Learning Personal Human Biases and Representations for Subjective Tasks in Natural Language Processing

The advances in deep learning have led to great achievements in many Natural Language Processing (NLP) tasks. With the nature of language, i.e., sequential data, most NLP tasks can be framed into the sequence learning framework, such as text generation. As one of the most important foundations for modern NLP techniques, autoregressive generation models have achieved dominant performance in a great deal of NLP tasks. Therefore, this thesis emphasizes improving the autoregressive generation model for different NLP tasks. While many tasks can naturally fit into the sequence learning framework, some of them, e.g., building discourse parsing tree, require sophisticated designs to fit into neural models. Therefore, this thesis firstly emphasizes a novel unified framework for discourse parsing, which builds a discourse tree in a top-down depth-first manner, and it frames the task as an autoregressive generation task with the goal of each step being the prediction of the node position given a piece of text. The proposed approach is proven effective with extensive empirical experiments. In addition, I extend the above framework by proposing a hierarchical decoder, which leverages the information from parents and siblings of the nodes that are currently processed. The proposed decoder utilizes the nature of the tree structure and further improves the experiment performance on both discourse parsing and dependency parsing tasks. On the other hand, the de facto strategies, i.e., cross entropy loss and teacher forcing, for training the autoregressive generation models have been shown problematic in certain aspects. For example, cross entropy loss, which is one of the widely leveraged training objective functions, often leads to text degeneration in text generation, and teacher forcing suffers from the exposure bias problem, where there exists a mismatch between the training and testing setup. For text degeneration, I introduce a class of diminishing attentions, which enforces the submodularity of the coverage calculated by cross attention in the sequence-to-sequence model. The proposed diminishing attentions achieve notable improvement on several neural text generation tasks, including text summarization, machine translation, and image paragraph generation. Further, I propose a novel training objective, ScaleGrad, to replace cross entropy, which significantly reduces the degeneration problem in different text generation tasks. In fact, ScaleGrad can be extended to problems beyond text degeneration. It provides wide flexibility to inject different inductive biases into the text generation model by directly modifying the gradient information in the output layer. Next, for the exposure bias problem, this thesis introduces a novel type of scheduled sampling based on training accuracy, which requires only minimal hyper-parameter tuning compared to existing scheduled sampling methods. Additionally, a novel imitation loss is proposed to further enforce the model’s generative behavior to match the teacher-forced behavior. Moreover, this thesis demonstrates that reducing exposure bias can improve the robustness of language models against repetition and toxic errors.

Text extraction is an important part of natural language processing (NLP) tasks. Most NLP tasks like text classification, machine translation, text-to-speech, text-based language identification, text summarization, and named-entity recognition involve the use of textual data. Such data is limited for low-resourced languages making it difficult to experiment advanced NLP techniques on these languages. This paper presents a Python-based toolkit for text analysis and text extraction from different types of images, documents, and audio files. The toolkit is built as a library that has functions that can be imported and utilized for text extraction.

Long Short-Term Memory (LSTM) networks are unique to exercise data in its memory cell with long-term memory as Natural Language Processing (NLP) tasks have inklings of intensive time and computational power due to their complex structures like magnitude language model Transformer required to pre-train and learn billions of data performing different NLP tasks. In this paper, a dynamic chaotic model is proposed for the objective of transforming neurons states in network with neural dynamic characteristics by restructuring LSTM as Chaotic Neural Oscillatory-Long-Short Term Memory (CNO-LSTM), where neurons in LSTM memory cells are weighed in substitutes by oscillatory neurons to speed up computational training of language model and improve text classification accuracy for real-world applications. From the implementation perspective, five popular datasets of general text classification including binary, multi classification and multi-label classification are used to compare with mainstream baseline models on NLP tasks. Results showed that the performance of CNO-LSTM, a simplified model structure and oscillatory neurons state in exercising different types of text classification tasks are above baseline models in terms of evaluation index such as Accuracy, Precision, Recall and F1. The main contributions are time reduction and improved accuracy. It achieved approximately 46.76% of the highest reduction training time and 2.55% accuracy compared with vanilla LSTM model. Further, it achieved approximately 35.86% in time reduction compared with attention model without oscillatory indicating that the model restructure has reduced GPU dependency to improve training accuracy.

