Quantum-inspired Neural Network for Conversational Emotion Recognition
Authors: Qiuchi Li, Dimitris Gkoumas, Alessandro Sordoni, Jian-Yun Nie, Massimo Melucci13270-13278
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our framework on two benchmarking conversational emotion recognition datasets, namely MELD (Poria et al. 2019a) and IEMOCAP (Busso et al. 2008a). The results show that the provided formal quantum view of conversational emotion recognition does not lead to the drop in performance: our model achieves comparable accuracy performances to state-of-the-art models on both datasets, with slightly improved values on particular metrics. Moreover, the introduced training algorithm for unitary matrix brings to affordable drop in efficiency. |
| Researcher Affiliation | Collaboration | 1 University of Padua, Padua, Italy 2 The Open University, Milton Keynes, UK 3 Microsoft Research Montreal, Montreal, Canada 4 Universit e de Montr eal, Montreal, Canada |
| Pseudocode | No | The paper describes the methodology with equations and textual explanations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The codes are implementation in Py Torch, publicly available on Git Hub 7. github.com/qiuchili/diasenti.git |
| Open Datasets | Yes | We evaluate our model on two benchmarking datasets, IEMOCAP (Busso et al. 2008b) and MELD (Poria et al. 2019b). |
| Dataset Splits | Yes | Dataset # dialogues # utterances train dev test train dev test IEMOCAP 96 24 31 6808 1702 1623 MELD 1039 114 280 9989 1109 2610 |
| Hardware Specification | Yes | The experiments are run on a Linux server with one NVidia Tesla V100 Graphic card. |
| Software Dependencies | No | The paper mentions that "The codes are implementation in Py Torch" but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | On both datasets, QMNN hyper-parameters are searched within embedding dimensions d {100, 120, 140, 160, 180, 200}, the size of last hidden layer in {32, 48, 64, 80}. Stochastic gradient descent (SGD) is used as the optimizer with a learning rate lr {0.001,0.002,0.005,0.008}. The unitary matrix training algorithm is also modified to an SGD fashion, where the general gradient G in Eq. 7 is replaced by the SGD gradient. The learning rate unitary-lr for updating the unitary matrix varies in {0.001, 0.002, 0.005, 0.008}. The batch size bs varies in {24, 48, 96} for MELD and {4, 8, 16} for IEMOCAP in proportion to the dataset scale. The dropout rate for the last hidden layer varies in {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}. |