A Hmong Corpus with Elaborate Expression Annotations

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4992–5000 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4992 A Hmong Corpus with Elaborate Expression Annotations David R. Mortensen, Xinyu Zhang, Chenxuan Cui, Katherine J. Zhang Carnegie Mellon University 5000 Forbes Ave Pittsburgh, PA 15213 United States of America dmortens@cs.cmu.edu, xinyuzh2@andrew.cmu.edu, cxcui@cs.cmu.edu, kjzhang@cmu.edu Abstract This paper describes the first publicly available corpus of Hmong [ISO 639-3: mww, hmj], a minority language of China, Vietnam, Laos, Thailand, Australia, and various countries in Europe and the Americas. The corpus has been scraped from a long-running Usenet newsgroup called soc.culture.hmong and consists of approximately 12 million tokens. This corpus (called SCH) is also the first substantial corpus to be annotated for elaborate expressions, a kind of four-part coordinate construction that is common and important in the languages of mainland Southeast Asia. We show that word embeddings trained on SCH can benefit tasks in Hmong (solving analogies) and that a model trained on it can label previously unseen elaborate expressions, in context, with an F1 of 90.79 (precision: 87.36, recall: 94.52). Keywords: corpus, Hmong, low-resource, coordination, non-compositional, elaborate expression 1. Introduction Hmong is a language of Southern China and Southeast Asia with about 2.7 million speakers1. A plurality of speakers are located in China, but starting in the eigh- teenth century, a substantial number of Hmong speak- ers migrated to Vietnam, then Laos and Thailand (Culas and Micraud, 1997). Following the Indochinese con- flict in the middle of the 20th century, many Hmong left Laos as refugees and emigrated to Western nations such as the United States, France, and Australia, carrying their language with them. Hmong remains a substan- tially under-resourced language. While there are some technologies for Hmong—for example, Google Trans- late (https://translate.google.com/) ostensibly supports it—there are few publicly available resources for the language. This paper presents the first sizable, publicly available data resource for Hmong: the SCH corpus (https: //github.com/dmort27/sch-corpus/). It is a col- lection of 20 years’ worth of posts from a Hmong- oriented Usenet group cleaned and filtered by language (using a high-precision regular expression classifier). It consists of almost 12 million tokens and should be of value both to researchers interested in producing lan- guage technologies for Hmong and to linguists inter- ested in further documenting the language. 2. The Hmong Language There is some confusion around the term “Hmong.” As used here, it refers to a dialect continuum that in- cludes the Hmong Daw (“White Hmong”) and Mong Leng/Mong Njua (“Green/Blue Hmong”) varieties that 1Here, “Hmong” refers to the set of lects that are mutually intelligible with Hmong Daw [mww] and Mong Leng [hmj] and excludes other Southwestern Hmongic languages. See Section 2 are spoken in Laos and Thailand, as well as the dis- tinct but mutually intelligible varieties that are spoken in Vietnam and China. Sometimes the term has also been used to include all of the members of the Miao nationality in China and this encompasses some …

Content truncated to 500 words