Programming Language Naturalization

Have u ever thought of a programming language which can be written in natural language. We came across different kind of applications which need graphs to be plotted and required data has to be stored and some complex actions have to be performed using Internet of things or on any other data. The above requirements can be accomplished using a programming Language which has to be written precisely following all the rules.

   

On the other hand, there is a method which can convert natural language into formal language. This can be done using semantic parsing. The ability of this parsing is limited and not as powerful as implementation through programming. Example for this is “Voxelurn”.

Example for natural language programming   

This concept is called “naturalization”. This bridges gap between natural language and core language. In any application development, we need to select a core language and we need to train the system with rules (conversion of natural language to core language).  Rules will be like ‘A’ means ‘M’. Initially, user has to instruct the system about the rules. After training the system will be able to understand the natural language (which is selected by the user) and can convert to core language such that it can be executed.     

“Voxelurn” was implemented by naturalizing “Voxel world”. Voxel has lot of loops, structures, Conditionals. A group of people started training system and system can be used by other users. A lot of short forms and alternative syntax are added making system works with natural language instead of core language.

Some rules used to train the system:

Modeling and Learning:

·         ‘X’ utterances to be mapped to core language.

·         ‘Y’ is tokens in core language.

·         ‘S’ set of tokens we know in natural language that can be mapped to core language.

·         'S’' is the new set which is reconstructed after learning new items.

The conversion of ‘X’ to ‘Y’ can be done using a derivation tree which has been constructed by the feature set.

Features:

Derivations are based on the weights given to different type of features.

Rule Features:

“Id” is for specific rules.

“Type” feature is to identify whether it is a core language or induced language.

Social Features: These features capture linguistic similarities and dissimilarities among users.

“Author” is to generalize the precise things to be accepted.

“Friends” is mapping between authors and users which means users tend to expect the rules which are defined by authors those are in their community.

“Self” this specifies whether it is written by the user

Spam Features:

Considering adjacent tokens of the border to work effectively using context specific

Scope Features:

This specifies how each community prefers to have scoping features

Parameter Estimation:      

        

                After each utterance or token, user will be given all the possibilities of next state and user can choose one among them. Then system performs an online AdaGrad update on the parameters according to gradient of loss function.

Conclusion: 

            In Future, this kind of conversion can be implemented for other programming languages. Semantic parsing, Grammar plays an important role.

References:

[1] Sida I. Wang, Samuel Ginn, Percy Liang, Christopher D. Manning, Naturalizing a Programming Language via Interactive Learning

[2] S. Zettlemoyer and M. Collins. 2005. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars.

[3] S. Zettlemoyer and M. Collins. 2007. Online learning of relaxed CCG grammars for parsing to logical form.