hypothesis and rescore at least some of the scored rows.
• implementations may include a non-transitory computer readable storage medium storing instructions executable by a processor to perform a method as described above.
• implementations include a system including memory and one or more processors operable to execute instructions, stored in memory, to perform a method as described above.
• a method that, in some environments accurately recognizes an intended meaning of a complete phrase.
• This method includes invoking an interpretation pattern that expresses a complete phrase with a meaning as a sequence of elements that include one or more words from a natural language and a table element that invokes an extended pattern table and receiving a text string of tokens that express an intended meaning, wherein the token include a combination of at least one ambiguously expressed element and a dependent element that correlates with the ambiguously expressed element, and further including one or more supplemental elements.
• the method can further include applying logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table against at least information from a group consisting of (1) a dependent element from the text string, (2) an optional element in the text string, and (3) supplemental information not included in the text string.
• the method further includes applying logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table against valid dependent values in a dependent value table and a dependent element from the text string.
• the method can further include applying logic expressed in a general programming language to process supplemental information not included in the text string and rescore at least some of the scored rows.
• That it can further include applying logic expressed in a general programming language to process and rescore at least some of the scored rows from the dependent value table against optional elements in the text string; and selecting at least one intended meaning using at least the rescored rows of the extended pattern table.
• implementations of this an additional methods described below also may include non-transitory computer readable storage medium storing instructions executable by a processor to perform a method is described.
• implementations may include the system including memory and one or more processors operable to execute instructions, stored in the memory, to perform a method is described.
• the proviso in this paragraph is hereby applied the to the implementations in this section.
• an automated method of building a natural language understanding application includes receiving at least one electronic record containing programming code that interprets sequence of input tokens by extending a general purpose programming language with interpret-block and interpret-statement data structures.
• the interpret-block data structures include at least one of the interpret-statements and zero or more variables returned by the interpret-block.
• the interpret-statements include a pattern of one or more tokens, and zero or more action instructions.
• the action instructions perform logic not achieved by pattern matching and/or assign values to the variables of the interpret-block.
• the method further includes parsing the received program code to produce an executable representation of the interpret-block and interpret-statement data structures.
• This method mother implementations of the technology disclosed can each optionally include one or more the following features.
• At least one token of the interpret expression can be another interpret-block.
• the returned parameters from other interpret-blocks are made available to the action statements inside the interpret block.
• At least one token of the interpret expression can be a statistical language model. It also can be a wildcard. It can be a table of token expressions with fixed returned values for each row of the table and without any action statements. At least one sub-expression of the interpret expression is allowed to have repetitions. At least one of a minimum and a maximum number of repetitions of a sub-expression can be specified.
• the outgoing weights at each token can be normalized to add up to 1. The normalization of outgoing weights can be performed at sub nodes instead of tokens to reflect the way the expression is modularized.
• this method also can be practiced in a non-transitory computer readable storage medium or by a system.
• a method in another implementation, includes scoring a partial transcription of input. Practicing this method includes instantiating in memory at least one data structure derived from programming code that interprets token list using a general purpose programming language extended with interpret-block and interpret-statement data structures.
• the interpret-block data structures include at least one of the interpret-statements and one or more values returned by the interpret-block.
• the interpret-statements include patterns that are built from words in a target natural language, from at least one extended pattern table, and from references to additional interpret-blocks and action instructions in the general purpose programming language that are triggered by a match between parts of an input text and the patterns.
• the extended pattern table matches and scores at least part of the token list against multiple rows of in the extended pattern table, at least some of the rows declaratively expressing weighted alternative expressions of ambiguously expressed elements.
• the action instructions assign values to the variables of the interpret-block, which values attribute meaning to the token list.
• the method further includes receiving the token list and processing and scoring the token list against the data structure including scoring at least part of the token list against multiple rows of the extended pattern table, at least some of the rows declaratively expressing weighted alternative expressions of a particular ambiguously expressed element.
• the action instructions can further include logic expressed in the general programming language to process and rescore at least some of the scored rows from the extended pattern table using at least information from a group consisting of: (1) a dependent element in the token list that has meaning in a context set by an ambiguously expressed element in the token list; (2) an optional element in the token list; and (3) supplemental information receive in addition to the token list.
• the action instructions further can include logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table comparing valid dependent values in an auxiliary value table with the dependent element.
• the action instructions further can include logic expressed in a general programming language to process and rescore at least some of the scored rows from the extended pattern table using the optional element.
