Seq2seq SPLAT

This is the code repository for our BlackBoxNLP 2023 paper "Why bother with geometry? On the relevance of linear decompositions of Transformer embeddings." It mainly contains utility classes to linearly decompose input sentences, along with some analysis scripts and result files.

Use this code

The file extract_marianmt.py defines a Decomposer class. That's your main entrypoint. Construct a Decomposer, then call it with the source and target sentences. That will return a L x T x 5 x D tensor where:

• L is either 1 or the number of layers in the model,
• T is the target sequence length in tokens,
• 5 corresponds to the terms we decompose embeddings into, and
• D is the embedding size of the model.

Note that the file also defines some short hands for decomposition indices (I, S, T, F, C)

An example of usage is provided in dump_sims.py, which defines a few similarity metrics and dumps in a CSV all similarity measurements for an input parallel corpus.

Similar scripts for token-wise decomposition inspired by Oh & Schuler (2023) can also be found in this repo.

Citation

If this code has been useful to you in any way, please consider citing our publication:

@inproceedings{mickus-vazquez-2023-bother,
    title = "Why Bother with Geometry? On the Relevance of Linear Decompositions of Transformer Embeddings",
    author = "Mickus, Timothee  and
      V{\'a}zquez, Ra{\'u}l",
    editor = "Belinkov, Yonatan  and
      Hao, Sophie  and
      Jumelet, Jaap  and
      Kim, Najoung  and
      McCarthy, Arya  and
      Mohebbi, Hosein",
    booktitle = "Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.blackboxnlp-1.10",
    doi = "10.18653/v1/2023.blackboxnlp-1.10",
    pages = "127--141",
    abstract = "A recent body of work has demonstrated that Transformer embeddings can be linearly decomposed into well-defined sums of factors, that can in turn be related to specific network inputs or components. There is however still a dearth of work studying whether these mathematical reformulations are empirically meaningful. In the present work, we study representations from machine-translation decoders using two of such embedding decomposition methods. Our results indicate that, while decomposition-derived indicators effectively correlate with model performance, variation across different runs suggests a more nuanced take on this question. The high variability of our measurements indicate that geometry reflects model-specific characteristics more than it does sentence-specific computations, and that similar training conditions do not guarantee similar vector spaces.",
}