Work orders, equipment information, technical records and best practices documents contain within them a wealth of insights related to Equipment Maintenance which can be unlocked with Natural Language Processing tasks like classification, clustering, named entity recognition or part of speech tagging. But obtaining large enough labelled data sets in Equipment Maintenance domain manually is prohibitive and very expensive. This lack of labeled Equipment Maintenance data can be overcome with Large Language Models (LLMs) such as GPT-3, BERT that are pretrained transformer networks, considered state-of-the-art when it comes to Natural Language Processing (NLP) tasks. However, the vocabulary understood by these LLMs are mostly from English Language and need to be fine-tuned to understand industry and organization specific vocabulary and acronyms. This paper explores the potential of a domain specific LLM model for oil and gas industries. Data that are of good quality and that provide a comprehensive overview of industry are collected. This corpus of text contains documents like work orders, equipment data and technical documents. A custom tokenizer is trained on this data to identify domain specific terminology. A comparative study is done with other off-the-shelf tokenizers: BERT and RoBERTa, to compare the effectiveness of the tokenization. With millions of work orders and equipment documents, training pipelines had to parallelized so that training can occur on multiple GPUs. A comprehensive study of multiple training methods is done in this paper. Model and tokenizer developed were packaged and archived to be consumed in machine learning pipelines to specific use-cases across the organization. For an organization adopting digital transformation, the availability of an organization specific LLM is an enabler to extract insights from millions of documents containing free text. The applicability of such models spans across multiple disciplines like Maintenance, Reliability, Safety etc. and streamlines the development of highly accurate and robust text analytics.

Document classification is a prevalent task in Natural Language Processing (NLP), which has an extensive range of applications in the biomedical domains such as biomedical literature indexing, automatic diagnosis codes assignment, tweets classification for public health topics, and patient safety reports classification. Nevertheless, manual classification of biomedical articles published every year into specific predefined categories becomes a cumbersome task. Hence, building an automatic document classification for biomedical databases emerges as a significant task among the scientific community. In recent years, Deep Learning (DL) models like Deep Neural Networks (DNN), Convolution Neural Networks (CNN), Recurrent Neural Networks (RNN), and Ensemble Deep Learning models are widely used in the area of text document classification for better classification performance compared to Machine Learning (ML) algorithms. The major advantage of using DL models in document classification is that it provides rich semantic and grammatical information for document representation through pre-trained word embedding. Hence, this paper investigates the deployment of the various state-of-the-art DL based classification models in automatic classification of benchmark biomedical datasets. Finally, the performance of all the aforementioned constitutional classifiers is compared and evaluated through the well-defined performance evaluation metrics such as accuracy, precision, recall, and f1-measure.

Towards a Novel Weakly Supervised Joint Approach of Named Entity Recognition and Normalization for Noisy Text

In this paper, we present the preliminary results on the analysis of deep learning terms used for natural language processing (NLP) tasks. We propose a statistical analysis of papers published from 2012 to 2015 in the main ACL conferences. Our aim is investigating which DL term, and consequently which DL method, is mostly used for each specific NLP task, since its introduction in the field. In order to do this, our first contribution is the development of two terminological lists, referring respectively to DL methods for text analysis and NLP tasks. The list of deep learning terms contains 41 terms and acronyms, as well as a NLP term list contains 145 terms and acronyms. From our corpus, the frequencies of various terms have been extracted with respect to the ACL conference and the publication year. After the preliminary data analysis, we decided to restrict the extraction process to abstract texts. We applied multivariate techniques called correspondence analysis in order to visualize and evaluate the joint behavior of our variables.

In this study, we explore the potential of machine learning (ML) models after different text representation methods on the balanced IMDB dataset, which is widely regarded as a gold standard in sentiment analysis, one of the Natural Language processing (NLP) tasks. On the open source IMDB movie reviews dataset, we first undertake data cleaning and text representation with data preprocessing steps. Then, we apply sentiment classification using different ML models. In order to evaluate the models, we used precision (P), recall (R), F1-score (F1), and area under curve (AUC), as well as receiver operating characteristic (ROC). It is worth noting that text feature extraction with Bidirectional Encoder Representations from Transformers (BERT) provided the highest performance in all models, with the SVM model offering particularly promising results. In this model, we observed the following results: ACC 0.9033, F1 0.9308, R 0.9015, R 0.9015, P 0.9072, AUC 0.9638, and ROC 0.96. These findings suggest that NLP techniques and, in particular, machine learning models that employ BERT may offer high levels of accuracy and reliability in text classification problems. It would be beneficial for future studies to validate these findings using BERT on different NLP tasks. This would help to evaluate the effectiveness and applicability of the models in practice.

Towards a Novel Weakly Supervised Joint Approach of Named Entity Recognition and Normalization for Noisy Text

Sarcasm detection is a challenging task in Natural Language Processing (NLP) due to its reliance on contextual, contrastive, and figurative cues. While deep learning models such as BERT have advanced contextual representation learning, they often rely on computationally expensive end-to-end fine-tuning. This paper proposes a hybrid framework that integrates BERT embeddings with a custom self-attention mechanism for feature extraction, combined with XGBoost classification for efficient decision making. The self-attention layer highlights sarcasm-relevant tokens, producing contextualized, weighted sentence representations, which are then fed into XGBoost to establish robust decision boundaries without neural classification overhead. We further employ Stratified 6-fold cross validation to ensure balanced evaluation and reliable performance reporting. Experiments on benchmark sarcasm datasets demonstrate that the proposed approach achieves superior F1-scores compared to traditional baselines, validating the effectiveness of combining contextual embeddings with gradient-boosted classification. This work highlights the potential of attention-augmented embeddings with ensemble learning for advancing sarcasm detection in social media text.

