This project was done as a summer research scholarship with The University of Auckland. Click on the poster below to open it, or keep reading below.
Introduction
Background
This project is part of the Science for Technological Innovation National Science Challenge. Missing and incomplete records mean that thousands of Māori descendants who are rightful shareholders in ancestral land can not be tracked down. Parininihi ki Waitotara (PKW) is one such Māori incorporation which has lost track of around half of its 10,300 owners. Algorithms to search death certificates, newspapers, and Māori Land Court records have been developed to search for missing shareholders. However, many connections are missed during this search due to family members that have both an English name and a Māori Name. These names are often the result of a borrowing process which modifies English names to follow the structure of the Māori language. In general, this process is called ‘borrowing’, and borrowed words are called ‘loanwords’ Some example borrowing pairs:
These borrowings reveal three things:
- Māori words always end in a vowel sound.
- Māori words do not typically have adjacent consonants.
- There is no one ‘correct’ way of borrowing a word. But what about more complicated borrowings?
Aims
Begin the development of a software package capable of the following:
- Applying general rules to match sounds between English and Māori word pairs.
- Analysing known loanwords to determine rules used to modify borrowed English words.
- Discovering unknown rules used during the borrowing process.
- Determining the likelihood of any pair of English and Māori words being a borrowing.
Methods
IPA Transcription
Many English names are not pronounced as they appear. This makes the matching of sounds between borrowing pairs difficult for a computer. The International Phonetic Alphabet (IPA) is used to solve this problem. IPA is an alphabet where symbols represent sounds used in human speech. IPA was used to transcribe borrowing pairs so that the modelling is sound-based.
Transcribing words with IPA means borrowings can now be more easily matched based on sound composition rather than spelling.
Mapping with English/Māori:
Mapping with IPA:
Phase 1: Rule Discovery
The first phase aims to learn a set of possible mapping rules from public borrowings databases, including those in Ngā Kupu Arotau - eweri tāima: Loanwords in Māori 1842-1952 by Ka’ai, T. & Moorfield, J.. (2009).
Most borrowings have multiple ways sounds can be matched between the words.
In this instance the schwa (ə) is lost in the borrowing. This is due to the argument that schwas are just filler sounds and aren’t actually enounced.
In this case the schwa becomes the second ɐ sound in Arama.
It’s also possible that the appended ɐ sound is actually dependent on the previous vowel and not the previous consonant.
These alternate mappings can be described as different paths through a network where each arc represents a different mapping:
The coloured arcs represent three instances of the same result that could be produced via different rules. These arcs are created by searching for applicable regular expressions (regex) like these:
Phase 2: Optimisation
Phase 2 is currently in progress, and aims to maximise the‘fit’ to the data while using the least amount of rules possible. This is currently being done by assigning each mapping:1. A probability based on its frequency in phase one.2. A penalty value based on how specific the rule is.The sum of shortest path lengths for each borrowing pair,balanced against the number of mappings is used as theobjective function for gradient descent.The aim of this is to remove more specific rules when themappings can be explained just as well by more generalones. However, generating and solving millions of shortest pathproblems is proving to be a computationally expensivemethod so other solutions will be investigated.
Conclusions
Development of the software package has begun, and the following features have been implemented:
- Automated fetching of IPA transcriptions for English words from the Wiktionary.com API.
- Software for generating and solving shortest path problems from a set of general rules.
- Generated list of possible borrowing mappings.
- Partial implementation of rule optimisation routines.
More work is required to improve the set of borrowingrules used to explain known mappings.