Benefits of Data Augmentation for NMT-based Text Normalization of User-Generated Content

Marathi is spoken across Maharashtra and nearby regions in many local forms such as Varhadi, Puneri, Kolhapuri, Marathwada and coastal varieties like Malvani and Konkani influenced Marathi. These dialects differ in pronunciation, vocabulary and sometimes grammar, while most digital systems still expect clean Standard Marathi or Hindi text. When users speak or write in their natural dialect, systems such as educational portals, government websites and chatbots may fail to understand the input or return poor quality output. This paper describes a small but complete framework for handling such cases. The proposed Indic Language Dialect Detector accepts both text and speech in selected Marathi regional dialects and produces normalized text in Standard Marathi and, optionally, in Hindi. For text input, the system fine-tunes a multilingual BERT-style encoder on pairs of dialect sentences and their normalized versions and uses this representation for both dialect classification and text normalization. For speech input, a wav2vec-style automatic speech recognition (ASR) model first converts audio to text, which is then passed through the same BERT-based module. A simple web interface connects these components and lets users type text or record audio and see the dialect label and normalized output. The system is evaluated on a small curated dataset collected from speakers of five dialects. We report dialect classification accuracy, normalization quality using sequence-level metrics, and qualitative examples. Although the dataset is limited, results suggest that combining modern text and speech models with basic rule-based handling is a practical way to support dialect users in low-resource Indian language settings.

Social media has become a rich data source for natural language processing tasks with its worldwide use; however, it is hard to process social media data due to its informal nature. Text normalization is the task of transforming the noisy text into its canonical form. It generally serves as a preprocessing task in other NLP tasks that are applied to noisy text. In this study, we apply two approaches for Turkish text normalization: Contextual Normalization approach using distributed representations of words and Sequence-to-Sequence Normalization approach using neural encoder-decoder models. As the approaches applied to Turkish and also other languages are mostly rule-based, additional rules are required to be added to the normalization model in order to detect new error patterns arising from the change of the language use in social media. In contrast to rule-based approaches, the proposed approaches provide the advantage of normalizing different error patterns that change over time by training with a new dataset and updating the normalization model. Therefore, the proposed methods provide a solution to language change dependency in social media by updating the normalization model without defining new rules.

Text normalization and sanitization are intrinsic components of Natural Language Inferences. In Information Retrieval or Dialogue Generation, normalization of user queries or utterances enhances linguistic understanding by translating non-canonical text to its canonical form, on which many state-of-the-art language models are trained. On the other hand, text sanitization removes sensitive information to guarantee user privacy and anonymity. Existing approaches to normalization and sanitization mainly rely on hand-crafted heuristics and syntactic features of individual tokens while disregarding the linguistic context. Moreover, such context-unaware solutions cannot dynamically determine whether out-of-vocab tokens are misspelt or are entity names. In this work, we formulate text normalization and sanitization as a multi-task text generation approach and propose a neural hybrid pointer-generator network based on multi-head attention. Its generator effectively captures linguistic context during normalization and sanitization while its pointer dynamically preserves the entities that are generally missing in the vocabulary. Experiments show that our generation approach outperforms both token-based text normalization and sanitization, while the hybrid pointer-generator improves the generator-only baseline in terms of BLEU4 score, and classical attentional pointer networks in terms of pointing accuracy.

TagNorm: Tagalog Text Normalization with Slang Word Detection and Classification of Type of Slang Word using Support Vector Machine was designed to normalize informal word and transform it to their canonical forms. The study aimed to improve text normalization in the Filipino domain in which the informal texts such as Alphanumeric Abbreviation and Phonetic, Stylistic variation, Combined Words, Misspelled words/ typographical errors (except rumbled letters), Omission of Vowels and Slang Words can decrease the performance of Natural Language Processing. Using the Facebook data the accuracy of the system in terms of informal text types the researchers got an accuracy of 91.23% in Alphanumeric Abbreviation and Phonetic, 77.61% in Repetition of Characters, 85.19% in Stylistic variation, 69.64% in Combined Words, 72.41% in Misspelled words/ typographical errors, 90.7S% in Omission of Vowels and 93.89% in Slang Words, while the data from previous study obtained an accuracy of 86.96% in Alphanumeric Abbreviation and Phonetic, 66.67% in Repetition of Characters, 57.14% in Stylistic variation, 80% in Combined Words, 85.29% in Misspelled words/ typographical errors, 93.15 in Omission of Vowels and 75% in Slang Words..

