Generative Large Language Models (LLMs) have recently achieved remarkable advancements in translation tasks. However, low-resource or unseen language words are underrepresented in the LLM vocabulary where words are frequently constructed from subwords that may not align with their morphological structures. In this study, we present a novel two-stage morphological knowledge injection strategy into LLMs to address the difficulties presented by low-resource, agglutinative languages focusing on English-Manipuri translation as a case study. The first stage involves extending the LLM vocabulary with morphemes and segmenting Manipuri words into fine-grained morphological units that preserve morpheme boundaries in the input representation. The second stage introduces a chain-of-morphemes (CoM) tuning approach that divides the translation process into two steps: first, generating word roots for semantic translation, and then applying morphological inflection. This methodology decouples semantic and morphological processing, improving translation quality. Evaluation on the WMT 2023 English-Manipuri dataset using mGPT and BLOOM models demonstrates that our approach outperforms existing baseline methods. Our findings emphasize the potential advantages of incorporating morphological knowledge in LLM-based machine translation for morphologically rich languages.

Agglutinative languages often have morphologically complex words (MCWs) composed of multiple morphemes arranged in a hierarchical structure, posing significant challenges in translation tasks. We present a novel Knowledge Distillation approach tailored for improving the translation of such languages. Our method involves an encoder, a forward decoder, and two auxiliary decoders: a backward decoder and a morphological decoder. The forward decoder generates target morphemes autoregressively and is augmented by distilling knowledge from the auxiliary decoders. The backward decoder incorporates future context, while the morphological decoder integrates target-side morphological information. We have also designed a reliability estimation method to selectively distill only the reliable knowledge from these auxiliary decoders. Our approach relies on morphological word segmentation. We show that the word segmentation method based on unsupervised morphology learning outperforms the commonly used Byte Pair Encoding method on highly agglutinative languages in translation tasks. Our experiments conducted on English-Tamil, English-Manipuri, and English-Marathi datasets show that our proposed approach achieves significant improvements over strong Transformer-based NMT baselines.

This paper presents a novel back transliteration dataset capturing native language text originally composed in the Roman/Latin script, harvested from popular social media platforms, along with its corresponding representation in the native Assamese script. Assamese, categorized as a low-resource language within the Indo-Aryan language family, predominantly spoken in the north-east Indian state of Assam, faces a scarcity of linguistic resources. The dataset comprises a total of 60,312 Roman-native parallel transliterated sentences. This paper diverges from conventional forward transliteration datasets consisting mainly of named entities and technical terms, instead presenting a novel transliteration dataset cultivated from three prominent social media platforms, Facebook, Twitter(currently X), and YouTube, in the backward transliteration direction. The paper offers a comprehensive examination of ten state-of-the-art word-level transliteration models within the context of this dataset, encompassing transliteration evaluation benchmarks, extensive performance assessments, and a discussion of the unique challenges encountered during the processing of transliterated social media content. Our approach involves the initial use of two statistical transliteration models, followed by the training of two state-of-the-art neural network-based transliteration models, evaluation of three publicly available pre-trained models, and ultimately fine-tuning one existing state-of-the-art multilingual transliteration model along with two pre-trained large language models using the collected datasets. Notably, the Neural Transformer model outperforms all other baseline transliteration models, achieving the lowest Word Error Rate (WER) and Character Error Rate (CER), and the highest BLEU (up to 4 gram) score of 55.05, 19.44, and 69.15, respectively.

Word embeddings and Language models are the building blocks of modern Deep Neural Network-based Natural Language Processing. They are extensively explored in high-resource languages and provide state-of-the-art (SOTA) performance for a wide range of downstream tasks. Nevertheless, these word embeddings are not explored in languages such as Assamese, where resources are limited. Furthermore, there has been limited study into the performance evaluation of these word embeddings for low-resource languages in downstream tasks. In this research, we explore the current state of Assamese pre-trained word embeddings. We evaluate these embeddings’ performance on sequence labeling tasks such as Parts-of-speech and Named Entity Recognition. In order to assess the efficiency of the embeddings, experiments are performed utilizing both ensemble and individual word embedding approaches. The ensembling approach that uses three word embeddings outperforms the others. In the paper, the outcomes of the investigations are described. The results of this comparative performance evaluation may assist researchers in choosing an Assamese pre-trained word embedding for subsequent tasks.

