Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/10—Text processing
  • G06F40/12—Use of codes for handling textual entities
  • G06F40/151—Transformation

Definitions

• deductive program synthesis a complete high-level specification is translated into the corresponding low-level program where each step of translation is validated using some axioms. This approach requires users to provide a complete specification, which in some cases may potentially be harder than writing the program itself. This has caused the inductive synthesis approaches to become more popular recently.
• inductive program synthesis approach a program is synthesized from an incomplete specification such as a specification consisting of a set of input-output examples. It has been used recently for synthesizing programs from various domains ranging from low-level pointer manipulating code to spreadsheet macros.
• Ranking technique embodiments described herein generally involve ranking candidate transformation programs generated through an inductive program synthesis procedure using just a small number of user-supplied input-output examples.
• a computer is used for ranking candidate transformation programs to establish a ranked group of one or more transformation programs each of which produces an output string in a user-desired form from each input string entered by a user, consistent with each of one or more user-supplied input-output examples.
• each transformation program is made up of program expressions in a domain-specific language, which are in turn made up of sub-expressions.
• Ranking schemes are established from offline training data for sub-expressions found in the candidate transformation programs. Each of these ranking schemes produces a likelihood score indicative of the likelihood that the sub-expression is part of a program capable of producing an output string in the user-desired form from each input string entered by a user, which is consistent with each of the user- supplied input-output examples.
• a likelihood score is established using the ranking scheme established for that subexpression.
• the likelihood scores are computed in the order from smaller sub-expressions to larger ones. An overall ranking score is then computed for each candidate
• the overall ranking associated with each of the candidate transformation programs is used to select a prescribed number of the top ranking candidate transformation programs. These top ranking programs are then designated as the aforementioned ranked group of one or more transformation programs. It is noted that in one implementation the prescribed number is one, and so only the highest-ranking candidate transformation program is selected and designated.
• FIG. 1 is a flow diagram generally outlining one embodiment of a process for ranking candidate transformation programs.
• FIG. 2 is a flow diagram generally outlining an implementation of the part of the process of Fig. 1 involving establishing ranking schemes for atomic expressions.
• FIG. 3 is a flow diagram generally outlining an implementation of the part of the process of Fig. 2 involving generating training data from a training set of tasks.
• FIG. 4 is a flow diagram generally outlining an implementation of the part of the process of Fig. 3 involving generating training data from each substring expression in sets of positive and negative training sub-expressions.
• FIG. 5 is a flow diagram generally outlining an implementation of the part of the process of Fig. 1 involving establishing a likelihood score for each sub-expression of a candidate transformation program for which a ranking scheme has been established using that ranking scheme.
• FIG. 6 is a flow diagram generally outlining one embodiment of a process for performing string transformations using the highest-ranking candidate transformation program.
• FIG. 7 is a table depicting a spreadsheet having an input column containing input strings representing a person's last name, and an output column having one example output string containing the corresponding input string name with the string "Mr.” placed in front of it.
• FIG. 8 is a table depicting a spreadsheet having an input column containing addresses including a city name, and an output column having one example output string containing a part of the corresponding input string corresponding to the city name.
• FIG. 9 is a diagram depicting a general purpose computing device constituting an exemplary system for implementing ranking technique embodiments described herein.
• input-output example refers to a tuple of character strings that form the input and a character string that forms the output.
• the input represents an example of the input that the user wishes to transform
• the output string represents an example of the output that the user wishes to be produced from the input.
• Ranking technique embodiments described herein use statistical and machine learning techniques to learn a desired ranking function for use in inductive program synthesis.
• a training dataset of positive and negative examples are
• ranking functions are learned that assign an expression in a program in a domain specific language to a likelihood measure. These ranking functions are then used to compute likelihoods of learnt programs from a very small number of input-output examples.
