Recognizing intents with slots using OpenNLP.
This is an example of using OpenNLP to train a system to accept natural language input, particularly via a speech-to-text source, and return a recognized action with arguments. The system uses document categorization to determine the action for inputs and entity recognition to determine the arguments. The training system requires a directory containing separate files for each possible action, in this case the actions in a fictitious weather application:
- example/weather/train
- current-weather.txt - get the current weather
- hourly-forecast.txt - get the hourly forcast
- five-day-forecast.txt - get a five day forecast
Each training file contains one example per line with any possible arguments surrounded by mark up to indicate the name of the parameter:
file: five-day-forecast.txt
...
how dos the weather look for this Thursday in <START:city> Boston <END>
is it going to snow this week in <START:city> Chicago <END>
show me the forecast for <START:city> Denver <END>
...
The training systems is run passing in the training file directory and any parameter name used in the training files.
$ mvn clean compile exec:java -Dexec.args="example/weather/train city"
...
Training complete. Ready.
>show me the weather for chicago
action=current-weather args={ city=chicago }
>will it rain tonight
action=hourly-forecast args={ }
>how does it look in seattle
action=hourly-forecast args={ }
>what are the conditions in new york
action=current-weather args={ city=new york }
>how does this weekend look in boston
action=five-day-forecast args={ city=boston }
>give me the five day forecast
action=five-day-forecast args={ }