The explosive increase in Internet usage has attracted technologies for automatically mining the user-generated contents (UGC) from Web documents. These UGC-rich resources have raised new opportunities and challenges to carry out the opinion extraction and mining tasks for opinion summaries. The technology of opinion extraction allows users to retrieve and analyze people’s opinions scattered over Web documents. Opinion mining is a process which is concerned with the opinions generated by the consumers about the product. Opinion Mining aims at understanding, extraction and classification of opinions scattered in unstructured text of online resources. The search engines performs well when one wants to know about any product before purchase, but the filtering and analysis of search results often complex and time-consuming. This generated the need of intelligent technologies which could process these unstructured online text documents through automatic classification, concept recognition, text summarization, etc. These tools are based on traditional natural language techniques, statistical analysis, and machine learning techniques. Automatic knowledge extraction over large text collections like Internet has been a challenging task due to many constraints such as needs of large annotated training data, requirement of extensive manual processing of data, and huge amount of domain-specific terms. Ambient Intelligence (AmI) in wed-enabled technologies supports and promotes the intelligent e-commerce services to enable the provision of personalized, self-configurable, and intuitive applications for facilitating UGC knowledge for buying confidence. In this chapter, we will discuss various approaches of Opinion Mining which combines Ambient Intelligence, Natural Language Processing and Machine Learning methods based on textual and grammatical clues.

For the longest time, translation was a labor-intensive process that required just human effort. While human translation remains the most reliable method of textual content translation, it takes longer and is more expensive if done for each individual piece of information. Several Machine Translation (MT) approaches have recently emerged to facilitate the migration of any content across languages, especially for the low resource languages such as the Arabic Language. Given that Arabic is one of the world's most widely spoken languages, the task of Arabic machine translation has recently gotten a lot of interest from the scientific community. Indeed, the amount of study devoted to these low resources languages has resulted in some significant accomplishments; however, the status of Arabic MT systems falls short of the quality obtained for other languages. As a result, this survey examines the origins and main development timeline of MT approaches, investigates the significant branches, and categorizes different study orientations. In addition, it gives a comprehensive overview of the key research works that have been completed in the field of Arabic Neural MT (ANMT) and discusses possible future research prospects in this discipline.

In Web2.0 era, social media platforms are bearing huge customer base and excessively abundant information resources. On one hand, information consumers spend a lot of time in information search. On the other hand, information providers are seeking effective methods to recognize potential customers, push targeting advertising and provide personalized information services. Generally, mining user-generated content (UGC) to discover user preferences becomes the main channel for user modeling and customer partitioning. However, on social media platforms, user preferences were often manifested in the user-defined tags, online social behaviors as well as the UGC texts. The paper proposed a social-media user partitioning model based on heterogeneous information fusion and ensemble clustering. In the model, online social behaviors and user-defined interest tags are combined with UGC texts respectively to generate basic partitions of social media users. Then, basic partitions are fused into a consensus partition based on the voting mechanism for the final user partitioning. Experiments on real world data sets demonstrate the effectiveness of the proposed model.

The rise of social media has led to vast amounts of user-generated content, with emotions ranging from joy to anger. Negative comments often target individuals, communities, or brands, prompting successful efforts to detect harmful speech such as hate speech, cyberbullying, and abuse. Recently, another type of speech referred to as ‘Hope Speech’ has gained attention from the research community. Hope speech consists of positive affirmations or words of reassurance, encouragement, consolation or motivation offered to the affected individual/ community during the lean periods of life. However, there has been relatively less research focused on the detection of hope speech, more particularly in low-resource languages. This paper, therefore, attempts to develop an ensemble model for detecting hope speech in some low-resource languages. Data for four different languages, namely English, Kannada, Malayalam and Tamil are obtained and experimented with different deep learning-based models. An ensemble model is proposed to combine the advantages of the better performing models. Experimental results demonstrate the superior performance of the proposed Ensemble (LSTM, mBERT, XLM-RoBERTa) model compared to individual models based on data from all four languages (weighted average F1-score for English is 0.93; for Kannada is 0.74; for Malayalam is 0.82; and for Tamil is 0.60). Thus, the proposed ensemble model proves to be a suitable approach for hope speech detection in the given low resource languages.

