This is a tool to visualize the distribution of attention in a text-based sequence-to-sequence task such as summarization. As you hover your mouse over the decoded words, the tool shows a heatmap of attention over the source words. A demo can be seen here (scroll down to "Example Output" section).
Additionally, for pointer-generator networks such as that described in this paper, the tool displays the generation probability of each decoded word. This tool was designed to work with the Tensorflow code for the paper.
To run the visualizer, run for python 2.7
python -m SimpleHTTPServer
for python 3.5
python -m http.server
from this directory then navigate to http://localhost:8000/
in browser. The visualizer will show some example data.
To visualize your own data, you need to replace attn_vis_data.json
with a similar file, either produced by this Tensorflow code, or by your own model. In particular attn_vis_data.json
should contain the following fields:
article_lst
: the article (or source text) as a list of wordsdecoded_lst
: the decoded (i.e. machine-generated) summary as a list of wordsabstract_str
: the reference summary as a single stringattn_dists
: a list same length asdecoded_lst
, containing lists of length "attention length", containing probabilities. Note attention length must be <= length ofarticle_lst
. e.g. your article may have 500 words but you only fed the first 200 words into the model, thus attention length is 200. In this case the visualizer will mark the truncation point in the article.p_gens
: a list same length asdecoded_lst
, containing the generation probabilities.
WARNING: Make sure that none of the strings in article_lst
, decoded_lst
, or abstract_str
contain <angled brackets>
. These will interfere with the HTML and can result in text not being displayed.