• the action instructions further include logic expressed in a general programming language to process the supplemental information not included in the token list and rescore at least some of the scored rows.
• this method also can be practiced in a non-transitory computer readable storage medium or by a system.
• method in another implementation, includes building a natural language understanding (abbreviated NLU) data structure.
• This method includes receiving at least one electronic record containing programming code that interprets an input text by extending a general purpose programming language with interpret-block and interpret-statement data structures.
• the interpret-block data structures include at least one of the interpret-statements and one or more variables returned by the interpret-block.
• the interpret-statements include patterns that are built from words in a target natural language, from at least one extended pattern table and from references to additional interpret-blocks and action instructions in the general purpose programming language that are triggered by a match between parts of an input text and the patterns.
• the extended pattern table matches and scores at least part of the input text against multiple rows of in the extended pattern table, at least some of the rows declaratively expressing weighted alternative expressions of ambiguously expressed elements.
• the action instructions assign values to the variables of the interpret-block, which values attribute meaning to the text.
• the method further includes parsing the received program code to produce a data structure representing the interpret-block and interpret-statement data structures.
• the pattern specified in the interpret-statement data structure can include a regular expression of the words and the additional interpret-blocks.
• the extended pattern table can be invoked by an antecedent event selected from a group at least consisting of: a match between part of the word hypothesis and at least one word in the natural language that is part of the pattern preceding the extended pattern table; and positioning of the extended pattern table as a first element of the pattern.
• the general purpose programming language can belong to a “C” programming language family.
• the set of the interpret-blocks collectively can define a vertical application of NLU.
• Values assigned to the variables of a particular interpret-block can be available to additional interpret-blocks and to a NLU processor at run time.
• Patterns in the interpret-statements in the set of interpret-blocks collectively can match substantially all of a vertical application vocabulary that is recognized by the vertical application of NLU.
• the method can further include receiving a plurality of sets of the interpret-blocks that define a plurality of vertical applications and parsing the plurality of sets of interpret-blocks.
• the interpret-block can further include at least one exclude-statement that contains an exclude pattern that is built from words in a target natural language and matching of the pattern in the exclude-statement causes an exit from the interpret-block without further processing of include-statements.
• the patterns of the include-statements include relative weights assignable to matches of patterns or partial patterns.
• this method also can be practiced as code stored on a non-transitory computer readable storage medium or on running a system.
• parser running on a processor that builds a representation of natural language understanding (abbreviated NLU).
• This parser includes, program instructions running on at least one processor that cause the processor to receive at least one electronic record containing programming code that interprets text or utterances by extending a general purpose programming language with interpret-block and interpret-statement data structures.
• the interpret-block data structures include at least one of the interpret-statements and one or move variables returned by the interpret-block.
• the interpret-statements include a pattern that is built from words in a target natural language or from references to additional interpret-blocks and action instructions in the general purpose programming language that are triggered by a match between parts of the text or utterances and the pattern.
• the action instructions assign values to the variables of the interpret-block, which values attribute meaning to the text or utterances.
• the parser parses the received program code to produce a parse tree that represents the interpret-block and interpret-statement data structures.
• the pattern specified in the interpret-statement data structure can include a regular expression of the words and the additional interpret-blocks.
• the general purpose programming language belongs to a “C” programming language family.
• the values assigned to the variables of a particular interpret-block are available to additional interpret-blocks and to a NLU processor runtime.
• a set of the interpret-blocks collectively can define a vertical application of NLU.
• the patterns in the interpret-statements in the set of interpret-blocks can collectively match substantially all of a vertical application vocabulary that is recognized by the vertical application of NLU.
• Operation of the parser can further include receiving a plurality of sets of the interpret-blocks that define a plurality of vertical applications and parsing the plurality of sets of interpret-blocks.
• the interpret-block can further include at least one exclude-statement that contains an exclude pattern that is built from words in a target natural language and matching of the pattern in the exclude-statement causes an exit from the interpret-block without further processing of include-statements.
• the patterns of the include-statements can further include relative weights assignable to matches of patterns or partial patterns.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Computational Linguistics (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Health & Medical Sciences (AREA)
• Artificial Intelligence (AREA)
• General Engineering & Computer Science (AREA)
• Software Systems (AREA)
• General Physics & Mathematics (AREA)
• Human Computer Interaction (AREA)
• Acoustics & Sound (AREA)
• Multimedia (AREA)
• General Health & Medical Sciences (AREA)
• Computing Systems (AREA)
• Machine Translation (AREA)
• Devices For Executing Special Programs (AREA)

Abstract

Description

Claims (14)

Priority Applications (2)

Applications Claiming Priority (5)

Related Parent Applications (1)

Related Child Applications (1)

Publications (2)

Family

ID=74882543

Family Applications (4)

Family Applications Before (2)

Family Applications After (1)

Country Status (1)

Cited By (1)

Families Citing this family (26)

Citations (306)