Abstract. The Lo's modified rescaled adjusted range test (R/S test) (Lo, 1991), GPH test (Geweke and Porter-Hudak, 1983) and two approximate maximum likelihood estimation methods, i.e., Whittle's estimator (W-MLE) and another one implemented in S-Plus (S-MLE) based on the algorithm of Haslett and Raftery (1989) are evaluated through intensive Monte Carlo simulations for detecting the existence of long-memory. It is shown that it is difficult to find an appropriate lag q for Lo's test for different short-memory autoregressive (AR) and fractionally integrated autoregressive and moving average (ARFIMA) processes, which makes the use of Lo's test very tricky. In general, the GPH test outperforms the Lo's test, but for cases where a strong short-range dependence exists (e.g., AR(1) processes with φ=0.95 or even 0.99), the GPH test gets useless, even for time series of large data size. On the other hand, the estimates of d given by S-MLE and W-MLE seem to give a good indication of whether or not the long-memory is present. The simulation results show that data size has a significant impact on the power of all the four methods because the availability of larger samples allows one to inspect the asymptotical properties better. Generally, the power of Lo's test and GPH test increases with increasing data size, and the estimates of d with GPH method, S-MLE method and W-MLE method converge with increasing data size. If no large enough data set is available, we should be aware of the possible bias of the estimates. The four methods are applied to daily average discharge series recorded at 31 gauging stations with different drainage areas in eight river basins in Europe, Canada and USA to detect the existence of long-memory. The results show that the presence of long-memory in 29 daily series is confirmed by at least three methods, whereas the other two series are indicated to be long-memory processes with two methods. The intensity of long-memory in daily streamflow processes has only a very weak positive relationship with the scale of watershed.

Machine Translation (MT) has recently been strengthened by the emergence of deep neural network techniques, which have brought the dream of language translation in community-based practices and clinical practices closer to reality. This chapter introduces the clinical practice of MT, especially in low resource languages. The following are some of the highlights of this chapter. How MT played a crucial role in good social reasons is focused on in this chapter. The chapter starts with the history of translation technologies that have played a vital role in various crisis and relief scenarios as the Haitian earthquake in 2010 and Translators without Borders (TWB), respectively. The Standby Task Force (SBTF) deployed the technology to tackle misinformation during Coronavirus Pandemic 2019. Next, the chapter explores the importance of the MT system, where it describes MT literacy as required across society to impact the healthcare of human beings. It describes the challenges of translation for clinical texts. Some of those are highlighted as (a) Domain: characteristics of clinical text differentiated from texts with other domains, (b) code mixed low-resource languages: noisy clinical data and lower quantity of data in the digital form available for computation, (c) lack of clinical domain corpora, and (d) very rich morphological structure: e.g., procedures (e.g., hepatico-cholangio-jejuno-stomy) and chemicals (e.g., Hydro-xy-nitro-di-hydro-thy-mine). Strategies or approaches to building MT on the clinical text can be categorized as: (1) Rule-based MT (early days), (2) Statistical MT (SMT, one of the successful methods in MT history), and (3) Neural MT (NMT, recent advancements of MT). Both SMT and NMT that work on same grounds by learning from text and are the primary focus of this chapter. L earning transparency for the patient's use of clinical data is an essential component. Much of today's healthcare data is unstructured text, such as clinical text, doctor's notes, nursing notes, lab reports, test reports, and more. A clinical practitioner is faced with a language barrier for communicating with a patient to interpret patient histories, examine a clinical diagnosis, or restate the recommended treatment method and follow-up to aid comprehension. The only way a medical practitioner encourages patients to ask questions or respond to queries is by directing them to input text with the help of MT. Finally, the chapter discusses contemporary techniques with Neural Machine Translation (NMT), namely multilingual, transfer learning on clinical data.

In order to solve the problem of insufficient human resources in the existing legal service industry, the automatic generation technology of legal documents based on machine learning and its legal risk analysis are proposed. In this paper, by synthesizing Encoder-Decoder model and Attention model of LSTM, and taking marital legal documents as the research dataset, we put forward a method of automatic generation of legal documents based on Encoder-Decoder model and Attention model, and take the generation of marital disputes legal documents as an example. The experimental results show that the convergence effect of Attention model is better than Encoder-Decoder model in the whole, which can reach more than 80%, and it has better effect on the processing of long text legal document information, and the prediction effect of Attention model is also better than Encoder-Decoder model. Therefore, the Attention model, in this experiment, generates legal documents with better accuracy, more in line with the original legal case, and initially meets the needs of general users of legal documents. Conclusion: This model optimizes the memory effect of LSTM model on long text, and can better complete the task of generating legal documents.