Pre-training and fine-tuning have become the de facto paradigm in many natural language processing (NLP) tasks. However, compared to other NLP tasks, neural machine translation (NMT) aims to generate target language sentences through the contextual representation from the source language counterparts. This characteristic means the optimization objective of NMT is far from that of the universal pre-trained models (PTMs), leading to the standard procedure of pre-training and fine-tuning does not work well in NMT. In this paper, we propose a novel framework to deep fuse the pre-trained representation into NMT, fully exploring the potential of PTMs in NMT. Specifically, we directly replace the randomly initialized Transformer encoder with a pre-trained encoder and propose a layer-wise coordination structure to coordinate PTM and NMT decoder learning. Then, we introduce a partitioned multi-task learning method to fine-tune the pre-trained parameter, reducing the gap between PTM and NMT by progressively learning the task-specific representation. Experimental results show that our approach achieves considerable improvements on WMT14 En2De, WMT14 En2Fr, and WMT16 Ro2En translation benchmarks and outperforms previous work in both autoregressive and non-autoregressive NMT models.

<title>Abstract</title> The Tenyidie language, a.k.a Angami language, is a Tibeto-Burman Language of Nagaland in the north-eastern part of India. Tenyidie is a low-resource language spoken by the Tenyimia Community and is considered a major language in Nagaland. Tenyidie is a tonal, Subject-Object-Verb, and highly agglutinative language. Among the many Natural Language Processing (NLP) tasks, part-of-speech tagging is one of the primary tasks which is used in building many other NLP tasks, such as dependency parsing, named entity recognition, machine translation, etc. The main aim of this paper is to develop the Part-of-speech (POS) Tagset and Tagger for the Tenyidie Language. To the best of our knowledge, there is no reported work on POS Tagset and POS Tagger for the Tenyidie language. The main contributions of this research are: creation of POS tagset for the Tenyidielanguage, creation of POS annotated dataset (11,540 sentences), and the verification of the proposed POS tagset and annotated dataset using the Machine Learning/Deep Learning techniques such as CRF, BLSTM, etc., including the state-of-the-art model such as BERT.

Frame semantics is a theory of meaning in natural language, which defines the structure of the lexical semantic resources known as framenets. Both framenets and frame semantics have proved useful for a number of natural language processing (NLP) tasks. However, in this connection framenets have often been criticized for their limited coverage. A proposed reasonable-effort solution to this problem is to develop domain-specific (sublanguage) framenets to complement the corresponding general-language framenets for particular NLP tasks, and in the literature we find such initiatives covering domains such as medicine, soccer, and tourism. In this paper, we report on building a framenet to cover the terms and concepts encountered in descriptive linguistic grammars (written in English) i.e. a framenet for the linguistic domain (LingFN) to complement the general-language BFN.KeywordsFrame semanticsFramenetsSemantics

Word embeddings, distributed word representations in a reduced linear space, show a lot of promise for accomplishing Natural Language Processing (NLP) tasks in an unsupervised manner. In this study, we investigate if the success of word2vec, a Neural Networks based word embeddings algorithm, can be replicated in an aggluginative language like Turkish. Turkish is more challenging than languages like English for complex NLP tasks because of her rich morphology. We picked ontology enrichment, again a relatively harder NLP task, as our test application. Firstly, we show how ontological relations can be extracted automaticaly from Turkish Wikipedia to construct a gold standard. Then by running experiments we show that the word vector representations produced by word2vec are useful to detect ontological relations encoded in Wikipedia. We propose a simple but yet effective weakly supervised ontology enrichment algorithm where for a given word a few know ontologically related concepts coupled with similarity scores computed via word2vec models can result in discovery of other related concepts. We argue how our algorithm can be improved and augmented to make it a viable component of an ontoloy learning and population framework.

This chapter presents a brief introduction to entropy guided transformation learning (ETL), a machine learning algorithm for classification tasks. ETL generalizes transformation based learning (TBL) by automatically solving the TBL bottleneck: the construction of good template sets. The main advantage of ETL is its easy applicability to natural language processing (NLP) tasks. This introductory chapter presents the motivation behind ETL and summarizes our experimental results. In Sect. 1.1, we first briefly detail TBL and explain its bottleneck. Next, we briefly present ETL and list some of its advantages. In Sect. 1.2, we first list some related works on the use of ETL for different NLP tasks. Next, we report a summary of our experimental results on the application of ETL to four language independent NLP tasks: part-of-speech tagging, phrase chunking, named entity recognition and semantic role labeling. Finally, in Sect. 1.3, we detail the structure of the book.KeywordsMachine learningEntropy guided transformation learningETL committeeTransformation based learningNatural language processingPart-of-speech taggingPhrase chunkingNamed entity recognitionSemantic role labeling