Intro

Frequency-Aware Cross-Entropy (FACE) is a simple yet effective algorithm that helps to improve the response diversity of Seq2Seq-based chatbots. The main idea is to assign token frequency-based weights to cross-entropy loss function, so as to suppress meaningless high-frequency tokens, which we believe to have caused generic responses like "I don't know". Read our paper for more details.

This repo contains the official implementation of the models proposed in paper Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss, together with the data we used for experiments.

Requirements

To use this programme, you need PyTorch 1.0.0 (tested with Python 3.6+) and the latest version of ParlAI framework (tested on commit c6745203).

After downloading and installing ParlAI, move the directory face in this repo to Your_Path_To_ParlAI/parlai/agents/. Read on for some examples of running experiments.

Arguments

To use different versions of FACE, try the following arguments:

• -wt or --weighing-time: Supported values are pre, post, and none, which correspond to "pre weight" (Eq. 10 in paper), "post weight" (Eq. 11) and "no frequency-based weights" (in cases of using CP* methods only), respectively.
• -ft or --frequency-type: Supported values are out, gt, and none, corresponding to "output frequency", "GT frequency" and "no frequency" (in cases of using CP* methods only).
• -cp or --confidence-penalty: Supports values like cp (Eq. 12), cpf (Eq. 13), cpfw (Eq. 14), and none if not intending to use confidence-penalty. N.B., cpfwn is a new version of cpfw that normalizes weight values to the range of [1, +inf], while cpfw weights can never approach 1.
• -b or --beta: Penalty strength specifically used by cp (Eq. 12).
• Seq2Seq: if all the values of -wt, -ft, -cp are set to none (and regardless of the value of -b), the programme reverts to its simplest form: Seq2Seq.

Examples

After installing ParlAI and FACE, change your working directory to ParlAI root.

The following example runs FACE with "pre weight" and "output frequency", as is reported our best-performming model (on OpenSubtitles dataset, with batch_size = 32, and a validation period of 30s of training):

python examples/train_model.py -t OpenSubtitles -m face -mf /tmp/model_face -bs 32 -vtim 30

To get a complete cheatsheet of ParlAI- or FACE-specific arguments, run the following command:

python examples/train_model.py -m face -h

Data

In cases when you want to reproduce the results reported in our paper, decompress data.zip. Then the subdirectories:

• OSDb: Contains validation and test sets used in our experiments for the OpenSubtitles dataset. We applogize that the training set is too large to share online.
• Tweet_IDs: Contains the Tweet IDs for Train, Valid, Test and Human_eval sets used in our experiments.

Citation

If you use the materials provided in this repo, please cite our following paper:

@inproceedings{jiang2019improving,
	Author = {Jiang, Shaojie and Ren, Pengjie and Monz, Christof and de Rijke, Maarten},
	Booktitle = {The Web Conference},
	Title = {Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss},
	Year = {2019}}

Baseline Models

I'm cleaning up the baseline models used in our paper, namely MMI and MHAM models and their variants. There will be separate repos for these models, but before that, you can find my current implementation here if you really want them :wink: