Local Word Discovery for Interactive Transcription

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2058–2067 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2058 Local Word Discovery for Interactive Transcription William Lane and Steven Bird Northern Institute Charles Darwin University Abstract Human expertise and the participation of speech communities are essential factors in the success of technologies for low-resource lan- guages. Accordingly, we propose a new com- putational task which is tuned to the available knowledge and interests in an Indigenous com- munity, and which supports the construction of high quality texts and lexicons. The task is illustrated for Kunwinjku, a morphologically- complex Australian language. We combine a finite state implementation of a published grammar with a partial lexicon, and apply this to a noisy phone representation of the sig- nal. We locate known lexemes in the signal and use the morphological transducer to build these out into hypothetical, morphologically- complex words for human validation. We show that applying a single iteration of this method results in a relative transcription den- sity gain of 17%. Further, we find that 75% of breath groups in the test set receive at least one correct partial or full-word suggestion. 1 Introduction In over a century of practice in descriptive linguis- tics, the pattern has been to prepare texts and a lex- icon to support the construction of a grammar. The grammar includes a description of the phonology and morphosyntax, which inform the representa- tion of the texts and lexicon, in a cyclic arrange- ment (Crowley, 2007, 139f). The three types of data are entwined in the so-called “Boasian trilogy” of texts, lexicon, and grammar. More recently, another tradition of working with little-studied languages has grown up in the lan- guage technology community. It frames these as “low resource languages,” lacking the text, speech and lexical resources that are needed for creating speech and language technologies (Krauwer, 2003). In many cases, these languages are not little-studied at all, it is just that the technological methods can only exploit texts and lexicons, not the grammar. This brings us to the question: how can we lever- age a grammar when working with a low resource language? In particular, how can we leverage a morphosyntactic description to accelerate the cre- ation of texts and a lexicon for a morphologically complex language? Our approach complements the practice of “learning to transcribe” (Bird, 2020), where non- speaker transcribers train themselves to recognize words in connected speech. We assume that tran- scribers are able to sparsely annotate spans of audio with any words they recognize. These words can be aligned with the output of an automatic phone recognizer, and the machine suggests new words conditioned on phones in the locus of known words (Fig. 1). We call this task local word discovery. In the case of low-resource languages like Kun- winjku (ISO gup), we do not have enough text to train a language model to guide the suggestion of words in the locus of previously recognized words. However, as a morphologically-complex …

Content truncated to 500 words