allenai/allennlp-website

The public website for AllenNLP.

NOTE: we no longer use this website at allennlp.org, as we created an AllenNLP team page at allenai.org. However, this repo is still served via GitHub pages to provide access to some paper pages that were made in the past.

CONTACT: michaels

AllenNLP Website

The AllenNLP website is for the most part a static website. There is no CMS driving any of the content. However, the site is using html abstraction (in the form of includes) to make it easier to edit and maintain. Includes are not a feature of static html, so this abstraction is being handled via Jekyll.

Since the AllenNLP website is hosted on GitHub Pages, Jekyll was the logical choice, as GitHub supports Jekyll out of the box. More about GitHub Pages and Jekyll

Introduction to Jekyll

From the Jekyll Homepage:

Jekyll is a simple, blog-aware, static site generator. It takes a template directory containing raw text files in various formats, runs it through a converter (like Markdown) and our Liquid renderer, and spits out a complete, ready-to-publish static website suitable for serving with your favorite web server.

Since GitHub Pages has deep integration with Jekyll, they make it very easy to run Jekyll sites on their servers. You don't need to configure anything or monkey with build processes. GitHub does that all automatically. You simply push your files and GitHub does the rest.

Here's a tutorial video that outlines Jekyll basics and how the system works.

Modifying the Website

In order to make changes to the website you need to checkout the master branch on the allennlp-website repo. The process for making changes is:

1. Edit file(s) locally
2. Test changes in browser at localhost
3. Commit changes and push to master
4. Test changes in browser at http://www.allennlp.org (usually takes at least 30s to update)

After you've committed your changes, push with the following command:

$ git push

Anatomy of the AllenNLP Jekyll Site (simplified)

📂 _includes        // Where html includes go
📂 _site            // Generated files. Do not edit.
📂 assets
📂 css
📂 js
_config.yml         // Main config file for the Jekyll site.
.gitignore          // Ignores changes to generated _site folder
CNAME               // Points GitHub Pages to allennlp.org
Gemfile             // Configs Jekyll version
Gemfile.lock        // Dependencies
index.html
...more html files

Install Jekyll Locally

Before pushing any changes to the website, you should test them locally in browser. In order to do that, you will need to install Jekyll on your machine.

The most important thing is making sure you have a supported version of Ruby installed.

• See Requirements on Jekyll's website.
• See Dependency Versions on GitHub Pages.

As of 08/15/17 it requires Ruby version 2.4.1. Note that OS X Sierra (10.12) ships with Ruby version 2.0 by default, so you may need to update. See Installing/Updating Ruby on OS X on the Ruby Homepage.

Once you've verified you are up-to-date with Ruby then you can install Jekyll with a single command:

$ gem install jekyll bundler

Running Jekyll Server

Navigate to the local directory where you've checked out the master branch and enter the command to run Jekyll.

$ cd allennlp-website
$ bundle exec jekyll serve
# => Now browse to http://localhost:4000

Now you should be able to see your edits reflected instantly..

allennlp-website's Issues

Models page side-nav links should auto-select based on scroll position

Similar to the scrolling nav experience on PRIOR project detail pages (example), we want to:

• Show users where they are in the page's table of contents
• Automate the building of side-nav menus based on page content

---

Spacing is off on the prediction entries on the website

Should look like

echo '{"passage": "A reusable launch system (RLS, or reusable launch vehicle, RLV) is a launch system which is capable of launching a payload into space more than once. This contrasts with expendable launch systems, where each launch vehicle is launched once and then discarded. No completely reusable orbital launch system has ever been created. Two partially reusable launch systems were developed, the Space Shuttle and Falcon 9. The Space Shuttle was partially reusable: the orbiter (which included the Space Shuttle main engines and the Orbital Maneuvering System engines), and the two solid rocket boosters were reused after several months of refitting work for each launch. The external tank was discarded after each flight.", "question": "How many partially reusable launch systems were developed?"}' > mc-examples.jsonl
allennlp predict \
    https://s3-us-west-2.amazonaws.com/allennlp/models/bidaf-model-2017.09.15-charpad.tar.gz \
    mc-examples.jsonl

