@conversationai/wpconvlib

Crowd-Sourcing Tools

This directory contains some tools to support Crowd Sourcing using CrowdFlower. In particular, some JavaScript to allow a prettier layout of the conversation structure.

Setup

This is a standard nodejs/yarn project. You need to have installed npm and node (Recommended to install using nvm, the node version manager).

Once you have npm, install yarn and typescript globally with:

npm install -g yarn typescript

Then, from this directory, you can setup the this node package with:

yarn install

Build for CrowdFlower and Local tests

Crowdflower has a special jquery setup, so its different for testing locally than when run in crowdflower. The difference is between concatenating these two files after the generated app.js (which has the main part of the code):

• src/jquery_crowdflower_require_hack.js
• src/jquery_localtest_require_hack.js

Build for crowdflower

yarn run build:crowdflower

Generated files end up in build/dist/crowdflower.js. This is what should be copied to the crowdflower JS text box.

Test Locally

yarn run build:localtest

Generated files end up in build/dist/localtest.js.

yarn run serve:localtest

In your browser, then open the URL for the corresponding crowdsourcing job i.e.:

• http://localhost:8000/src/testdata/test1.html
• http://localhost:8000/src/testdata/test2.html

And observe the test conversations look right, and that no errors appear in the JavaScript console.

Deploy to CrowdFlower

CrowdFlower has 3 sections that need files from this directory need to be copied to the relevant UI input boxes in a CrowdFlower's Job page (afer you run the build command):

• CML: src/cml{N}.html
• CSS: src/style{N}.css
• JavaScript: build/dist/crowdflower.js

Where {N} is the corresponding crowdsourcing job.

Job 1: Conversation Selection Method: ask questions.

TODO(ldixon): update.

Bad:

• First N comments are OK (score <0.6)
• N+1 comment (by time) has score > 0.6
• Conversation = first N comments.

Good:

• Contains: all comments in a conv (by section)
• Such that: all comments score < 0.4

Good and bad are matched:

• Iterate over bad conv to find a match.
  • Good conv happenning on the same page.
  • Good conv must be >= bad conv size.
  • Select the closest good conv in time (choose the closest good conv, w.r.t. point in time to the bad conv)
  • Select fragment of good conv to match the bad conv in size.
• Each conv only appears once (remove from bucket as we match them)

Job 2: Conversation Selection Method: Ask on Tail of comments about is there a toxic comment.

TODO(ldixon): write and update.