Harnessing Cross-lingual Features to Improve Cognate Detection for Low-resource Languages

Abstract

Cognates are variants of the same lexical form across different languages; for example "fonema" in Spanish and "phoneme" in English are cognates, both of which mean "a unit of sound". The task of automatic detection of cognates among any two languages can help downstream NLP tasks such as Cross-lingual Information Retrieval, Computational Phylogenetics, and Machine Translation. In this paper, we demonstrate the use of cross-lingual word embeddings for detecting cognates among fourteen Indian Languages. Our approach introduces the use of context from a knowledge graph to generate improved feature representations for cognate detection. We then evaluate the impact of our cognate detection mechanism on neural machine translation (NMT), as a downstream task. We evaluate our methods to detect cognates on a challenging dataset of twelve Indian languages, namely, Sanskrit, Hindi, Assamese, Oriya, Kannada, Gujarati, Tamil, Telugu, Punjabi, Bengali, Marathi, and Malayalam. Additionally, we create evaluation datasets for two more Indian languages, Konkani and Nepali 1 . We observe an improvement of up to 18% points, in terms of F-score, for cognate detection. Furthermore, we observe that cognates extracted using our method help improve NMT quality by up to 2.76 BLEU. We also release 2 our code, newly constructed datasets and cross-lingual models publicly.

Weakly-supervised Deep Cognate Detection Framework for Low-Resourced Languages Using Morphological Knowledge of Closely-Related Languages

“…The extensive experiments (in Section 6) on three different cognate detection datasets across language families have showcased the efficacy of our weakly-supervised and supervised cognate detection framework. For example, on six different Indian language pairs, our weakly-supervised model (with morphological knowledge) has outperformed the state-of-the-art supervised model proposed by Kanojia et al (2020a) by an average of 9 points of F -score whereas, for Celtic language pairs, it outperformed by 8.6 points of F -score. At the same time, our supervised framework has produced a state-of-the-art performance by outperforming the existing supervised model by an average of 16 points of F -score.…”

“…Kanojia et al (2019a) performed a cognate detection task on Indian languages, which includes a large amount of manual intervention during identification. Kanojia et al (2019b) introduced a character sequence-based recurrent neural network for identifying cognates between Indian language pairs.…”

“…Nevertheless, there are word pairs which are false friends or partial cognates (Kanojia et al, 2020b). Partial cognates are similar words across languages but carry different meanings in different contexts (Kanojia et al, 2019b), thus making automatic cognate detection hard and challenging. Identifying these cognates requires extensive linguistic knowledge across languages, which is quite hard and expensive to annotate.…”

“…Supervised cognate detection frameworks understand the relationship between word pairs by concentrating on their phonetic or lexical similar-ities based on their annotated positive or negative labels (Jäger et al, 2017;Rama, 2016). Recently, researchers tried to exploit contextual multilingual word embedding techniques to identify cognates which produced better results than only concentrating on phonetic transcriptions (Kanojia et al, 2020a). Although such methods had good results, annotating labels is quite expensive and tedious for many under-resourced languages.…”

A large and evolving cognate database

“…In historical linguistics, cognate identification methods are mainly based on three types of similarity measures: semantic, phonetic, and orthographic. For information on semantic similarity, special-purpose multilingual dictionaries, such as the wellknown Swadesh List, are used, and more recently cross-lingual word embeddings (Kanojia et al, 2020(Kanojia et al, , 2021. For orthographic similarity, string metrics (Hauer & Kondrak, 2011;St Arnaud et al, 2017) are often employed, e.g.…”