Skip to content

Latest commit

 

History

History
114 lines (93 loc) · 3.61 KB

README.md

File metadata and controls

114 lines (93 loc) · 3.61 KB
Raw

Callsign naive recognition

UP

  • Detecting nearby plane in the transcript
  • API searching for nearby callsigns inside the transcript
  • Author: Juraj Dedič

Mechanism

  • For each airplane of the nearby callsigns list:
    • callsign example: RYR83H
    • sentence example: Ryanair eighty three hotel descend flight level one hundred
    • (Expansion) - It generates possible variations of the callsign in it's spoken form:
      • possible variations example:
        • Romeo Yankee Romeo eight three Hotel
        • Romeo Yankee Romeo eighty three Hotel
        • Ryanair eighty three Hotel
        • Ryanair Romeo eighty three Hotel
      • (Scoring) for each variation it scores it based on windowing algorithm:
        • The algorithm uses a window, which is an N-gram of the transcript.
        • Where N is the number of tokens of the callsign
        • The algorithm finds the maximum score after iterating through each position of the window
      • The most probable variation is returned for each callsign
    • The callsign scores are then compared
    • The best matching callsign is chosen
    • If there are multiple callsigns with the same score, then the longest match will be returned

About the script

  • The way it was meant to work initially was to leverage the n-best hypothesis
  • Somehow using the way I intended did not gave me better results
  • That is why there are functions doing similar things (labeled as deprecated) but for CNET words
  • It was switched to use 1-best CNET sequence
  • It normalizes the input text
  • Also for some reason, I initially implemented the datasample endpoint, which maybe just made things more difficult

Files

  • callsignsExpansion.js is the script responsible for generating the variations for more accurate matching
  • callsignsService.js the main script doing the hard work
  • airlines.js mappings of ICAO callsigns
  • index.js the main JS file which should be run

API

The API accepts content type application/json. It listens for requests on following endpoints:

POST /datasample

  • Used for testing
  • The request body should contain whole datasample described in the ATCO JSON dataset readme
  • returns matches for each segment of the datasample

POST /:threshold?

  • Recommended primary endpoint
  • Accepts the transcript of a single segment
  • The request should contain one segment
  • threshold is an optional parameter (default = 0.6)
  • threshold is for filtering low probability results
  • Request body should have following form:
    {
        "text": "Ryanair eighty three hotel romeo descend flight level one hundred",
        "nearbyCallsigns": [
            "RYR8HR", "RYR93V", "OKFCA", "CSA183"
        ] 
    }
  • returns match object with the most probable variation of the contained callsign:
    {
        "callsign": "RYR8HR",
        "variation": "Ryanair eighty three hotel romeo",
        "p": 1.0
    }
  • If none of the potential callsigns has score higher than the threshold, it returns:
    {}

Installation and Running the script

The following section describes how to properly install the API and run it using Node.js.

Installation

Run one of the following commands inside this directory to install the script:

yarn install

or

npm install

Start the server

To run the server, use one of the following commands:

yarn start

or

npm run-script start

yarn start

Requirements

These applications are needed to successfully install and run the server script:

  • Node.js v18