• a trace (or concatenate) expression e denotes the concatenation of atomic expressions fi, ..,f n .
• An atomic expression f can either denote a constant string expression ConstStr(s) or a substring expression SubStr(v 3 ⁇ 4 pi, pi).
• the constant string expression ConstStr(s) denotes an expression that generates the constant string "s".
• the substring expression SubStr(v 3 ⁇ 4 pi, pi) represents a substring of the input string present in column vi whose left and right positions are represented by position pair expressions pi and pi respectively.
• the position expression CPos( ) refers to the k xh index in the given string from the left side (or right side), if the integer constant is positive (or negative).
• the position expression Pos(ri, ri, c) denotes a position in the input string whose left and right sides match the regular expressions ri and ri respectively, and it is the c & such match of the regular expressions in the string.
• r : TokenSeq(r 1; . . , T m ) [0026]
• the data structure maintains independent sub-expressions of an expression independently to represent an exponential number of expressions in polynomial space.
• the set of substring expressions SubStr (v it ⁇ p j ⁇ ., ⁇ Pk)kj maintains the set of left ( ⁇ p_/ ⁇ .) and right position expressions ( Pk)k) independently.
• the mapping W maps each edge in ⁇ to a set of SubStr and ConstStr expressions, thereby achieving sharing even between Concatenate expressions of different lengths (Each path in the dag from the start node i7 s to target node ⁇ t represents a set of Concatenate expressions obtained by taking any expression from each edge in the path and concatenating them.)
• the ranking technique embodiments described herein maintain this independence of sub-expressions and ranks the set of sub-expressions independently. Three techniques for ranking the set of sub-expressions are used: frequency-based ranking (for ranking regular expressions r and position expressions p), feature -based ranking (for ranking Substr expressions), and path-based ranking (for ranking the Concatenate expression).
• a frequency-based ranking is preformed to rank such expressions.
• ⁇ maps each possible expression value to its frequency as estimated from the training data.
• the likelihood score of a regular expression is defined to be directly proportional to its frequency score.
• the likelihood score of a position expression can be defined to be any linear function of its frequency score and square-root of the likelihood scores of its constituent regular expressions.
• a feature-based ranking is performed.
• a set of features 0(e) are computed from the underlying input-output example and the likelihood is obtained from a machine learning procedure.
• a support vector machine (SVM) binary- classification procedure is used in one embodiment.
• a path-based ranking is performed.
• this path-based ranking involves combining scores for individual atomic expressions found in a concatenate expression.
• the likelihood scores for the atomic expressions are combined by multiplying them, and in another implementation they are added together.
• the challenge is to align the edges in the two dags to compute the positive and negative expression values.
• the common edges between them constitutes the positive expression values and the edges that are present in Vk but not in V n constitutes the negative expression values.
• the DAG programs Vk and V n are run on the input string ik and the dag nodes are annotated with the indices of the output string ok using the label function L : ⁇ int.
• Expressions that appear in e l>m are denoted as positive expressions and expressions that appear in the set e i ⁇ m ⁇ fe ⁇ e i ⁇ m ⁇ A are denoted as the negative expressions.
• the set of positive and negative expression values is obtained for each input-output example pair.
• a frequency-based ranking approach is performed as both of them takes values from finite sets.
• a database of token sequences together with their frequency scores for token sequence expressions is also created.
• the frequencies are estimated from the set of positive training expressions.
• the frequencies can also be estimated for each different context in which a certain expression occurs.
• a position expression can either be a left position expression or a right position expression (denoting whether that position expression occurs as the first or second argument of a SubStr expression).
• a regular expression expression also occurs in two different contexts inside a position expression.
• One of the major ambiguities in learning string transformations comes from making the decision as to whether a substring in the output string is a constant string or a substring of an input string. Since making such a decision is not only dependent on the possible values for position pair and constant string expressions but also on the input and output strings, it is not possible to use frequency-based ranking in this case as there are infinitely many possible input and output strings. Instead, in one embodiment, a feature- based ranking approach is used to select between SubStr and ConstStr expressions. More particularly, a set of features is defined for each SubStr expression. Examples of features that can be used, include, but are not limited to, the following:
