Cleansing Jewel: A Neural Spelling Correction Model Built On Google OCR-ed Tibetan Manuscripts

Scholars in the humanities rely heavily on ancient manuscripts to study history, religion, and socio-political structures in the past. Many efforts have been devoted to digitizing these precious manuscripts using OCR technology, but most manuscripts were blemished over the centuries so that an Optical Character Recognition (OCR) program cannot be expected to capture faded graphs and stains on pages. This work presents a neural spelling correction model built on Google OCR-ed Tibetan Manuscripts to auto-correct OCR-ed noisy output. This paper is divided into four sections: dataset, model architecture, training and analysis. First, we feature-engineered our raw Tibetan etext corpus into two sets of structured data frames -- a set of paired toy data and a set of paired real data. Then, we implemented a Confidence Score mechanism into the Transformer architecture to perform spelling correction tasks. According to the Loss and Character Error Rate, our Transformer + Confidence score mechanism architecture proves to be superior to Transformer, LSTM-2-LSTM and GRU-2-GRU architectures. Finally, to examine the robustness of our model, we analyzed erroneous tokens, visualized Attention and Self-Attention heatmaps in our model.

A Perspectival Mirror of the Elephant: Investigating Language Bias on Google, ChatGPT, Wikipedia, and YouTube

Contrary to Google Search's mission of delivering information from "many angles so you can form your own understanding of the world," we find that Google and its most prominent returned results -- Wikipedia and YouTube, simply reflect the narrow set of cultural stereotypes tied to the search language for complex topics like "Buddhism," "Liberalism," "colonization," "Iran" and "America." Simply stated, they present, to varying degrees, distinct information across the same search in different languages (we call it 'language bias'). Instead of presenting a global picture of a complex topic, our online searches turn us into the proverbial blind person touching a small portion of an elephant, ignorant of the existence of other cultural perspectives. The language we use to search ends up as a cultural filter to promote ethnocentric views, where a person evaluates other people or ideas based on their own culture. We also find that language bias is deeply embedded in ChatGPT. As it is primarily trained on English language data, it presents the Anglo-American perspective as the normative view, reducing the complexity of a multifaceted issue to the single Anglo-American standard. In this paper, we present evidence and analysis of language bias and discuss its larger social implications. Toward the end of the paper, we propose a potential framework of using automatic translation to leverage language bias and argue that the task of piecing together a genuine depiction of the elephant is a challenging and important endeavor that deserves a new area of research in NLP and requires collaboration with scholars from the humanities to create ethically sound and socially responsible technology together.