Multiview LSA

This is the companion code for the paper: Multiview LSA: Representation Learning Via Generalized CCA, Pushpendre Rastogi, Benjamin Van Durme and Raman Arora, NAACL(2015).

@conference{rastogi2015multiview,
  Author = {Pushpendre Rastogi, Benjamin {Van Durme} and Raman Arora},
  Booktitle = {Proceedings of NAACL},
  Keywords = {mvlsa,multiview lsa,mvppdb},
  Title = {Multiview LSA: Representation Learning Via Generalized CCA},
  Year = {2015}
}

NOTES:

  1. We have made available the best performing embeddings in both /mat/ and /svmlight/ formats (Table 9: column MVLSA Combined) at DOI. Download the file combined_embedding_0.mat from that collection or combined_embedding_0.emb.ascii and combined_embedding_0.word.ascii If you decide to use other embeddings files then please note that the embedding matrices in the matlab file need to be aligned to the vocabulary and then normalized as shown in the code below.

     word=textread('$(VOCAB_500K_FILE)', '%s');
     load('$(EMB_FILE)');
     if exist('sort_idx')
        word=word(sort_idx); % VERY IMPORTANT STEP 1 !!!
     end;
     G = normalize_embedding(G); % VERY IMPORTANT STEP 2 !!!
     conduct_extrinsic_test_impl(G, ...);
    
  2. The code is hosted at github.com/se4u/mvlsa

Acknowledgements

This material is based on research sponsored by the Defense Advanced Research Projects Agency (DARPA) under the Deep Exploration and Filtering of Text (DEFT) Program (Agreement number FA8750-13-2-0017). We also thank Juri Ganitkevitch for providing the word aligned bitext corpus as part of the PPDB project

Detailed Description

MultiView LSA works in 4 stages

  1. EXTRACT_COUNT
  2. PROCESS_COUNT
  3. MVLSA
  4. EVALUATE

Every stage creates files for the next stage. To run a particular stage, just run the shell script associated with it. For example, run process_count.sh after running the EXTRACT_COUNT stage that dumps co-occurrence counts into sparse .mat files.

pipeline.sh would run all the steps below Assuming that you have downloaded the matlab files containing co-occurrence counts to the directory EXTRACT_COUNT_FOLDER and the VOCABWITHCOUNT_500K_FILE points to the vocabulary file that you downloaded. These configuration variables are specified in commonheader.mk.

EXTRACT_COUNT

This is a tedious process with lots of grunt work, and though the code is provided you might not have access to the underlying resources. For example I used word aligned bitext corpora that were used as inputs for PPDB amongst others like FrameNet.

To solve this problem just download the extracted co-occurrence counts as svmlight files or .mat file and go to the next step. Look at extract_count.sh and extract_count.mk for more details.

PROCESS_COUNT

There was some non-trivial tuning involved in getting good results with the Multiview LSA paper. See the file process_count.sh to see the best setting reported in the paper. You can simply run ./process_count.sh assuming that EXTRACT_COUNT_FOLDER defined in commonheader.mk is set properly and that you downloaded the files from the previos step into that folder. Note that my code works with matlab files even though I provide the svmlight from previous stage for convenience. Also note that this process can be trivially parallelized on clusters since the commands in process_count.sh are independent of each other.

MVLSA

The heart of the algorithm. At its simplest it is just an SVD of the concatenated matrices derived from previous step. But care is required so that the matrices are not simultaneously loaded in memory and missing values need special attention. Run:

./mvlsa.sh

This script would produce the best embeddings reported in the paper. Note that this process is the slowest step, one by one we would load the mat files produced in the previous stage and then incrementally update our estimate of the left singular vectors. We load one matlab file at a time to balance memory usage and disk access but that is easy to change to load more or less data. Note that this process needs close to 10GB memory and 2.5 minute per view to run on a 10 core machine. (roughly 2 hours) It is possible to decrease the run time and memory required through more intelligent IO. (pre-loading data, mmap ?) but that’s not built in. The best embeddings are also provided for download as:

 `v5_embedding_mc_CountPow025-trunccol12500_500~E@mi,300_1e-5_16.mat`

EVALUATE

Once we have trained the embeddings they should appear in the EMB_FOLDER defined in the commonheader.mk file. For example, running ./mvlsa.sh with default settings creates:

 `$EMB_FOLDER/v5_embedding_mc_CountPow025-trunccol12500_500~E@mi,300_1e-5_16.mat`

Now we can evaluate the embedding to reproduce the results reported in the paper. Run ./evaluate.sh to get results like the following

The TOEFL score over G with bare [80, 78, 72] is 0.900000
Now working on SCWS_FILENAME
The SCWS Pearson correlation over G (2002 out of 2003) is 0.655713
The SCWS Spearman correlation over G (2002 out of 2003) is 0.674836
Now working on RW_FILENAME
The RW Pearson correlation over G (1868 out of 2034) is 0.380015
The RW Spearman correlation over G (1868 out of 2034) is 0.411926
...
...

Note that running the evaluation if you did not extract the counts yourself requires you to download the vocabulary file that I used for creating these embeddings. and store it in the vocabulary file folder. Change variable VOCABWITHCOUNT_500K_FILE in file commonheader.mk to change the location of the vocabulary file. You can download the vocabulary file that we used to evaluate our embeddings as well. By default the script produces the results without combining with Glove embeddings, or Word2Vec embeddings, but the code provides enough hooks to easily do it.

MRDS

MRDS is a simple way to assign a minimum threshold t to a testset of size N. If the correlation coefficients of two competing methods differ by t then the difference is significant otherwise not. Please note that the above description hides details and you should see the paper for a more qualified and measured explanation. Run

./mrds.sh

At its default settings mrds.sh would produce the last column of Table~2 of the paper that describes the datasets used.

Possible Future Work

  1. Task specific representation learning through feedback guided weights.
  2. Which contexts give a boost (this is part of analysis) Basically we code PMI, PPMI, Glove’s Data dependent preprocessing as different views and then find which views get a high weight. I should decide the exact weighting strategy to use. Setting x-max really benefits the Semantic dataset however I can do a lot better in terms of weighting by carefully either either premultiply or postmultiply and then get basically a factored weighting.
  3. Finally do humans really break the performance intro matrices of statistics that are called views ?
  4. Tension between thresholding for noise removal and “missing value imputation for svd”.