• IsOutputLeftTok a boolean value that denotes whether the left position of the output substring can be recognized by a token
• IsOutputLeftConstant a boolean value that denotes whether the character at the left position of the output substring is a constant
• IsOutputRightTok a boolean value that denotes whether the right position of the output substring can be recognized by a token
• IsOutputRightConstant a boolean value that denotes whether the character at the right position of the output substring is a constant
• IsOutputTok a boolean value that denotes whether the output substring denotes a token
• IsInputLeftTok a boolean value that denotes whether the left position of the input substring can be recognized by a token
• IsInputRightTok a boolean value that denotes whether the right position of the input substring can be recognized by a token
• IsInputTok a boolean value that denotes whether the input substring is a token; i) Casing: the casing performed to obtain the output substring;
• LenSubstring the length of the substring
• RelLenlnSubstring the relative length of the substring to the length of the input string (lenSubstring/lenlnputString);
• RelLenOutSubstring the relative length of the substring to the length of the output string (lenSubstring/lenOutputString) .
• each feature vector is computed together with a class label (e.g., +1 for positive expressions and 0 for negative expressions).
• An off-the-shelf support- vector machine (SVM) procedure can then used to learn the binary classifier. More particularly, each feature vector created has a prescribed order of elements where each element (except one reserved for the class label) corresponds to a feature value
• a feature value having a first binary value e.g., 1
• a feature value having a second binary value e.g., 0
• the feature vector element reserved for the aforementioned class label is assigned one binary value (e.g., 1) if the SubStr expression is a positive SubStr expression, and another binary value (e.g., 0) if the SubStr expression is a negative SubStr expression.
• Each edge of a dag consists of a set of SubStr and ConstStr expressions.
• the feature based ranking assigns a likelihood score between +1 and 0 to each SubStr expression, while the likelihood score of a ConstStr expression is taken to be 0.5.
• Each path in a dag represents several programs that conforms to the given input- output example.
• Thelikelihood score w(e) of an edge e is defined as the highest likelihood score of any SubStr or ConstStr expression on that edge.
• the likelihood score of a path is defined as w(p) ⁇ rie E edges(p) w ( e suc h mat it is computed by multiplying (or adding) the likelihood scores of the corresponding edge expressions on the path.
• the Dijkstra's shortest path procedure is then used to compute the highest ranked path in the dag, and execute it to generate the highest ranked output string.
• the likelihood score of a path can also be a more sophisticated function of various features such as: the product/sum of the likelihood scores of all edges, the number of edges, the min/max likelihood score of any edge. Furthermore, this function can also be learned by using machine learning techniques.
• a computer such as any of the computing devices described in the Exemplary Operating Environments to follow
• a set of candidate transformation programs is received as input (process action 100).
• Each of the candidate transformation programs was inductively synthesized using conventional methods from the one or more user-supplied input-output examples, and produces an output string in a form exhibited by each user-supplied output example from each user-supplied input example. It is noted that, as described previously, each transformation program is made up of program expressions which in turn are made up of sub-expressions. Next, for each sub-expression of each candidate transformation program for which a ranking scheme has been established (from smaller to larger sub-expressions), a likelihood score is established using the ranking scheme established for that subexpression (process action 102).
• each of the ranking schemes produces a likelihood score indicative of the likelihood that the sub-expression is part of a program capable of producing an output string in the user-desired form from each input entered by a user, which is consistent with each of the user- supplied input-output examples.
• An overall ranking score is then computed for each candidate transformation program from the sub-expression likelihood scores established for that candidate transformation program (process action 104).
• a sub-expression can be either a regular expression or a position expression or an atomic expression or a concatenate expression.
• regular and position expressions employ frequency-based ranking (as this type of ranking is applicable to expressions that take values from a finite set of tokens).
• feature-based ranking is used for atomic expressions
• path-based ranking is used for concatenate. It is also noted that for the sake of efficiency, in one embodiment, ranking schemes are established for regular expressions first, then position expressions, then atomic expressions, and finally concatenate expressions.
