accuracy seems off by a wide-margin wrt original word2vec

On the section 'capital-common-countries' of the questions-words.txt file from word2vec, using

```
./w2v test
```

I get 33.59% Top-1 accuracy vs 80.04% for the original C version released by Google.
