Learning to Embed Sentences Using Attentive Recursive Trees
Authors: Jiaxin Shi, Lei Hou, Juanzi Li, Zhiyuan Liu, Hanwang Zhang6991-6998
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model on three benchmarking tasks: textual entailment, sentiment classification, and author profiling. We show that AR-Tree outperforms previous Tree LSTM models and is comparable to other state-of-the-art sentence embedding models. |
| Researcher Affiliation | Academia | 1Tsinghua University 2Nanyang Technological University |
| Pseudocode | Yes | Algorithm 1 Recursive AR-Tree construction |
| Open Source Code | Yes | The implementation is made publicly available.1 1https://github.com/shijx12/AR-Tree |
| Open Datasets | Yes | We evaluated our model using the Stanford Natural Language Inference corpus (SNLI; (Bowman et al. 2015))...We used Stanford Sentiment Treebank (SST) (Socher et al. 2013)...The Author Profiling dataset consists of Twitter tweets and some annotations about age and gender of the user writing the tweet. Following (Lin et al. 2017) we used English tweets as input to predict the age range of the user |
| Dataset Splits | Yes | SNLI consists of 549,367/9,842/9,824 premise-hypothesis pairs for train/validation/test sets respectively...The age prediction dataset consists of 68,485/4,000/4,000 tweets for train/validation/test sets...For all experiments, we saved the model that performed best on the validation set as our final model and evaluated it on the test set. |
| Hardware Specification | Yes | The training on an NVIDIA GTX1080 Ti needs about 30 hours |
| Software Dependencies | No | The paper mentions optimizers like Adam and Adadelta but does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | We set α = 0.1, λ = 1e 5 in Eq. 4 through all experiments. For fair comparisons, we followed the experimental settings in (Choi, Yoo, and goo Lee 2017) on language inference and sentence sentiment analysis. For the author profiling task whose dataset is provided by (Lin et al. 2017), we followed their settings by contacting the authors. We considered their model, which is self-attentive but without tree structures, as a baseline, to show the effect of latent trees. We conducted Tf-idf Tree-LSTM experiment, which replaces the scoring function with tf-idf value while retaining all other settings, as one of our baselines. For all experiments, we saved the model that performed best on the validation set as our final model and evaluated it on the test set. |