• atomic expressions as indicated previously these can be either a constant string expression or a substring expression.
• establishing ranking schemes for atomic expressions in one embodiment first involves identifying whether the atomic expression is a constant string expression or a substring expression (process action 200).
• a training set of tasks is then input (process action 202).
• Each task set is made up of multiple input-output examples.
• Training data is generated from the training set (process action 204), and used to train a machine learning classifier to assign a likelihood score to each substring expression and each constant string expression (process action 206).
• the machine learning classifier is a support vector machine (SVM) binary-classifier, and it is trained to assign a likelihood score to each substring expression having a value in a range from 0 to 1 depending on how likely it is that the sub-expression is part of the intended program.
• the classifier assigns a prescribed likelihood score value (e.g., 0.5) to each constant string expression.
• a previously unselected input-output example in the task is selected (process action 300).
• a set of transformation programs is inductively synthesized from the selected input- output example (process action 302).
• Each of the transformation programs produces an output string in a form exhibited in the output of the example from a tuple of input strings in the input of the example.
• the tuple of input strings refers to the one or more parts of the input in an input-output example that is used to generate the output of the example.
• a set of positive training sub-expressions is established from the set of
• establishing the set of positive training sub-expressions from the set of transformation programs involves designating each sub-expression found in the intersection of all sets of transformation programs synthesized from the input-output examples in the task as a positive training sub-expression.
• establishing a set of negative training sub-expressions from the set of transformation programs involves designating each sub-expression in the set of transformation programs that is not identified as positive training sub-expression as a negative training sub-expression.
• a dictionary is generated that maps sub-expressions that generate an output substring that takes values from a set having a finite number of predefined tokens (i.e., the regular and position expressions) to a frequency-based score based on their frequencies in the sets of positive training sub-expressions (process action 308).
• training data is generated for feature -based ranking (process action 310). It is then determined if all the input-output examples in the task have been selected and processed (process action 312). If not, then process actions 300 through 312 are repeated. Otherwise the procedure ends for the selected task. The foregoing procedure is repeated for all the remaining tasks and their input-output examples.
• a set of one or more features (such as those described previously) is input (process action 400).
• Each of the features is indicative of whether a substring in the output string is produced by a one of the substring expressions in the sets of positive and negative training sub-expressions.
• a feature vector is generated (process action 402).
• each feature vector is a prescribed order of elements where each element, except one reserved for a class label, corresponds to a feature value respectively associated with one of the aforementioned features.
• each feature vector is generated by, for each of the features in an order corresponding to the prescribed order of elements, assigning a feature value having a first binary value (e.g., 1) if the feature is exhibited in a substring in the output string produced by the substring expression under consideration. More specifically, the feature value is assigned to the feature vector element associated with the exhibited feature.
• a feature value having a second binary value (e.g., 0) is assigned to the feature vector element associated with the feature under consideration if that feature is not exhibited in a substring in the output string produced by the substring expression.
• the feature vector element reserved for the aforementioned class label is assigning one binary value (e.g., 1) if the substring expression is in the set of positive training sub-expressions, and a different binary value (e.g., 0) if the substring expression is in the set of negative training sub-expressions.
• this involves combining likelihood scores for individual atomic expressions found in a concatenate expression. In one implementation this combining is accomplished by multiplying the scores, and in one implementation it is accomplished by adding the scores.
• establishing a likelihood score for each sub-expression of a candidate transformation program for which a ranking scheme has been established using that ranking scheme involves the following. Referring to Fig. 5, first the likelihood for each regular expression is identified using a frequency-based ranking scheme (process action 500). Then, the likelihood of each position expression is identified also using a frequency-based ranking scheme (process action 502). As described previously, this frequency-based ranking can involve employing a dictionary that maps regular expressions and position expressions to a frequency-based score. Next, the likelihood of each atomic expression is identified using a feature-based ranking scheme (process action 504). As described previously, an atomic expression is either a constant string expression or a substring expression.
