This repository contains code for the paper "Personalization in Goal-Oriented Dialog" by Chaitanya Joshi, Fei Mi and Boi Faltings. We introduce a set of 5 tasks for testing end-to-end dialog systems in a goal-oriented setting with a focus on personalization of conversation. We also provide baselines using various models and publicize our implementations and experimental results through this repository.
Check out the accompanying blog post here.
Update: The paper was published at the NIPS 2017 Workshop on Conversational AI!
The Personalized Dialog dataset can be downloaded using build_data.py. Alternatively, it is accessable using this link or through the ParlAI framework for dialog AI research.
Each of the tasks can also be generated from the bAbI Dialog tasks using the files in the scripts/ directory.
Please refer to the README file inside each model directory for exact installation and usage instructions. In general, we used Python 3.6 and TensorFlow 0.12.1 for our experiments. (Apologies for not providing the exact package versions of other libraries!)
We provide two variants of the dataset- a full set with 6000/12000 dialogs each for training, development and testing; and a small set restricted to 1000 dialogs each for training, development and testing. We also include a second test set (with suffix -OOV.txt) that contains dialogs including entities not present in training and development sets.
The file format for each task is as follows:
1 user_attributes
ID user_utterance [tab] bot_utterances
...
The IDs for a given dialog start at 1 and increase. When the IDs in a file reset back to 1 you can consider the following sentences as a new dialog. The line starting with ID 1 contains attributes associated with a user profile and subsequent lines contain various turns forming the dialog. When the bot speaks two times in a row, we used the special token "" to fill in for the missing user utterance.
For example (for Task 1):
1 male elderly
2 hello greetings sir what may i assist you with today
3 may i have a table in a moderate price range with italian food for eight excellent sir i will start the request now
4 <SILENCE> may i know where the restaurant should be located
5 bombay please excellent sir please give me a moment to provide you with options
6 <SILENCE> api_call italian bombay eight moderate
The goal of the tasks is to predict the bot utterances, that can be sentences or API calls (sentences starting with the special token "api_call").
Along with the train, dev and test sets, we also include a knowledge base file (personalized-dialog-kb-all.txt) that contain all entities appearing in dialogs for tasks 1-5. We also include a file containing the candidates to select the answer from (personalized-dialog-candidates.txt) for tasks 1-5, that is simply made of all the bot utterances in train, dev, test for these tasks.
In addition to the small and full datasets, we also provide a split-by-profile dataset where each directory contains 1000 dialogs each for training, development and testing for a specific user profile. This set can be used to analyze multi-task learning capabilities of models.
We provide our implementations of three models- Supervised Embeddings (supervised-embedding/), Memory Networks (MemN2N/) and Memory Networks with split-memory architecure (MemN2N-split-memory/). Each directory contains scripts, experimental logs and model checkpoints. Instructions on using a models are given in its README.
The dataset is released under Creative Commons Attribution 3.0 Unported license. A copy of this license is included with the data.
- Antoine Bordes, Y-Lan Boureau, Jason Weston, "Learning End-to-End Goal-Oriented Dialog", arXiv:1605.07683 [cs.CL].
- Chaitanya K. Joshi, Fei Mi, Boi Faltings, "Personalization in Goal-Oriented Dialog", arXiv:1706.07503 [cs.CL].