mm_action_prediction

DSTC Track 4: SIMMC | Sub-Task #1: Multimodal Assistant API Prediction

This directory contains the code and the scripts for running the baseline models for Sub-Task #1: Multimodal Assistant API Prediction.

This subtask involves predicting the assistant actions through API calls along with the necessary arguments using dialog history, multimodal context, and the current user utterance as inputs. For example, enquiring about an attribute value (e.g., price) for a shared furniture item is realized through a call to the SpecifyInfo API with the price argument. A comprehensive set of APIs for our SIMMC dataset is given in the paper.

Evaluation

Currently, we evaluate action prediction as a round-wise, multiclass classification problem over the set of APIs, and measure the accuracy of the most dominant action. In addition, we also use action perplexity (defined as the exponential of the mean loglikelihood of the dominant action) to allow situations where several actions are equally valid in a given context. We also measure the correctness of the predicted action (API) arguments using attribute accuracy (for Furniture) and f1 score (for Fashion). Specifically, the following API classes and attributes are evaluated.

SIMMC-Furniture

API	API Attributes
SearchFurniture	furnitureType, color
FocusOnFurniture	position
SpecifyInfo	attributes
Rotate	direction
NavigateCarousel	navigateDirection
AddToCart	-
None	-

Each of the above attributes is a categorical variable, modeled as multiclass classification problem, and evaluated using attribute accuracy. Note: minPrice and maxPrice attributes corresponding to the SpecifyInfo action for Furniture are excluded in the current evaluation.

SIMMC-Fashion

API	API Attributes
SearchDatabase	attributes
SearchMemory	attributes
SpecifyInfo	attributes
AddToCart	-
None	-

Each of the attributes takes multiple values from a fixed set, modeled as multilabel classification problem, and evaluated using attribute F1 score.

The code to evaluate Sub-Task #1 is given in tools/action_evaluation.py. The model outputs are expected in the following format:

[
	{
		"dialog_id": ...,  
		"predictions": [
			{
				"action": <predicted_action>,
				"action_log_prob": {
					<action_token>: <action_log_prob>,
					...
				},
				"attributes": {
					<attribute_label>: <attribute_val>,
					...
				}
			}
		]
	}
]

where attribute_label corresponds to the API attribute(s) predicted for each API (refer to the table above) and attribute_val contains the list of values taken by the key attribute_label.

NOTE: We plan to extend the Multimodal Assistant API Prediction from the most dominant assistant action to allow the prediction of a series of multiple actions per turn. Please follow the Latest News section in the main README of the repository for updates.

For more details on the task definition and the baseline models we provide, please refer to our SIMMC paper:

@article{moon2020situated,
  title={Situated and Interactive Multimodal Conversations},
  author={Moon, Seungwhan and Kottur, Satwik and Crook, Paul A and De, Ankita and Poddar, Shivani and Levin, Theodore and Whitney, David and Difranco, Daniel and Beirami, Ahmad and Cho, Eunjoon and Subba, Rajen and Geramifard, Alborz},
  journal={arXiv preprint arXiv:2006.01460},
  year={2020}
}

NOTE: The paper reports the results from an earlier version of the dataset and with different train-dev-test splits, hence the baseline performances on the challenge resources will be slightly different.

Installation (Same across all sub-tasks)

• Git clone the repository:

$ git lfs install
$ git clone https://github.com/facebookresearch/simmc.git

NOTE: We recommend installation in a virtual environment (user guide). Create a new virtual environment and activate it prior to installing the packages.

• Install the required Python packages:
  • Python 3.6+
  • PyTorch 1.5+
  • Transformers

Baselines

The baselines for API Prediction (Sub-Task #1) and Assistant Response Generation & Retrieval (Sub-Task #2) are jointly learnt. For these baselines, the following are the additional dependencies:

pip install absl-py
pip install numpy
pip install json
pip install nltk
pip install spacy

Code also uses spaCy's en_vectors_web_lg dataset for GloVE embeddings. To install:

python -m spacy download en_vectors_web_lg

Code also uses NLTK's punkt. To install:

python
>>> import nltk
>>> nltk.download('punkt')

Overview

Contains the following baselines models:

1. History-agnostic Encoder (HAE)
2. Hierarchical Recurrent Encoder (HRE)
3. Memory Network Encoder (MN)
4. Transformer-based History-agnostic Encoder (T-HAE)
5. TF-IDF-based Encoder (TF-IDF)
6. LSTM-based Language Model (LSTM)

Please see our paper for more details about the models.

Code Structure

• options.py: Command line arguments to control behavior
• train_simmc_agent.py: Trains SIMMC baselines
• eval_simmc_agent.py or eval_genie.py: Evaluates trained checkpoints
• loaders/: Dataloaders for SIMMC
• models/: Model files
  • assistant.py: SIMMC Assistant Wrapper Class
  • encoders/: Different types of encoders
    • history_agnostic.py
    • hierarchical_recurrent.py
    • memory_network.py
    • tf_idf_encoder.py
  • decoder.py: Response decoder, language model with LSTM or Transformers
  • action_executor.py: Learns action and action attributes
  • carousel_embedder.py: Learns multimodal embedding for furniture
  • user_memory_embedder.py: Learns multimodal embedding for fashion
  • positional_encoding.py: Positional encoding unit for transformers
  • self_attention.py: Self attention model unit
  • {fashion|furniture}_model_metainfo.json: Model action/attribute configurations for SIMMC
• tools/: Supporting scripts for preprocessing and other utilites
• scripts/: Bash scripts to run preprocessing, training, evaluation

Pretraining

Run scripts/preprocess_simmc.sh with appropriate $DOMAIN (either "furniture" or "fashion") to run through the following steps:

1. Extract supervision for dominant Assistant Action API
2. Extract word vocabulary from the train split
3. Read and embed the shopping assets into a feature vector
4. Convert all the above information into a multimodal numpy input array for dataloader consumption
5. Extract action attribute vocabulary from train split

Please see scripts/preprocess_simmc.sh to better understand the inputs/outputs for each of the above steps.

Training and Evaluation

To train a model or evaluate a saved checkpoints, please check examples in scripts/train_simmc_model.sh.

You can also train all the above baselines at once using scripts/train_all_simmc_models.sh. For description and usage of necessary options/flags, please refer to options.py or one of the above two scripts.

Results

The baselines trained through the code obtain the following results for Sub-Task #1.

SIMMC-Furniture

Model	Action Accuracy	Action Perplexity	Attribute Accuracy
TF-IDF	77.1	2.59	57.5
HAE	79.7	1.70	53.6
HRE	80.0	1.66	54.7
MN	79.2	1.71	53.3
T-HAE	78.4	1.83	53.6

SIMMC-Fashion

Model	Action Accuracy	Action Perplexity	Attribute Accuracy
TD-IDF	78.1	3.51	57.9
HAE	81.0	1.75	60.2
HRE	81.9	1.76	62.1
MN	81.6	1.74	61.6
T-HAE	81.4	1.78	62.1

Rules for Sub-task #1 Submissions

• Disallowed Input: belief_state, system_transcript, system_transcript_annotated, state_graph_1, state_graph_2, and anything from future turns.
• If you would like to use any other external resources, please consult with the track organizers ([email protected]). Generally, we allow the use of publicly available pre-trained language models, such as BERT, GPT-2, etc.