Finding Implicatures from User Utterances

This application aims to find implicatures from user utterances in smart home environments.
According to the smart home ontology proposed by Corno et al., there are six categories of environment properties in a smart home: air quality, brightness, humidity, noise level, security, and temperature.
Powered by the Google T5 large model pretrained with UnifiedQA dataset, this application returns if exactly one of these properties is too low or high. That is, the output of the application is one of the followings: LowAirQuality, LowBrightness, LowHumidity, LowNoise, LowSecurity, LowTemperature, HighAirQuality, HighBrightness, HighHumidity, HighNoise, HighSecurity, or HighTemperature.

Starting the application

Before you start the application, please configure the environment variables on .env file.

Environment Variable	Description
BASE_MODEL	The base language model for the application.
PORT	The port number for the application.
CACHE_DIR	The cache directory of the huggingface transformers library.

The easiest way to prepare the application is to use Docker. The following command will create a Docker image.

docker-compose -f docker/Dockerfile --env-file ./.env build

After the build is complete, please run the following command to execute the application.

docker-compose -f docker/Dockerfile --env-file ./.env up

The application could take some time (possibly < 1 min.) to initialize the NLP model. It could take much longer time for the first run because the application downloads required cache files.

Outputs

Predicted implicatures will be returned in JSON format, with predict field. For example, the response for the following request,

curl --request POST \
  --header "Content-Type: application/json" \
  --data '{"query": "I feel vulnerable in this room"}' \
  localhost:8000

should return the response of {"predict": "LowSecurity"} (assuming that the PORT is set to 8000).

Samples

Here are some sample inputs and outputs with BASE_MODEL="allenai/unifiedqa-t5-large".

Input	Output
The wind from the air conditioner is not that chilling.	HighTemperature
The room condition makes me sweat.	HighHumidity
Feeling like someone is here.	LowSecurity
I think the room is too bright now.	HighBrightness
I want to breathe some fresh air.	LowAirQuality
The sound is irritating me.	HighNoise

Limitations

It is surprising that the NLP model works quite well without any fine-tuning. However, the application will not properly working with some inputs, such as

• Input: "Where did the sound come from"
  • Expected Output: HighNoise
  • Actual Output: LowNoise
• Input: "I feel that I have to wear a mask"
  • Expected Output: LowAirQuality
  • Actual Output: HighSecurity
• Input: "The air is so soggy"
  • Expected Output: HighHumidity
  • Actual Output: LowHumidity