The links to the data files in the README appear to be broken. Can you please update those?

Hi there. I am implementing the paper myself by referring to that repertoire. I have one question.

In the training code you wrote, I understood that you save Ulist, Vlist respectively as training results, but in the visualization code, you use an embedding weight file named 'emb_frobreg~ '.
I didn't see any code that results in that file. Is there a separate process for calculating it ?

thank you for your time :)

I am trying to run your visualization code on your embeddings. I get this error:

/DynamicWord2Vec/visualization$ python tsne_of_results.py
Traceback (most recent call last):
File "tsne_of_results.py", line 19, in
fid = open('data/wordlist.txt','r')
FileNotFoundError: [Errno 2] No such file or directory: 'data/wordlist.txt'

Do you still have that file available somewhere?

Step 1: I loaded embeddings from the folder "./embeddings". The provided folder has only 26 files while one is missing for NYT alignment quality. I copy the last file embeddings_25.m as embeddings_26.m to have 27 embeddings in total.
The results for the alignment quality test1 are (MRR, P@1, P@3, P@5, P@10 respectively)
0.1027 0.0494 0.1042 0.1340 0.1962
while the reported one is:
0.4222 0.3306 0.4854 0.5488 0.6191
For test2: I get
0.1161 0.0449 0.1079 0.1775 0.2989
while the reported one is:
0.1444 0.0764 0.1596 0.2202 0.3820

Step 2: When I tried to loaded the pre-trained static word embedding and PMI matrices to train it by using the provided code (train_time_CD_smallnyt.py). The hyperparamters are the same in the paper and the code, and I also tried smaller batch size, which seems to be more stable. The best performance is a little bit better than Step 1.

Step 3: I tried to train the word embedding and calculate PMI matrices by myself. The performance was not improved.

Could you please provide some tips for these issues? or provide the evaluation for the alignment quality task to check whether I implemented the evaluation properly?

best,
Benyou

Hi,
Sorry for this question but after I train the model, I will get some U and V s as pickle files. However, I notice that the embeddings used in visualization and other tasks are matlab data files. How do we get to that? Can you please direct me to a resource to do that?

Hi,

Is it possible to generate on my own data if i have individual embeddings generated by word2vec of temporal data? I have a total of ~10 embeddings. Which script should i use for this. is it ok to modify accordingly?

thank you for your time

Could you renew the links?

Hi,
I am trying to get dynamic embeddings for my dataset. I have the data file in the format word,context,ppmi. But while running the main train script, I get the following error:

Traceback (most recent call last):
  File "train_time_CD_smallnyt.py", line 115, in <module>
    pmi_seg = pmi[:,ind].todense()
AttributeError: 'matrix' object has no attribute 'todense'

Can you help me understand why this is happening?

Hi, I am trying to implement the results in the paper. Some issues I met hindered me:

1. The provided wordlist.txt does not match the evaluation set: only 2294 out of 11028 words can be found in test1 (the alignment quality task) and 5 out of 445 words can be found in test2. The size of word vocabulary in the scripts is 20936, but the provided emb_static.mat has 20000 words.
2. The provided pickled files for baselines have 20 time slices, but the NYT data itself has 27 time slices. This makes it difficult to reproduce the baseline results as well.
   Thanks for your attention in advance