The main goal of the parser is to translate a label representing a SL gloss, written in HamNoSys format, into a form that can be used for DL-based classification. We used a decision tree to decompose the notation into numerical multilabels for the defined classes. The parser analyzes a series of symbols and matches each symbol with the class that describes it (while assigning it the appropriate number) or removes it.
The Hamburg Sign Language Notation System (HamNoSys) is a phonetic transcription system that has been widespread for more then 20 years. HamNoSys does not refer to different national finger alphabets and can therefore be used internationally. It can be divided into six basic blocks, as presented in the below Figure (upper panel). The first two out of six blocks - symmetry operator and non-manual features - are optional. The remaining four components - handshape, hand position, hand location, and movement - are mandatory. More detailed structure is also presented in the figure (lower panel).
HamNoSys font is required to be installed, to properly visualise provided examples. It can be downloaded directly from DSG Corpus website.
In our implementation, four blocks (symmetry operator, location left/right, location top/bottom, distance from the body) refer to the overall human posture, while five blocks (handshape base form, handshape thumb position, handshape bending, hand position extended finger direction, and hand position palm orientation) relate to a single hand.
List of blocks and their ranges:
- Symmetry operator class consists of 9 symbols (0 to 8).
- Handshape - Baseform class consists of 12 symbols (0 to 11).
- Handshape - Thumb position class consists of 4 symbols (0 to 3).
- Handshape - bending class consists of 6 symbols (0 to 5).
- Handposition - Extended finger direction class consists of 18 symbols (0 to 17).
- Handposition - Palm orientation class consists of 8 symbols (0 to 7).
- Handposition - Left/Right class consists of 5 symbols (0 to 4).
- Handposition - Top/Bottom class consists of 37 symbols (0 to 36).
- Handposition - Distance class consists of 6 symbols (0 to 5).
The figure presents the numerical values and assigned to them HamNoSys symbols.
In its most basic form, parsing new annotation file boils down to:
$ python parse-hamnosys.py -sf <source_file> -df <destination_file> -ef <error_file>
In repositry we prepared some example files, which were produced by running:
$ python parse-hamnosys.py -sf hamnosys_example.txt -df hamnosys_parsed.txt -ef error.txt
Default input file format is defined by the HearAI project requirements. As in hamnosys_example.txt parser requires a file that has 6 columns separated with a space sign " ". Due to this, if any of the columns contains space, it must be removed or replaced (for example with "_" sign) before passing to the parser. Parsers operates only on HamNoSys notation, that is stored in the last (6th) column. It shall start with HamNoSys sign (No quote nor apostrophe sign is allowed). Input file columns and description:
- Name - name of a video file that given notation refers to
- Start - sign start time (on a video)
- End - sign end time (on a vide)
- Dict - word number in a dictionary
- Word - translation to a spoken language
- Hamnosys - Notation
Default estination file consists of following columns separated by the space " " sign:
- Name - name of a video file that given notation refers to, directly copied from source file
- Start - sign start time (on a video), directly copied from source file
- End - sign end time (on a vide), directly copied from source file
- Symmetry operator - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- NonDom first - Used when notation starts with sign
- Dominant - Handshape - Baseform - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Dominant - Handshape - Thumb position - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Dominant - Handshape - bending - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Dominant - Handposition - extended finger direction - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Dominant - Handposition - palm orientation - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Dominant - Handshape - Baseform2
- Dominant - Handshape - Thumb position2
- Dominant - Handshape - Bending2
- Dominant - Handposition - Extended finger direction2
- Dominant - Handposition - Palm orientation2
- NONDominant - Handshape - Baseform
- NONDominant - Handshape - Thumb position
- NONDominant - Handshape - Bending
- NONDominant - Handposition - Extended finger direction
- NONDominant - Handposition - Palm orientation
- NONDominant - Handshape - Baseform2
- NONDominant - Handshape - Thumb position2
- NONDominant - Handshape - Bending2
- NONDominant - Handposition - Extended finger direction2
- NONDominant - Handposition - Palm orientation2
- Handposition - LR - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Handposition - TB - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
- Handposition - Distance - Number that represents one of the classes (please refer to hamnosys_dicts.txt), parsed from notation
If you find this code useful in your research, please consider citing:
@misc{majchrowska2022hamnosys,
doi = {10.48550/ARXIV.2204.06924},
url = {https://arxiv.org/abs/2204.06924},
author = {Majchrowska, Sylwia and Plantykow, Marta and Olech, Milena},
title = {Handling sign language transcription system with the computer-friendly numerical multilabels},
publisher = {arXiv},
year = {2022},
copyright = {arXiv.org perpetual, non-exclusive license}
}
The result of the work was presented at Communication in Human-AI Interaction Workshop @ IJCAI 2022.
File hamnosys_example.txt
was created using Korpusowy słownik polskiego języka migowego
Joanna Łacheta, Małgorzata Czajkowska-Kisil, Jadwiga Linde-Usiekniewicz, Paweł Rutkowski (red.), 2016, Korpusowy słownik polskiego języka migowego, Warszawa: Wydział Polonistyki Uniwersytetu Warszawskiego, ISBN: 978-83-64111-49-5 (publikacja online).