This article aims to understand different transliteration behaviors of Romanized Assamese text on social media. Assamese, a language that belongs to the Indo-Aryan language family, is also among the 22 scheduled languages in India. With the increasing popularity of social media in India and also the common use of the English Qwerty keyboard, Indian users on social media express themselves in their native languages, but using the Roman/Latin script. Unlike some other popular South Asian languages (say Pinyin for Chinese), Indian languages do not have a common standard romanization convention for writing on social media platforms. Assamese and English are two very different orthographical languages. Thus, considering both orthographic and phonemic characteristics of the language, this study tries to explain how Assamese vowels, vowel diacritics, and consonants are represented in Roman transliterated form. From a dataset of romanized Assamese social media texts collected from three popular social media sites: (Facebook, YouTube, and X (formerly known as Twitter)), 1 we have manually labeled them with their native Assamese script. A comparison analysis is also carried out between the transliterated Assamese social media texts with six different Assamese romanization schemes that reflect how Assamese users on social media do not adhere to any fixed romanization scheme. We have built three separate character-level transliteration models from our dataset. One using a traditional phrase-based statistical machine transliteration model, (1) PBSMT model and two separate neural transliteration models, (2) BiLSTM neural seq2seq model with attention, and (3) Neural transformer model. A thorough error analysis has been performed on the transliteration result obtained from the three state-of-the-art models mentioned above. This may help to build a more robust machine transliteration system for the Assamese social media domain in the future. Finally, an attention analysis experiment is also carried out with the help of attention weight scores taken from the character-level BiLSTM neural seq2seq transliteration model built from our dataset.

The objective of the current study is to explore a quantitative frequency domain technique to evaluate rhythm in spontaneous speech data of 19 native speakers of Mising and Assamese, two low-resourced languages spoken in Assam, North-East India. The concept of analyzing speech rhythm using amplitude modulation (AM) low-frequency (LF) spectrum, also known as rhythm formant analysis (RFA), is initially put forth by Gibbon and Li [ 17 ]. We propose three features from rhythm formants of the LF spectrum and also explore discrete cosine transform (DCT)–based characterization of the retrieved LF spectrum. We aim to distinguish the rhythm of Assamese and two Mising dialects, namely Pagro and Delu, with the aid of machine learning techniques fed with the derived features as input. We have observed that the features are efficient in classifying Assamese vs Pagro and Assamese vs Delu with an accuracy of 92.73% and 91.15%, respectively. The experimental analysis further reveals that Assamese is rhythmically closer to Delu than Pagro.

Ensemble system for part-of-speech (POS) tagging is beneficial for many resource-poor languages that do not have enough annotated training data to train Deep Learning (DL, also named Deep Neural Network)-based POS taggers. An Ensemble system is a better choice to incorporate the linguistic features of a language and leverage the benefits of various types of POS taggers. In this work, we present our experiment of developing an ensemble tagger for Assamese, a low-resource, morphologically rich scheduled language of India, spoken by more than 15 million people. Despite the success of modern neural-network-based models in sequence tagging tasks, it has yet to receive attention in developing tasks such as POS in a resource-poor language such as Assamese. We develop a POS tagging model based on the BiLSTM-CRF architecture with a corpus of 404k tokens. We cover several word embeddings during training. Among all the experiments, the top two POS tagging models achieve tagging F1 scores of 0.746 and 0.745. We observe that the DL-based taggers are not able to achieve decent accuracy. It may be due to the inability to capture the linguistic features of the language or due to comparatively less annotated data. So, we build another POS tagger using a rule-based approach considering several morphological phenomena of the language and get an F1 score of 0.85. Subsequently, we integrate the top two DL-based taggers with the rule-based ones and develop a new POS tagger using an ensemble approach, of which we get an improved F1 score of 0.925. Performance improvement of our new ensemble POS taggers over the baseline taggers suggests that integration of the taggers combines the qualities of all taggers in the new tagger. Therefore, this study also states ensemble taggers are more suitable for highly inflectional, morphologically rich resource-poor languages.

Part of Speech (POS) tagging is crucial to Natural Language Processing (NLP). It is a well-studied topic in several resource-rich languages. However, the development of computational linguistic resources is still in its infancy despite the existence of numerous languages that are historically and literary rich. Assamese, an Indian scheduled language, spoken by more than 25 million people, falls under this category. In this paper, we present a Deep Learning (DL)-based POS tagger for Assamese. The development process is divided into two stages. In the first phase, several pretrained word embeddings are employed to train several tagging models. This allows us to evaluate the performance of the word embeddings in the POS tagging task. The top-performing model from the first phase is employed to annotate another set of new sentences. In the second phase, the model is trained further using the fresh dataset. Finally, we attain a tagging accuracy of 86.52 in F1 score. The model may serve as a baseline for further study on DL-based Assamese POS tagging.

