Skip to content

Latest commit

 

History

History
71 lines (64 loc) · 4.42 KB

README.md

File metadata and controls

71 lines (64 loc) · 4.42 KB
Raw

Monte Carlo Tree Search for Multiple Words Substitutions

This code is the original implementation to replicate the Monte Carlo Tree Search (MCTS) simulations as described in the paper "Assessing Robustness of Text Classification through Maximal Safe Radius Computation", accepted as long paper publication at EMNLP-Findings (EMNLP2020).
ACL: https://www.aclweb.org/anthology/2020.findings-emnlp.266/
ArXiv: https://arxiv.org/abs/2010.02004

Dependencies (versions reported have been tested)

You need to install python3, then you have to install the following dependencies (e.g., by pip3): we list the packages versions that we used, but other versions could work:

  • Numpy 1.17.2
  • Nltk 3.4.5
  • Pandas 0.25.1
  • Progress 1.5
  • Tensorflow 1.14.0
  • Torch 1.3.1
  • Torchtext 0.4.0

Further to this, you need to download the GloVe and GloVeTwitter pre-trained embeddings respectively from http://nlp.stanford.edu/data/glove.6B.zip and http://nlp.stanford.edu/data/glove.twitter.27B.zip. Once archives have been unzipped, move them inside the ./data/embeddings folder. Please note that naming should be consistent with variables in ub_CNN_models.py and ub_LSTM_models.py files, i.e., glove.6B.<DIMS>d.txt and glove.twitter.27B.<DIMS>d.txt where is the dimensionality of the embedding (can be specified as a parameter, see the next section). For example, to test GloVe50d, the embedding's name should be glove.6B.50d.txt.

Instructions

To test all the experiments for CNNs, launch from console:
python3 ub_CNN_models.py
While for LSTM models, run:
python3 ub_LSTM_models.py

Please note that you can pass several arguments to these scripts:
-s, --sims: number of Monte Carlo simulations per-vertex
-m, --max-depth: maximum depth of the tree (i..e, number of perturbations)
-e, --eps: max-distance (in L2 norm) for collecting neighbors
-l, --lrate: UTC learning rate
-d, --discount: UTC discount factor
-ed, --emb-dims: embeddings dimension
-hu, --lstm-hu: LSTM hidden units (applicable only to `ub_LSTM_models.py`)

The first time it may take a while as MCTS collects all the neighbors for every word in an input text and it saves them in a file named 'neighbors_<TEST_NUMBER>.pkl' in 'obj' folder (<TEST_NUMBER> is the the i-th input text from test set).

Cite

If you are considering citing this algorithm as part of your work, please use the following bibtex: 

@inproceedings{la-malfa-etal-2020-assessing,
    title = "Assessing Robustness of Text Classification through Maximal Safe Radius Computation",
    author = "La Malfa, Emanuele  and
      Wu, Min  and
      Laurenti, Luca  and
      Wang, Benjie  and
      Hartshorn, Anthony  and
      Kwiatkowska, Marta",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.266",
    pages = "2949--2968",
    abstract = "Neural network NLP models are vulnerable to small modifications of the input that maintain the original meaning but result in a different prediction. In this paper, we focus on robustness of text classification against word substitutions, aiming to provide guarantees that the model prediction does not change if a word is replaced with a plausible alternative, such as a synonym. As a measure of robustness, we adopt the notion of the maximal safe radius for a given input text, which is the minimum distance in the embedding space to the decision boundary. Since computing the exact maximal safe radius is not feasible in practice, we instead approximate it by computing a lower and upper bound. For the upper bound computation, we employ Monte Carlo Tree Search in conjunction with syntactic filtering to analyse the effect of single and multiple word substitutions. The lower bound computation is achieved through an adaptation of the linear bounding techniques implemented in tools CNN-Cert and POPQORN, respectively for convolutional and recurrent network models. We evaluate the methods on sentiment analysis and news classification models for four datasets (IMDB, SST, AG News and NEWS) and a range of embeddings, and provide an analysis of robustness trends. We also apply our framework to interpretability analysis and compare it with LIME.",
}