• the feature-based ranking can involve assigning a prescribed likelihood score value (e.g., 0.5). And when an atomic expression is a substring expression, the feature based ranking can involve assigning a likelihood score value (e.g., ranging from 0 to 1) using a machine learning classifier that has been trained to assign a likelihood score to each substring expression.
• a likelihood score value e.g., ranging from 0 to 1
• the likelihood of each concatenate expression is then identified using path-based ranking scheme (process action 506). As described previously, this path-based ranking can involve combining scores for individual atomic expressions found in a concatenate expression.
• the ranking associated with each of the candidate transformation programs is used to select a prescribed number of the top ranking candidate transformation programs, and designate these top-ranking programs as the aforementioned ranked group of one or more transformation programs. It is noted that in one
• the prescribed number is one, and so only the highest-ranking candidate transformation program is selected and designated.
• a set of candidate transformation programs that were inductively synthesized from one or more user-supplied input-output examples are input (process action 600).
• each of the candidate transformation programs produces an output string in a form exhibited by each user-supplied output example from each user-supplied input example.
• the set of candidate transformation programs is ranked to identify a highest-ranked transformation program (process action 602). This highest-ranked transformation program is then applied to a user-supplied input string to produce an output string (process action 604).
• a spreadsheet user has a series of names in an input column and wants to add the title "Mr.” in front of each name.
• the user provided the input-output example shown in Fig. 7.
• the challenge is to learn the desired transformation program from the given input- output example by deciding which substrings in the output string "Mr. Roger" are constant strings and which are substrings of the input string "Roger”.
• the output substring o 1 [0. .0] ⁇ M is not present in the input string, it has to be a constant string.
• the output substring o- l. .1] ⁇ r can come from two different substrings in the input string (i- ⁇ [0. .0] ⁇ R and i [4. .4] ⁇ r).
• the ranking technique embodiments described herein learn all possible expressions to compute the substring "r" in the output string that includes (i) a position pair expression to extract the substring i-jO. .0] and perform lower case operation to produce "r” from "R”, and (ii) a position pair expression to extract the substring i [4. .4], and the constant string expression "r".
• the substrings of output string that do not exist in the input strings are guaranteed to be a constant string, whereas the substrings that do exist in the input strings can be either a constant string or substrings of the input string (although experience has shown they are more likely to be coming from the input string). For example, the output substring o 1 [4.
• Relative length of substring The relative length of substring "r" as compared to the output string is also quite small, i.e., 1/10;
• a spreadsheet user has series of addresses in a column and wants to extract the city names from them.
• the user provided the input- output example shown in Fig. 8.
• the ranking technique embodiments described herein could learn more than 100 different SubStr expressions to extract the substring "Cambridge” from the input string "243 Flyer Drive, Cambridge, MA 02145", some of which are:
• position pair expression pi does not produce any output string from the input string "64 128th St, Seattle, WA 98102".
• the position pair expression pz generates the desired output string for each of the position pair expressions (i.e., Cambridge, Los Angeles, Seattle and San Francisco, respectively).
• the ranking technique embodiments described herein would rank the position pair expression pz higher than the other position pair expressions to generate the desired output string from the one input-output example.
• FIG. 9 illustrates a simplified example of a general-purpose computer system on which various embodiments and elements of the ranking technique
• FIG. 9 shows a general system diagram showing a simplified computing device 10.
• Such computing devices can be typically be found in devices having at least some minimum computational capability, including, but not limited to, personal computers, server computers, hand-held computing devices, laptop or mobile computers, communications devices such as cell phones and PDA's, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, audio or video media players, etc.
• the device should have a sufficient computational capability and system memory to enable basic computational operations.
