attendee-search

attendee-search is a command-line like tool to find EAG attendees with specific interests.

It performs a "fuzzy search" over text fields in SwapCard profiles. This functionality is much better than using the search features on SwapCard. For some people, it's probably better than using Ctrl-F in Excel.

This tool is especially useful for finding attendees with niche interests.

This tool has been created quickly (this afternoon), please contact me if you have any questions or issues.

Installation

1. Have a Python 3.7+ environment
2. Install the packages in the requirements.txt:
   • Packages include transformers, pandas and tqdm
3. Download the Excel (*.xlsx) file from SwapCard that contains attendee data (this should have approximately 1,500 rows)
4. Run main.py

Note that the tool expects the Excel file that is downloaded from SwapCard. Don't modify this .xlsx file, just download and move it to the directory with main.py. The Excel file is a little idiosyncratic, and this tool expects the column names to be on a certain row (the 4th row).

How to use the tool

Run main.py. You should get a text prompt, with a few simple instructions.

When you run the tool for the first time, it will generate and save an "embedding file" for the attendee data.

I haven't included screenshots in this public repo, because I don't want to show attendee information. If you're an EA who wants help or more information, please contact me (e.g. using the communication method that led you to this repo).

Drawbacks

Compared to the alternatives, the tool works well in finding attendees with more niche interests.

However, right now, the tool isn't very useful if you're using a query like AI safety or animal welfare that produce a lot of results. When a search involves canonical EA categories and hundreds of results, I think the value of the current tool over SwapCard or Excel is uncertain.

While the tool reliably returns results closely related to many simple queries, queries that include a number of unrelated words can produce inconsistent results. Entering multiple minor variations of a query is useful e.g. searching for spanish, spanish speakers and spanish outreach can help make sure more attendees with these interests are found.

Possibilities for development

There’s many improvements that could be made:

1. Basic preprocessing. I just passed entire text fields from SwapCard into the embedder, instead of trying to parse sentences or paragraphs. Better preprocessing would improve functionality.
2. Fine tuning or finding better embeddings. An off the shelf embedding was used. Getting data and fine-tuning a model to understand EA semantics and terminology would work better. Even just trying out different "off-the-shelf" embeddings would probably be useful.
3. Basic augmentation. Simple “augmentation” ideas, like using a set of automatically generated similar search terms, would make the tool work more robustly.
4. GUI and installer. Putting this into a GUI and having an installer for non-development users would be great.

The "machine learning" in this tool is just a semantic search that uses 2 lines from Hugging Face, loading a BERT embedding. This is functional, but I’m guessing it’s only about "30% effective" as it could be, and could be greatly improved with little effort.

Comments on configuration

Limited configuration settings has been implemented as global variables. A small number of users might find it useful to change these:

In main.py:

class AttendeeSearch:
    RESULTS_PER_PAGE = 5
    THRESHOLD_COSINE_SCORE = 0.30  # search results with cosine score below this aren't considered

    USE_ANY_XLSX_FILE = True
    ROW_OF_SPREADSHEET_CONTAINING_COLUMN_NAMES = 3

Contact Info

You can reach out to me on Slack or SwapCard. You also find contact information on SwapCard.