Mining people's semantic trajectory behaviours from geotagged photographs

Emotional content is a crucial ingredient in user-generated videos. However, the sparsity of emotional expressions in the videos poses an obstacle to visual emotion analysis. In this paper, we propose a new neural approach, Bi-stream Emotion Attribution-Classification Network (BEAC-Net), to solve three related emotion analysis tasks: emotion recognition, emotion attribution, and emotion-oriented summarization, in a single integrated framework. BEAC-Net has two major constituents, an attribution network and a classification network. The attribution network extracts the main emotional segment that classification should focus on in order to mitigate the sparsity issue. The classification network utilizes both the extracted segment and the original video in a bi-stream architecture. We contribute a new dataset for the emotion attribution task with human-annotated ground-truth labels for emotion segments. Experiments on two video datasets demonstrate superior performance of the proposed framework and the complementary nature of the dual classification streams.

BackgroundThe increasing use of social media platforms has given rise to an unprecedented surge in user-generated content, with millions of individuals publicly sharing their thoughts, experiences, and health-related information. Social media can serve as a useful means to study and understand public health. Twitter (subsequently rebranded as “X”) is one such social media platform that has proven to be a valuable source of rich information for both the general public and health officials. We conducted the first study applying Twitter data mining to autism screening.ObjectiveWe aimed to study the feasibility of autism screening from Twitter data and discuss the ethical implications of such models.MethodsWe developed a machine learning model to attempt to distinguish individuals with autism from their neurotypical peers based on the textual patterns from their public communications on Twitter. We collected 6,515,470 tweets from users’ self-identification with autism using “#ActuallyAutistic” and a separate control group. To construct the data set, we targeted English-language tweets using the search query “#ActuallyAutistic” posted from January 1, 2014 to December 31, 2022. We encrypted all user IDs and stripped the tweets of identifiable information such as the associated email address prior to analysis. From these tweets, we identified unique users who used keywords such as “autism” OR “autistic” OR “neurodiverse” in their profile description and collected all the tweets from their timelines. To build the control group data set, we formulated a search query excluding the hashtag “#ActuallyAutistic” and collected 1000 tweets per day during the same time period. We trained a word2vec model and an attention-based, bidirectional long short-term memory model to validate the performance of per-tweet and per-profile classification models. We deleted the data set and the models after our analysis.ResultsOur tweet classifier reached a 73% accuracy, a 0.728 area under the receiver operating characteristic curve score, and an 0.71 F1-score using word2vec representations fed into a logistic regression model, while the user profile classifier achieved an 0.78 area under the receiver operating characteristic curve score and an F1-score of 0.805 using an attention-based, bidirectional long short-term memory model.ConclusionsWe have shown that it is feasible to train machine learning models using social media data to predict use of the #ActuallyAutistic hashtag, an imperfect proxy for self-reported autism. While analyzing textual differences in naturalistic text has the potential to help clinicians screen for autism, there remain ethical questions that must be addressed for such research to move forward and to translate into the real world. While machine learning has the potential to improve behavioral research, there are still a plethora of ethical issues in digital phenotyping studies using social media with respect to user consent of marginalized populations. Achieving this requires a more inclusive approach during the model development process that involves the autistic community directly in the ideation and consent processes.

This investigation investigates the influence of a variety of data augmentation techniques on sentiment analysis in low-resource languages, with a particular emphasis on Bulgarian, Croatian, Slovak, and Slovene. The following primary research topic is addressed: is it possible to improve sentiment analysis efficacy in low-resource languages through data augmentation? Our sub-questions look at how different augmentation methods affect performance, how effective WordNet-based augmentation is compared to other methods, and whether lemma-based augmentation techniques can be used, especially for Croatian sentiment tasks. The sentiment-labelled evaluations in the selected languages are included in our data sources, which were curated with additional annotations to standardise labels and mitigate ambiguities. Our findings show that techniques like replacing words with synonyms, masked language model (MLM)-based generation, and permuting and combining sentences can only make training datasets slightly bigger. However, they provide limited improvements in model accuracy for low-resource language sentiment classification. WordNet-based techniques, in particular, exhibit a marginally superior performance compared to other methods; however, they fail to substantially improve classification scores. From a practical perspective, this study emphasises that conventional augmentation techniques may require refinement to address the complex linguistic features that are inherent to low-resource languages, particularly in mixed-sentiment and context-rich instances. Theoretically, our results indicate that future research should concentrate on the development of augmentation strategies that introduce novel syntactic structures rather than solely relying on lexical variations, as current models may not effectively leverage synonymic or lemmatised data. These insights emphasise the nuanced requirements for meaningful data augmentation in low-resource linguistic settings and contribute to the advancement of sentiment analysis approaches.