Sense-Spectrum

These are the codes of the Sense Spectrum model proposed in our IJCNN 2022 paper Dense Embeddings Preserving the Semantic Relationships in WordNet (In case one cannot have access to the official paper, the pre-print version is available here).

In this model, we provided low-dimensional dense embeddings for WordNet noun and verb synsets, where the hypernym-hyponym relationship between two synsets can be preserved in the embedding. Since our synset embedding operates like a spectrum, we call our embedding system Sense Spectrum. We have to specify that this model can only provide the spectra (embeddings) for noun and verb synsets in WordNet. At this moment, we are not able to embed adjective or adverb synsets, since there is no hypernym-hyponym relationship among them.

To implement our model, one may follow these steps:

1. Clone this repository via git clone https://github.com/canlinzhang/Sense-Spectrum.git to your local file.

2. Train the sense spectrum for noun and verb synsets: python train_sense_spectra.py. Only CPU is required for the training. On a Macbook Pro with Apple M1 chip, it takes around 7 hours for training. Also, one can directly download the trained noun and verb spectra via this link. After downloading, please put the two pickle files into the synset_spactra folder.

3. Evaluate the Hypernym Intersection Similarity (HIS) measurement using the SimLex-999 dataset: python evaluate_HIS_dist_on_SimLex-999.py. The HIS measurement out-performs the traditional synset similarity measurements of WordNet by achieving higher correlation covariance score on the SimLex-999 noun and verb word pairs. One may read our paper for more details.

Finally, one can obtain the noun and verb spectrum for a given word. For example, we can obtain the noun and verb synset spectrum for the word 'dog' by implementing the following script:

from obtain_sense_spectra import ObtainSpectrum
Class = ObtainSpectrum('synset_spactra/noun_synset_spectra.pickle', 'synset_spactra/verb_synset_spectra.pickle')

noun_spec, verb_spec = Class.obtain_spectrum_for_word('dog')

Here, noun_spec is the average of the noun synset spectra related to 'dog'. To be specific, by evaluating nltk wordnet, we can see the noun synset related to 'dog' are dog.n.01, frump.n.01, dog.n.03, cad.n.01, frank.n.02, pawl.n.01 and andiron.n.01. Then, the above script will obtain the spectrum for each of these synsets, average the obtained spectra dimension-wise, and then output the average spectra as noun_spec. The verb_spec is obtained similarly from the verb synsets related to 'dog', which is only chase.v.01.

Besides, if one already gets the synset (say, 'dog.n.01') and wants to obtain the corresponding spectrum, simply implement the same class as above and run

spec = Class.obtain_spectrum_for_synset('dog.n.01')

To cite our paper, please use:

C. Zhang and X. Liu, "Dense Embeddings Preserving the Semantic Relationships in WordNet," 2022 International Joint Conference on Neural Networks (IJCNN), Padua, Italy, 2022, pp. 01-08, doi: 10.1109/IJCNN55064.2022.9892238.