We present the AsNER, a named entity annotation dataset for low resource Assamese language with a baseline Assamese NER model. The dataset contains about 99k tokens comprised of text from the speech of the Prime Minister of India and Assamese play. It also contains person names, location names and addresses. The proposed NER dataset is likely to be a significant resource for deep neural based Assamese language processing. We benchmark the dataset by training NER models and evaluating using state-of-the-art architectures for supervised named entity recognition (NER) such as Fasttext, BERT, XLM-R, FLAIR, MuRIL etc. We implement several baseline approaches with state-of-the-art sequence tagging Bi-LSTM-CRF architecture. The highest F1-score among all baselines achieves an accuracy of 80.69% when using MuRIL as a word embedding method. The annotated dataset and the top performing model are made publicly available.

Reduplication is a productive morphological process widely used in a substantial number of languages in the world. Reduplication is a well-studied phenomenon, and several typological works have provided evidence for different types of reduplication in most of the languages around the world. Addressing reduplication plays a vital role in the efficiency of POS tagger, sentiment analysis, as well as other NLP tasks. However, it is an understudied area in computational linguistics, especially in low-resource languages like Assamese. This article first describes different types of reduplication and their shapes that occur in Assamese. Second, an exhaustive set of reduplication formation rules is compiled that is incorporated to build a system to identify reduplication in Assamese text. The results of the experiments performed on three different domain datasets showed that the rule-based system can identify reduplicated expressions with an average precision, recall, and F1 scores of 94.19%, 98.07%, and 96.07%, respectively. Third, it is shown that the Assamese reduplication processes can be captured through a two-way finite-state transducer (2-way FST). Finally, two broad categories of reduplicative processes along with their corresponding 2-way FST model are presented.

WordNets built for low-resource languages, such as Assamese, often use the expansion methodology. This may result in missing lexical entries and missing synonymy relations. As the Assamese WordNet is also built using the expansion method, using the Hindi WordNet, it also has missing synonymy relations. As WordNets can be visualized as a network of unique words connected by synonymy relations, link prediction in complex network analysis is an effective way of predicting missing relations in a network. Hence, to predict the missing synonyms in the Assamese WordNet, link prediction methods were used in the current work that proved effective. It is also observed that for discovering missing relations in the Assamese WordNet, simple local proximity-based methods might be more effective as compared to global and complex supervised models using network embedding. Further, it is noticed that though a set of retrieved words are not synonyms per se, they are semantically related to the target word and may be categorized as semantic cohorts.

Mizo is an under-studied Tibeto-Burman tonal language of the North-East India. Preliminary research findings have confirmed that four distinct tones of Mizo (High, Low, Rising and Falling) appear in the language. In this work, an attempt is made to automatically recognize four phonological tones in Mizo distinctively using acoustic-prosodic parameters as features. Six features computed from Fundamental Frequency (F0) contours are considered and two classifier models based on Support Vector Machine (SVM) & Deep Neural Network (DNN) are implemented for automatic tonerecognition task respectively. The Mizo database consists of 31950 iterations of the four Mizo tones, collected from 19 speakers using trisyllabic phrases. A four-way classification of tones is attempted with a balanced (equal number of iterations per tone category) dataset for each tone of Mizo. it is observed that the DNN based classifier shows comparable performance in correctly recognizing four phonological Mizo tones as of the SVM based classifier.

This paper focuses on the phonetic analysis of Korean and Rabha fricatives and Angami nasals. Though aspirated consonants have been studied earlier, very few studies were found for the comparative study of aspirated fricatives and aspirated nasals. Previous literature has suggested the presence of aspirated fricatives. As there are limited studies on aspirated nasals, this paper tries to investigate the properties of aspirated nasals by comparing them with the aspiration in Korean aspirated fricative and analyses the feature that might distinguish between the aspirated and unaspirated counterparts of both the consonants. Features such as Intensity, Duration, Centre of Gravity (COG), F1 onset and Spectral tilt (H1-H2) are used to investigate whether there is a distinction between the aspirated and unaspirated fricatives and nasals. Results confirm that COG is a distinctive acoustic cue to discriminate the aspirated and unaspirated counterparts.