Wan2Vec: Embeddings Learned on Word Association Norms

Tracking #: 2036-3249

Helena Gomez-Adorno
Gemma Bel-Enguix
Jorge Reyes-Magaña
Gerardo Sierra

Responsible editor: 
Guest Editors Knowledge Graphs 2018

Submission type: 
Full Paper
Word embeddings are powerful for many tasks in natural language processing. In this work, we learn word embeddings using weighted graphs from word association norms (WAN) with the node2vec algorithm. Although building WAN is a difficult and time-consuming task, training the vectors from these resources is a fast and efficient process. This allows us to obtain good quality word embeddings from small corpora. We evaluate our word vectors in two ways: intrinsic and extrinsic. The intrinsic evaluation was performed with several word similarity benchmarks,WordSim-353, MC30, MTurk-287, MEN-TR-3k, SimLex-999, MTurk-771 and RG-65, and different similarity measures achieving better results than those obtained with word2vec, GloVe, and fastText, trained on a huge corpus. The extrinsic evaluation was done by measuring the quality of sentence embeddings using transfer tasks: sentiment analysis, paraphrase detection, natural language inference, and semantic textual similarity.
Full PDF Version: 


Solicited Reviews:
Click to Expand/Collapse
Review #1
Anonymous submitted on 30/Oct/2018
Review Comment:

The original comments and suggestions were thoroughly addressed and I believe this revision is much more fair and comprehensive in its presentation of the approach, experiments, and relevant discussion. The quality of writing remains good for the first version and the results are much easier to follow in the revision. There are a couple areas with words extending beyond column widths and inconsistencies in capitalization (e.g. node2vec and Node2Vec).

I'm not convinced the proposed Wan2Vec approach presents significant performance advantages when compared to FastText and other approaches (especially when comparing models with the same dimensionality). In certain situations where WANs are available there may be some performance and computational advantages. Regardless, such a result can still be valuable and informative for an approach that has received relatively little exploration.

Review #2
Anonymous submitted on 30/Oct/2018
Minor Revision
Review Comment:

This second version of the paper is definitely stronger than the previous one. It contains a richer set of experiments and numerous insights. The authors have also included a broad set of references, which were instead missing in the previous version. The writing is easy to follow. Despite some doubts on the scalability of the approach, I would recommend the publication with minor revisions.

It is unfortunate but expected (because of the smaller vocabulary) that wan2vec performs relatively worse on the extrinsic evaluation. The authors should shortly mention how the sentence representations are created in the SentEval toolkit. Is it possible that other ways to create the sentence representation would have worked better? Can the learned word representations be used to train vectors for words that are unseen in WAN?

About the low performance of APSyn: is it possible that the authors have implemented the version used for count-based vectors (i.e. 38) rather than the version for embeddings (i.e. 39)? If this is the case, they should either mention it (it is known that such measure did not work well on embeddings) or implement the other version, which simply includes a smoothing parameters (there should be a freely accessible repository with the implementation).

Review #3
By Thamme Gowda submitted on 13/Nov/2018
Review Comment:

Originality: (4/5)::
This paper uniquely distinguishes itself by taking a different approach to building word embedding, by using resources grounded by psycho-linguistics. The other popular word embedding techniques require a huge corpus of text and computational resources, whereas the method proposed in this paper only relies on fewer high-quality resources (however these resources are hard to obtain and the authors correctly clarify it). The authors analyze their proposed methods from various theoretical and usability standpoints. The evaluation performed on the datasets as well as metrics inform the readers about how this system compares with other popular systems and also informs a suite of experiments that are commonly used by the community.

The significance of results: (3/5)::
The authors report an extensive evaluation of their method and report all the results (that cover a major portion of this paper), including positive as well as some not-so-positive results. The proposed embeddings outperform other unsupervised embeddings on the tasks that evaluate at the word level (word similarity and relatedness task). Authors also analyze and report the limitation of the proposed method on other more complex tasks that work at the sentence level. From the practical standpoint, the embeddings are often the part of a larger NLU/NLG system (like those mentioned in extrinsic evaluation), and this method does not advance the SOTA there, however, stands in the same ballpark as others.

Quality of Writing: (4.5/5)::
Paper is nicely written, with proper references and footnotes wherever needed. The claims made by authors are convincing and provided with sufficient arguments and empirical evidence.

Further Suggestions:
Readers would greatly appreciate a link to download the Wan2Vec embeddings (100d, 128d, 300d) at the footnote of the first page.