Update publications on the website

Matt P. et al: Deep Contextualized Word Embedding (aka ELMo)
Mohit et al: Adversarial Example Generation with Syntactically Controlled Paraphrase Networks
Gabi et al: Supervised Open Information Extraction

bert srl didn't registered

when I tried

        from allennlp.predictors.predictor import Predictor
        predictor = Predictor.from_path("https://s3-us-west-2.amazonaws.com/allennlp/models/srl-model-2018.05.25.tar.gz")
        predictor = Predictor.from_path("https://s3-us-west-2.amazonaws.com/allennlp/models/bert-base-srl-2019.06.17.tar.gz")
        predictor.predict(
          sentence="Did Uriah honestly think he could beat the game in under three hours?"
        )

on the latest version 0.8.4, but got the error:

Traceback (most recent call last):
File "preprocessing/srl_storyline_processing/srl_bert_test.py", line 2, in
predictor = Predictor.from_path("https://s3-us-west-2.amazonaws.com/allennlp/models/bert-base-srl-2019.06.17.tar.gz")
File "/Users/yangjinrui/anaconda3/envs/envi2/lib/python3.7/site-packages/allennlp/predictors/predictor.py", line 144, in from_path
return Predictor.from_archive(load_archive(archive_path), predictor_name)
File "/Users/yangjinrui/anaconda3/envs/envi2/lib/python3.7/site-packages/allennlp/models/archival.py", line 230, in load_archive
cuda_device=cuda_device)
File "/Users/yangjinrui/anaconda3/envs/envi2/lib/python3.7/site-packages/allennlp/models/model.py", line 327, in load
return cls.by_name(model_type)._load(config, serialization_dir, weights_file, cuda_device)
File "/Users/yangjinrui/anaconda3/envs/envi2/lib/python3.7/site-packages/allennlp/common/registrable.py", line 58, in by_name
raise ConfigurationError("%s is not a registered name for %s" % (name, cls.name))
allennlp.common.checks.ConfigurationError: 'srl_bert is not a registered name for Model'

Automate alternating section background color on Models page

On the models page, the section backgrounds for each model section were designed to alternate white, light gray, etc. Since this style needed to be applied manually, the current background scheme is kind of random, as people have been updating the page undoubtedly not knowing about this design intention.

There's a simple way to automate this so that going forward, people updating the page don't have to keep track of which background color should be applied to a given section:

Add a social sharing card

Include authors in contributed model list

We need to add authors to the contributed model list.

We should include the following information:

Deep enhanced representation for implicit discourse relation recognition
Hongxiao Bai, Hai Zhao

Ontology alignment in the biomedical domain using entity definitions and context
Lucy Lu Wang, Chandra Bhagavatula, Mark Neumann, Kyle Lo, Chris Wilhelm, Waleed Ammar

SciTail: A textual entailment dataset from science question answering
Tushar Khot and Ashish Sabharwal and Peter Clark

Bilateral Multi-Perspective Matching for Natural Language Sentences
Zhiguo Wang, Wael Hamza, Radu Florian
Contributed by: HandsomeZebra

Support deep linking on https://allennlp.org/models

Presently it's not possible to link directly to the user gallery.

Rework onboarding by highlighting a single tutorial

@joelgrus is working on creating a new introductory tutorial. Once we have that, we want to re-work the tutorial page so a single tutorial is highlighted, with links out to the more advanced tutorials.

---

Related PRs:

• allennlp-website: Adding new Tutorial experience (#73)
• allennlp: WIP: one tutorial to rule them all (#1613)

Add a link to the demo corresponding with each model

On https://allennlp.org/models we list each of the models, but don't have any way of connecting back to the demo.

We could add a forth "button" that opens a window an pulls up the corresponding demo for that model.

Add the NER model to the website.

http://allennlp.org/models

Tutorial doesn't utilize tqdm

On https://allennlp.org/tutorials, the intro states "We'll use tqdm to track the progress of our training" but tqdm is not utilized in the example code.

Make section padding consistent on the home page

the favicon is really pixelated

we should make a smoother one

Add Event2Mind model section to website

Once we make the new Event2Mind demo live on prod, we should add the associated model description / prediction / training / evaluation content to https://allennlp.org/models.

---

That content can be edited here: https://github.com/allenai/allennlp-website/blob/master/_data/models.yml

Broken link

The "View demo" link under "State of the art models" on the main page is broken.

Make logo list on home page look more professional

As is, these logos are pretty big and kind of inconsistent in their layout:

---

Usually logo lists are more uniform, follow some sort of a grid, and are composed of smaller individual logos:

---

Related issue: #79

Add Terms / Privacy Policy links to footer

It seems this is becoming standard issue for AI2 web properties, so we may as well include the legal stuff on AllenNLP as well.

