Rafael Brown’s Post

When I see an early stage #health start up saying they do #GenAI in 80+ languages, including Urdu, Arabic, or Swahili, am skeptical. When I speak to Arabic speakers, they tell me #ChatGPT performs poorly in Arabic for even basic tasks, how could you do #healthcare #generativeAI well in 80+ languages?

我識講廣東話...! (Ngo5 sik1 gong2 gwong2 dung1 waa2!) 😊 Google Translate adds Cantonese and 109 other languages with AI advancements • Google Translate has expanded its language support to include Cantonese, along with 109 other languages, through advancements in generative Artificial Intelligence (AI). This marks the largest-ever expansion of the app's language capabilities, made possible by Google's PaLM 2 large language model. • Cantonese, widely spoken in Chinese communities worldwide and in Hong Kong, has been a highly requested language for translation. However, due to its frequent overlap with Mandarin in written form, finding data and training models for Cantonese has been challenging. • Google aims to support the 1,000 most spoken languages globally and has also added Punjabi and several African languages, including Fon, Kikongo, Luo, Ga, Swati, Venda, and Wolof. The company plans to continue expanding its language offerings and improving translation accuracy by collaborating with expert linguists and native speakers. Good Read: Article about the topic; Full Story https://lnkd.in/gtwRYyGE

TextDetox CLEF 2024 We are glad to invite you to participate in the first of its kind multilingual Text Detoxification shared task! https://lnkd.in/dyWcuuS3 TL;DR Task formulation: transfer a text style from toxic to neutral (i.e. what a f**k is this about? -> what is this about?) 9 Languages: English, Spanish, