Abstract

Pre-training a language model and then fine-tuning it for downstream tasks has demonstrated state-of-the-art results for various NLP tasks. Pre-training is usually independent of the downstream task, and previous works have shown that this pre-training alone might not be sufficient to capture the task-specific nuances. We propose a way to tailor a pre-trained BERT model for the downstream task via task-specific masking before the standard supervised fine-tuning. For this, a word list is first collected specific to the task. For example, if the task is sentiment classification, we collect a small sample of words representing both positive and negative sentiments. Next, a word’s importance for the task, called the word’s task score, is measured using the word list. Each word is then assigned a probability of masking based on its task score. We experiment with different masking functions that assign the probability of masking based on the word’s task score. The BERT model is further trained on MLM objective, where masking is done using the above strategy. Following this standard supervised fine-tuning is done for different downstream tasks. Results on these tasks show that the selective masking strategy outperforms random masking, indicating its effectiveness.

Abstract

Due to the lack of a large annotated corpus, many resource-poor Indian languages struggle to reap the benefits of recent deep feature representations in Natural Language Processing (NLP). Moreover, adopting existing language models trained on large English corpora for Indian languages is often limited by data availability, rich morphological variation, syntax, and semantic differences. In this paper, we explore the traditional to recent efficient representations to overcome the challenges of a low resource language, Telugu. In particular, our main objective is to mitigate the low-resource problem for Telugu. Overall, we present several contribu- tions to a resource-poor language viz. Telugu. (i) a large annotated data (35,142 sentences in each task) for multiple NLP tasks such as sentiment analysis, emotion identification, hate-speech detection, and sarcasm detection, (ii) we create different lexicons for sentiment, emotion, and hate-speech for improving the effi- ciency of the models, (iii) pretrained word and sentence embeddings, and (iv) different pretrained language models for Telugu such as ELMo-Te, BERT-Te, RoBERTa-Te, ALBERT-Te, and DistilBERT-Te on a large Telugu corpus consisting of 8,015,588 sentences (1,637,408 sentences from Telugu Wikipedia and 6,378,180 sentences crawled from different Telugu websites). Further, we show that these representations significantly improve the performance of four NLP tasks and present the benchmark results for Telugu. We argue that our pretrained embeddings are competitive or better than the existing multilingual pretrained models: mBERT, XLM-R, and IndicBERT. Lastly, the fine-tuning of pretrained models show higher performance than linear probing results on four NLP tasks with the following F1-scores: Sentiment (68.72), Emotion (58.04), Hate-Speech (64.27), and Sarcasm (77.93). We also experiment on publicly available Telugu datasets (Named Entity Recognition, Article Genre Classification, and Sentiment Analysis) and find that our Telugu pretrained language models (BERT-Te and RoBERTa-Te) outperform the state-of-the-art system except for the sentiment task. We open-source our corpus, four different datasets, lexicons, embeddings, and code https://github.com/Cha14ran/DREAM-T. The pretrained Transformer models for Telugu are available at https://huggingface.co/ltrctelug

Abstract

Document summarization aims to create a precise and coherent summary of a text document. Many deep learning summarization models are developed mainly for English, often requiring a large training corpus and efficient pre-trained language models and tools. However, English summarization models for low-resource Indian languages are often limited by rich morphological variation, syntax, and semantic differences. In this paper, we propose GAE-ISUMM, an unsupervised Indic summarization model that extracts summaries from text documents. In particular, our proposed model, GAE-ISUMM uses Graph Autoencoder (GAE) to learn text representations and a document summary jointly. We also provide a manually-annotated Telugu summarization dataset TELSUM, to experiment with our model GAE-ISUMM. Further, we experiment with the most publicly available Indian language summarization datasets to investigate the effectiveness of GAE-ISUMM on other Indian languages. Our experiments of GAE-ISUMM in seven languages make the following observations: (i) it is competitive or better than state-ofthe-art results on all datasets, (ii) it reports benchmark results on TELSUM, and (iii) the inclusion of positional and cluster information in the proposed model improved the performance of summaries. We open-source our dataset and code 1 . Index Terms—component, formatting, style, styling, insert

Abstract

We present the Hindi-Telugu Parallel Corpus of different technical domains such as Natural Science, Computer Science, Law and Healthcare along with the General domain. The qualitative corpus consists of 700K parallel sentences of which 535K sentences were created using multiple methods such as extract, align and review of Hindi-Telugu corpora, end-to-end human translation, iterative back-translation driven post-editing and around 165K parallel sentences were collected from available sources in the public domain. We present the comparative assessment of created parallel corpora for representativeness and diversity. The corpus has been pre-processed for machine translation, and we trained a neural machine translation system using it and report state-of-the-art baseline results on the developed development set over multiple domains and on available benchmarks. With this, we define a new task on Domain Machine Translation for low resource language pairs such as Hindi and Telugu. The developed corpus (535K) is freely available for non-commercial research and to the best of our knowledge, this is the well curated, largest, publicly available domain parallel corpus for Hindi-Telugu.

Abstract

Word Problem Solving remains a challenging and interesting task in NLP. A lot of research has been carried out to solve different genres of word problems with various complexity levels in recent years. However, most of the publicly available datasets and work has been carried out for English. Recently there has been a surge in this area of word problem solving in Chinese with the creation of large benchmark datastes. Apart from these two languages, labeled benchmark datasets for low resource languages are very scarce. This is the first attempt to address this issue for any Indian Language, especially Hindi. In this paper, we present HAWP (Hindi Arithmetic Word Problems), a dataset consisting of 2336 arithmetic word problems in Hindi. We also developed baseline systems for solving these word problems. We also propose a new evaluation technique for word problem solvers taking equation equivalence into account.

