It is a chatbot with the "Open Domain Long Form Question and Answering" (ODLFQA) function, in which a user can ask anything and the chatbot engine can give a long answer. The first version has been developed for Kata.ai chatbot framework. Another frameworks (Rasa or DeepPavlov) will be tested.
The chatbot's ODLFQA engine gathers relevant information from a variety of sources (e.g. paragraphs from relevant Wikipedia pages), synthesizes the information, and creates an easy-to-read original summary of the relevant elements. The engine uses various advanced techniques such as pre-trained language models (e.g. BART, T5) for conditional text generation and a new method for finding information in large data sources (such as Wikipedia) using the REALM or DPR model.
Big thanks to Yacine Jernite from Huggingface who provides the code for "Open Domain Long Form Question and Answering" as jupyter notebook.
Currently the Chatbot is available on Telegram messenger as @UncoolAIBot or you can join the Telegram group AMAPlanet. Following chatbot's commands are available now:
-
/start
It is the welcome action when you join to the chatbot
-
/about
Explain about the chatbot
-
/ask
Ask a question, example: /ask how old is the universe?
If you talk with the chatbot using private message, you can ask a question without the /ask command
-
/joke
Just give a random joke available on internet for entertainment.
It is running on kata.ai server infrastructure which works as the frontend and "hub" between the chat messenger (Telegram in this case) and the Chatbot's ODLFQA Engine
The engine has been dockerized and available at https://hub.docker.com/r/wirawan/eli5.
-
Create a directory /eli5/.cache (or another directory where you have access) before you run the docker for the first time, It needs around 25GB disk space:
% mkdir -p /eli5/.cache -
run the docker image wirawan/eli5:
% docker run --gpus all -p 8000:8000 -v /eli5/.cache:/root/.cache wirawan/eli5.
It will run the engine on port 8000 and provide Rest API interface on http://<hostname>:8000/ask/ and its documentation:
http://<hostname>:8000/docs/.
We will use this URL address http://<hostname>:8000/ask/ in chatbot
config file bot.yml in the action get_eli5.
The first time we run docker run --gpus all -p 8000:8000 -v /eli5/.cache:/root/.cache wirawan/eli5,
it will download all necessary models and datasets, and prepare it for the inferencing. This will take more
than 2 hours depend on your network and server performance. This long process will run only once. All
processed files are saved in local host file system /eli5/.cache, so we can use it for the following run without
long pre-processing time.
- a Windows/Linux server with docker, 100GB Disk space and 16GB RAM
- NVidia GPU (16GB RAM)
The current implementation uses English Wikipedia as data source and the model was pretrained also with Reddit dataset. Therefore it has bias due to its dataset and can give ugly or politically incorrect answer for some questions.