Add Georgia Tech to affiliates

https://allennlp.org/

Simplify maintenance of Models page content by migrating to a data-driven Jekyll template

This kind of optimization can be done on a per-page basis. No need for a sitewide refactor. No need to change site infrastructure, as this capability is already supported by our GitHub Pages / Jekyll implementation.

PRIOR website is a good example of data-driven Jekyll templates.

HTML boilerplate lives in the template:
https://github.com/allenai/vision-website/blob/master/datasets.html

Data lives in a .yml file:
https://github.com/allenai/vision-website/blob/master/_data/datasets.yml

Users need only edit the data file unless someone needs to make a change to the template mechanics.

Add additional logos to the website

We have several logo listed on https://allennlp.org/, but more organizations are using AllenNLP. For example, We would also like to add:

• Harvard
• UPenn
• https://www.ed.ac.uk/
• Cambridge
• Stony Brook University

---

Related issue: #54

Add Terms and Conditions

From @carissas: Hello again Dustin and Michael,

We're back with one more legal bit! We now have a new Terms and Conditions document that needs to be linked on all of our web properties, in the footer alongside Privacy Policy. We will be fully legit after this is added to PRIOR and AllenNLP's websites.

Could you please add this link to your respective footers/feets?

Link: https://allenai.org/terms.html

Hyperlink in footer example:

Issues with dependency parsing in models.html

I think that there are a some issues with how dependency parsing appears in models.html:

• It appears below the user gallery and how to contribute (instead of with the other models).
• There's no link to it from the sidebar.
• It doesn't have a View Demo link.

where to put pretrained ELMo for other domains

On the ELMo page, there is a place for pre-trained models and a place for contributed models; neither of these tables seems like the right place to add domain-specific ELMo, e.g. the ELMo trained on PubMed by @matt-peters available on S3.

The pre-trained model table reports each model's F1 scores for SRL and Constituency Parsing, so it wouldn't make sense to put the domain-specific ELMo in that table as we wouldn't report performance of the domain-specific model on generic text (I assume). On the other hand, the PubMed ELMo was not 'contributed' by a third party.

Should there be a new section for domain-specific pre-trained ELMo models on the page? Or, should the existing sections be broadened in scope?

The dataset in the tutorial is missing (400 Bad Request)

Hi everyone,

I've been following the tutorial here: https://allennlp.org/tutorials

At the step where the train and test dataset are being loaded, I locally get an exception:

OSError: HEAD request failed for url https://raw.githubusercontent.com/allenai/allennlp with status code 400

allennlp-website/_includes/default-tutorial.html

Line 245 in 922994c

'/master/tutorials/tagger/training.txt'))

My sneaking suspicion is that the link is outdated. Do you have any idea where it went to?

There should be some sort of global / common navigation across different properties.

The AllenNLP web presence continues to grow, and as it does it becomes increasingly difficult to navigate between different contexts. For instance, we have https://allennlp.org, https://demo.allennlp.org and https://docs.allennlp.org. All of these have their own distinct styles and top-level navigation, which makes it hard for a user to navigate between them. The experience doesn't feel very connected / seamless.

In the long-term we should consider using a consistent collection of styles (probably those defined in @allenai/varnish), rather than having bespoke styles per site. This will make it less noticeable when navigating between properties.

In the shorter term we should add some global navigation that makes it easy for users to jump to the key, top-level properties. Here's one such proposal:

Incorporate new AI2-branded web header into AllenNLP website

We're blocked on this until Nick prototypes something with Semantic Scholar.

Oren would like a stronger AI2 emblem than what we currently have on AllenNLP. In particular, he would like us to incorporate
into the header.

Add WTQ model to the models page

https://allennlp.org/models does not include semantic parsing.

Create a "model gallery" of user codebases that implement models in AllenNLP.

Each model record would include:

• Paper name
• Link to GitHub Repo
• Link to paper
• Contributors (maybe)

---

Will not be prominently featured when first launched, but will likely become more prominent in the future.

Use pre-processed styles instead of main.css

The original site was developed with Jade and Less. When moving to Jekyll, the compiled CSS was used since Jekyll doesn't support Less. It does, however, support SCSS which is very similar.

We should convert the original Less to SCSS and use that in place of the main.css to improve maintainability and reduce CSS conflicts.

Clean how how programatic / bash predictions examples are presented

Presently they are each presented in the same box with comments separating the two. Is there a better design? Would some syntax highlighting make them more presentable?