A Game Theoretic Approach to Class-wise Selective Rationalization
Authors: Shiyu Chang, Yang Zhang, Mo Yu, Tommi Jaakkola
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the method in singleand multi-aspect sentiment classification tasks and demonstrate that the proposed method is able to identify both factual (justifying the ground truth label) and counterfactual (countering the ground truth label) rationales consistent with human rationalization. |
| Researcher Affiliation | Collaboration | Shiyu Chang1,2 Yang Zhang1,2 Mo Yu2 Tommi S. Jaakkola3 1MIT-IBM Watson AI Lab 2IBM Research 3CSAIL MIT {shiyu.chang,yang.zhang2}@ibm.com yum@us.ibm.com tommi@csail.mit.edu |
| Pseudocode | No | The paper describes the CAR framework and its training procedure textually and with equations, but does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code for our method is publicly available2. 2https://github.com/code-terminator/classwise_rationale |
| Open Datasets | Yes | To evaluate both factual and counterfactual rationale generation, we consider the following three binary classification datasets. The first one is the single-aspect Amazon reviews [10] (book and electronic domains)... We also evaluate algorithms on the multiaspect beer [23] and hotel reviews [29] that are commonly used in the field of rationalization [8, 19]. |
| Dataset Splits | No | The paper mentions 'dev set' and 'training' but does not explicitly provide details on the train/validation/test dataset splits (e.g., percentages or specific counts for each split). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., CPU, GPU models, or memory specifications). |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies (e.g., programming languages, libraries, or frameworks). |
| Experiment Setup | Yes | The hidden state size of all LSTMs is set to 100. The first term constraints on the sparsity of the rationale. It encourages that the percentage of the words being selected as rationales is close to a preset level α. The second term constraints on the continuity of the rationale. λ1, λ2 and α are hyperparameters. The h0( ) and h1( ) functions in equation (2) are set to h0(x) = h1(x) = x. The training scheme involves the following alternate stochastic gradient descent. Specifically, the target factual sparsity level is set to around ( 2%) 20% for the Amazon dataset and 10% for both beer and hotel review. |