Mask & Focus: Conversation Modelling by Learning Concepts

Authors: Gaurav Pandey, Dinesh Raghu, Sachindra Joshi8584-8591

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the utility of Mask & Focus, we evaluate it on two datasets. Mask & Focus achieves significant improvement in performance over existing baselines for conversation modelling with respect to several metrics.
Researcher Affiliation Industry Gaurav Pandey, Dinesh Raghu, Sachindra Joshi IBM Research, New Delhi, India {gpandey1, diraghu1, jsachind}@in.ibm.com
Pseudocode No The paper describes its methods and models in prose and with diagrams (Figure 2), but it does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access to source code, such as a repository link or an explicit statement of code release.
Open Datasets Yes The proposed model is evaluated for generating responses on the Ubuntu Dialogue Corpus (Lowe et al. 2015).
Dataset Splits Yes Table 1 depicts some statistics for this dataset: Training Pairs 499,873 20,000 Validation Pairs 19,560 10,000 Test Pairs 18,920 10,000
Hardware Specification No The paper describes model architectures and dimensions (e.g., '500-dimensional word embeddings', 'hidden size of 1,000') but does not specify any concrete hardware details such as CPU/GPU models or memory used for experiments.
Software Dependencies No The paper mentions 'We implemented Mask & Focus using the Pytorch library (Paszke et al. 2017)', but it does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes We use 500-dimensional word embeddings for all our experiments. For the Ubuntu dataset, the utterance and utterance concept encoders are single-layer bidirectional encoders, where each direction has a hidden size of 1, 000. ... We use a fixed vocabulary size of 20, 000. ... We used the Adam optimizer with a learning rate of 1.5 10 4. A batch size of 10 conversations is used for training. ... To prevent the model from overfitting, we use early stopping with log-likelihood on validation set as evaluation criteria.