• the computational capability is generally illustrated by one or more processing unit(s) 12, and may also include one or more GPUs 14, either or both in communication with system memory 16.
• the processing unit(s) 12 of the general computing device may be specialized microprocessors, such as a DSP, a VLIW, or other micro-controller, or can be conventional CPUs having one or more processing cores, including specialized GPU- based cores in a multi-core CPU.
• the simplified computing device of FIG. 9 may also include other components, such as, for example, a communications interface 18.
• the simplified computing device of FIG. 9 may also include one or more conventional computer input devices 20 (e.g., pointing devices, keyboards, audio input devices, video input devices, haptic input devices, devices for receiving wired or wireless data transmissions, etc.).
• the simplified computing device of FIG. 9 may also include other optional components, such as, for example, one or more conventional display device(s) 24 and other computer output devices 22 (e.g., audio output devices, video output devices, devices for transmitting wired or wireless data transmissions, etc.).
• display device(s) 24 e.g., audio output devices, video output devices, devices for transmitting wired or wireless data transmissions, etc.
• typical communications interfaces 18, input devices 20, output devices 22, and storage devices 26 for general-purpose computers are well known to those skilled in the art, and will not be described in detail herein.
• the simplified computing device of FIG. 9 may also include a variety of computer readable media.
• Computer readable media can be any available media that can be accessed by computer 10 via storage devices 26 and includes both volatile and nonvolatile media that is either removable 28 and/or non-removable 30, for storage of information such as computer-readable or computer-executable instructions, data structures, program modules, or other data.
• Computer readable media may comprise computer storage media and communication media.
• Computer storage media refers to tangible computer or machine readable media or storage devices such as DVD's, CD's, floppy disks, tape drives, hard drives, optical drives, solid state memory devices, RAM, ROM, EEPROM, flash memory or other memory technology, magnetic cassettes, magnetic tapes, magnetic disk storage, or other magnetic storage devices, or any other device which can be used to store the desired information and which can be accessed by one or more computing devices.
• Retention of information such as computer-readable or computer-executable instructions, data structures, program modules, etc., can also be accomplished by using any of a variety of the aforementioned communication media to encode one or more modulated data signals or carrier waves, or other transport mechanisms or
• modulated data signal or “carrier wave” generally refer to a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal.
• communication media includes wired media such as a wired network or direct-wired connection carrying one or more modulated data signals, and wireless media such as acoustic, RF, infrared, laser, and other wireless media for transmitting and/or receiving one or more modulated data signals or carrier waves. Combinations of the any of the above should also be included within the scope of communication media.
• program modules include routines, programs, objects, components, data structures, etc., that perform particular tasks or implement particular abstract data types.
• the embodiments described herein may also be practiced in distributed computing environments where tasks are performed by one or more remote processing devices, or within a cloud of one or more devices, that are linked through one or more communications networks.
• program modules may be located in both local and remote computer storage media including media storage devices.
• the aforementioned instructions may be implemented, in part or in whole, as hardware logic circuits, which may or may not include a processor.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Health & Medical Sciences (AREA)
• Computational Linguistics (AREA)
• General Health & Medical Sciences (AREA)
• Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Artificial Intelligence (AREA)
• Machine Translation (AREA)
• Stored Programmes (AREA)
• Document Processing Apparatus (AREA)
• Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
• Complex Calculations (AREA)

Priority Applications (4)

Applications Claiming Priority (2)

Publications (2)

Family

ID=49517708

Family Applications (1)

Country Status (6)

Cited By (1)

Families Citing this family (29)

Family Cites Families (23)

• 2012
  • 2012-10-17 US US13/653,581 patent/US9002758B2/en active Active
• 2013
  • 2013-10-17 WO PCT/US2013/065497 patent/WO2014062948A2/en not_active Ceased
  • 2013-10-17 KR KR1020157009902A patent/KR102129640B1/ko active Active
  • 2013-10-17 CN CN201380054348.2A patent/CN104737155B/zh active Active
  • 2013-10-17 EP EP13786055.7A patent/EP2909740B1/en active Active
  • 2013-10-17 JP JP2015537832A patent/JP6285943B2/ja not_active Expired - Fee Related

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events