DDRel: A New Dataset for Interpersonal Relation Classification in Dyadic Dialogues

Authors: Qi Jia, Hongru Huang, Kenny Q. Zhu13125-13133

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also construct session-level and pair-level relation classification tasks with widely-accepted baselines. The experimental results show that both tasks are challenging for existing models and the dataset will be useful for future research.
Researcher Affiliation Academia Qi Jia, Hongru Huang, Kenny Q. Zhu* Shanghai Jiao Tong University Shanghai, China Jia qi@sjtu.edu.cn, onedesire@sjtu.edu.cn, kzhu@cs.sjtu.edu.cn
Pseudocode No The paper describes the baseline models (Random, Majority, CNN, LSTM, BERT) but does not provide any pseudocode or algorithm blocks.
Open Source Code Yes The code and dataset are avaliable at Github 3. 3https://github.com/Jia Qi SJTU/Dialogue Relation Classification
Open Datasets Yes In this paper, we propose a new dyadic dialogue dataset for interpersonal relation classification called DDRel. The dataset consists of 6300 dialogue sessions from movie scripts crawled from IMSDb between 694 pairs of speakers, annotated with relationship labels by human. ... The code and dataset are avaliable at Github 3. 3https://github.com/Jia Qi SJTU/Dialogue Relation Classification
Dataset Splits Yes The whole dataset is split into train/development/test sets by 8:1:1 as shown in Table 2.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, or cloud instance specifications) used to run the experiments.
Software Dependencies No The paper mentions using GloVe embeddings and BERT models, citing their respective papers, but does not provide specific version numbers for software dependencies such as Python, PyTorch, or TensorFlow.
Experiment Setup Yes CNN: '300-dimension pre-trained Glove (2014) embeddings are used and freezed during training. Following the setting of Kim (2014), we use three convolution layers with kernel size equaling 3, 4, and 5... A dropout layer with probability 0.5... The loss function is the negative log likelihood loss. Stochastic gradient descent is used for parameter optimization with the learning rate equaling 0.01.' LSTM: 'Adam Delta as optimizer with learning rate 0.0003.' BERT: 'Adam as optimizer with learning rate 1e 6. We fine-tune the model for 32 epochs at most with early stopping patience equaling 3.'