Concha

An interactive command shell who deals with Spanish natural language and outer world.

Why Concha?

Because it is a shell... and a bad word in Latin America.

Concha aims to use Spanish natural language as a programming language, not just because it is cool, but because Spanish is a pretty ruled language (it is officially handled by the Real Academia Española, or RAE.

Which are the main goals

There are two main goals:

• There are many natural languages who can answer questions related to an "internal world". Concha must interact with outer world, mainly through HTTP calls in the most flexible way possible.
• There are plenty of chatbots who interact with your phone to set up an alarm or to buy things in your favorite online, but those are pre-trained tricks. Concha must be able to be taught about new tricks. This is not deep learning but superficial teaching.

Current status

Currently Concha can:

• Be taught about new tricks through HTTP services, described in JSON format according to CoNNL Universal Dependences
• React upon action requests sent in plain Spanish documents according to stored tricks. All of them are stored as resources for future reference and usage.
• Define external actions to interact ith external HTTP GET and POST services
• Define internal actions to TREAT a document or part of it in order to do direct modifications or to recursively explore new approaches to the full document.
• Build unattended constructs of different tricks in a recursive manner to solve documents who have dependent clauses.
• Respond in plain Spanish combining parts of incoming document or intermediate tricks output.

Current extra-bonusses:

• There's a Servers mock to play with as an external orld playground
• It was meant to be Spanish only, but... what the heck! it speaks 40 more languages as well, English included (you have to download the proper model for the SyntaxNet Universal Parser).

Currently Concha has some limitations:

• It has no persistence. All tricks are forgotten once it stops.
• It is synchronous
• It calls in an extremely innefficient way to external parsers (who runs several paralel TensorFlow models loading from scratch every call), so response times can go far beyond 10 seconds.

Comparison to Shellscript It is not a programming language yet. If we compare it to a common shell script it only does:

• Command acceptance (HTTP endpoint acts as a prompt).
• Process execution (HTTP).
• Processes pipelining (|) joining outputs with inputs (done in an autonomous way in this case)
• Process execution return value (following HTTP coding).

Installation

Well, this is pretty simple. Just clone the project parallel to TensorFlow/SyntaxNet and run concha.py in python3 from concha/concha execution directory.

1. Have tensorflow/models/research/syntaxnet installed on whatever X directory.
2. Follow the SyntaxNet installation instructions detailed here
3. There should be some directory structure like this: X/models/research/syntaxnet/syntaxnet. If you have problems with the created pip installation, you can re-try the final phase doing sudo -H pip install --ignore-installed --upgrade /path/to/binary.whl
4. Clone this very repository over X directory.
5. There should be some directory structure like this: X/concha/concha and you are done.

Credits

• Files parse.sh and context.pbtx on concha/concha directory are part of universal-parsey as mentioned in SyntaxNet documentation, but they are not included in the official repository. You can find them in mldb.ai organization at this repo (it should be located at models/research/syntaxnet/syntaxnet/models/parsey-universal/)
• Laguage models located at concha/lang_models are referenced in the SyntaxNet documentation, but they claim you must download them. Given they are exposed under the same Syntaxnet Apache 2.0 License, I've included them here for your convenience.

Execution

Follow the Jupyter notebook tutorial in your own cloned repository.

Final goals

When we, humans, teach each other (and I'm not talking about learning, but teaching), we use tonnes of natural language piled through centuries to transfer wisdom, to tell how to do things, to register queryable facts... Evolution must not be that wrong, otherwise we should be talking in P-Code or Assembler. In next to come versions, Concha will:

• Be taught in plain spanish (without intermediate formal declaration of new tricks).
• Ask for help to other Concha instances to solve hat it has been asked for.
• Teach other Concha instances using Spanish.