Abstract

Cross-lingual dubbing of lecture videos requires the transcription of the original audio, correction and removal of disfluencies, domain term discovery, text-to-text translation into the target language, chunking of text using target language rhythm, text-to-speech synthesis followed by isochronous lipsyncing to the original video. This task becomes challenging when the source and target languages belong to different language families, resulting in differences in generated audio duration. This is further compounded by the original speaker’s rhythm, especially for extempore speech. This paper describes the challenges in regenerating English lecture videos in Indian languages semi-automatically. A prototype is developed for dubbing lectures into 9 Indian languages. A mean-opinion-score (MOS) is obtained for two languages, Hindi and Tamil, on two different courses. The output video is compared with the original video in terms of MOS (1-5) and lip synchronisation with scores of 4.09 and 3.74, respectively. The human effort also reduces by 75%.

Abstract

In computer assisted pronunciation training (CAPT), robust automatic models are critical for pronunciation assessment and mispronunciation detection and diagnosis (MDD). In the modelling, besides the audio data of second language (L2) learners, CAPT requires manually annotated ratings of over- all pronunciation quality, and the MDD uses manually anno- tated phonetic transcriptions. Though the pronunciation qual- ity and the mispronunciation are interdependent, to the best of our knowledge, none of the existing corpora contains both ratings and phonetic transcriptions. This could be due to the cost involved in obtaining phonetic transcriptions. However, a corpus with both kinds of information could benefit the re- searchers to obtain robust models by exploring the interde- pendencies. For addressing this, we develop voisTUTOR 2.0 corpus considering the existing voisTUTOR corpus referred to as voisTUTOR 1.0. We obtain phonetic transcriptions man- ually from a linguist for the entire Indian L2 learners’ En- glish audio data (26529 utterances approximately 14 hours) in voisTUTOR 1.0 for which overall quality ratings and bi- nary scores of factors influencing the pronunciation quality are available. A preliminary analysis of voisTUTOR 2.0 sug- gests that the phonetic errors correlated with the ratings and the binar

Abstract

Previous research has found that Acoustic Models (AM) of an Automatic Speech Recognition (ASR) system are susceptible to dialect variations within a language, thereby adversely affecting the ASR. To counter this, researchers have proposed to build a dialectspecific AM while keeping the Language Model (LM) constant for all the dialects. This study explores the effect of dialect mismatched LM by considering three different Telugu regional dialects: Telangana, Coastal Andhra, and Rayalaseema. We show that dialect variations that surface in the form of a different lexicon, grammar, and occasionally semantics can significantly degrade the performance of the LM under mismatched conditions. Therefore, this degradation has an adverse effect on the ASR even when dialect-specific AM is used. We show a degradation of up to 13.13 perplexity points when LM is used under mismatched conditions. Furthermore, we show a degradation of over 9% and over 15% in Character Error Rate (CER) and Word Error Rate (WER), respectively, in the ASR systems when using mismatched LMs over matched LMs

Abstract

Alzheimer’s Dementia is a progressive neurological disorder characterized by cognitive impairment. It affects mem- ory, thinking skills, language, and the ability to perform simple tasks. Detection of Alzheimer’s Dementia from the speech is considered a primitive task, as most speech cues are preserved in it. Studies in the literature focused mainly on the lexical features and few acoustic features for detecting Alzheimer’s disease. The present work explores the single frequency filtering cepstral coefficients (SFCC) for the automatic detection of Alzheimer’s disease. In contrast to STFTs, the proposed feature has better temporal and spectral resolution and captures the transient part more appropriately. This offers a very compact and efficient way to derive the formant structure in the speech signal. The experiments were conducted on the ADReSSo dataset, using the support vector machine classifier. The classification performance was compared with several baseline features like Mel-frequency cepstral coefficients (MFCC), perceptual linear prediction (PLP), linear prediction cepstral coefficient (LPCC), Mel frequency cepstral coefficients of LP-residual (MFCC-WR), ZFF signal (MFCC-ZF) and eGeMAPS (openSMILE). The experiments conducted on Alzheimer’s Dementia classification task show that the proposed feature performs better than conventional MFCCs. Among all the features, SFCC offers the best classification accuracy of 65.1% and 60.6% for dementia detection on cross- validation and test data, respectively. The combination of baseline features with SFCC features further improved the performance. Index Terms—Alzheimer’s disease, Cognitive impairment, Sin- gle frequency filtering cepstral coefficients.

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative dis- order of the central nervous system identified by motor and non- motor activities abnormalities. PD affects respiration, laryngeal, articulation, resonance, and prosodic aspects of speech production. Detection of PD from speech is a non-invasive approach useful for automatic screening. Perceptual attributes of speech due to PD are manifested as temporal variations in speech. In this regard, current work investigated the use of LSTM and BiLSTM networks with shifted delta cepstral (SDC) features to detect PD from speech. Further in BiLSTM networks, a multi-head attention mechanism is introduced, assuming that each head captures distinct information to detect PD. SDC features obtained from MFCCs, and SFFCCs are used for developing the PD detection system. The performance of the experiments is validated using the PC-GITA database. The experimental results revealed that BiLSTM networks give a rela- tive improvement of 4-5% over the LSTM networks. The use of a multi-head attention mechanism further improved the detection accuracy of the PD detection system, showing that it can capture various discriminative features.