TT-SEMPRE: Semantic Parsing for Compound Virtual Assistant Commands

This repository contains TT-SEMPRE, a fork of SEMPRE. TT-SEMPRE is a semantic parser that understands virtual assistant commands of the form when-get-do

Installation

Download dependencies:

./pull-dependencies core corenlp thingtalk

Build the core:

JAVAHOME=<path to java> ant thingtalk

Build the HTTP server:

JAVAHOME=<path to java> ant api

JAVAHOME should be set to the path to your Java installation, eg. /usr/lib/jvm/openjdk-1.8.0. Java 1.8 is required. A working C compiler must also be installed.

Running TT-SEMPRE interactively

./run.sh interactive -ThingpediaDatabase.dbPw <DATABASE_PW> -ThingpediaLexicon.subset <SUBSET>

Where <DATABASE_PW> is the password to the Thingpedia Database, and <SUBSET> is a space separated list of devices to limit the scope of Thingpedia (e.g. twitter instagram).

In interactive mode, it is possible to type sentences and check how they are parsed.

Training

./run-training.sh -ThingpediaDatabase.dbPw <DATABASE_PW> -ThingpediaLexicon.subset <SUBSET>

This command will fetch the data from the Thingpedia dataset and run a full session of training. The trained model (.params file) will be saved under ./almond.

You must have the Berkeley Aligner installed. By default the script looks in the parent directory of the TT-SEMPRE checkout. Set the environment variable BERKELEYALIGNER if you installed it elsewehre.

You can specify the environment variables TRAINING and TESTING (as a comma separated list of Thingpedia dataset names, eg thingpedia,online,turking-prim0) to control the datasets used for training and testing. The default covers the normal Thingpedia dataset used to train Almond.

If you specify the environment variables SEMPRE_USER and SEMPRE_SERVER, the trained model will be uploaded to the given server (using the given user through ssh) and the server will be reloaded.

Running the server

./run.sh server -ThingpediaDatabase.dbPw <DATABASE_PW> -ThingpediaLexicon.subset <SUBSET>

The server runs on port 8400 by default. Use -APIServer.port <X> and -APIServer.ssl_port <X> to change the ports and to enable TLS.

Querying the server

http://127.0.0.1:8400/query?q=<query>&locale=<lang>&limit=<x>

Set <query> to the sentence to parse, <lang> to the locale code (eg. en-US) and <x> to the maximum number of results to report. If limit is unspecified it defaults to 20. If the locale is unspecified it defaults to en-US (American English).

Result:

{"sessionId":"....",
 "candidates":[
    {"prob":0.5,"score":1,"answer":"..."},
    {"prob":0.5,"score":1,"answer":"..."},
 ]}