Distributed Representations for Arithmetic Word Problems

Authors: Sowmya S Sundaram, Deepak P, Savitha Sam Abraham9000-9007

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through an evaluation on retrieval over a publicly available corpus of word problems, we illustrate that our framework is able to consistently improve upon contemporary generic text embeddings in terms of schema-alignment.
Researcher Affiliation Academia Sowmya S Sundaram,1 Deepak P,2 Savitha Sam Abraham1 1Indian Institute of Technology, Madras, India 2Queen s University Belfast, UK {sowmya, savithas}@cse.iitm.ac.in, deepaksp@acm.org
Pseudocode No The paper describes the architecture and training process in text and diagrams, but does not include formal pseudocode or algorithm blocks.
Open Source Code No The paper does not state that its source code for the described methodology is publicly available, nor does it provide a link to a code repository.
Open Datasets Yes Dataset Though there are quite a few public datasets for math word problem solving (Koncel-Kedziorski et al. 2016), the only dataset which comprises word problems attached with unique schemas is the Single Op (Roy and Roth 2016) dataset with 562 word problems. Thus, we choose this dataset for our empirical study. The dataset is manually annotated5 by schema labels for each problem and these labels are used only during evaluation. 5http://tiny.cc/word-probs-data
Dataset Splits No The paper mentions that the dataset was augmented for training but does not provide specific details on train/validation/test splits (e.g., percentages or counts for a validation set).
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions using GloVe embeddings and the SpaCy dependency parser, but it does not specify version numbers for these or any other key software dependencies required for replication.
Experiment Setup Yes Pre-trained Glo Ve embeddings of 100 dimensions were used in all the experiments as the source of the word vectors in the embedding layer. Wherever LSTMs were used, the state size was kept at 256. The GCN network had 4 convolutional layers with 200 hidden states each. The word problems in the dataset was subjected to 50 applications of each operator... resulting in an augmented dataset of 27852 problems that will be used in training.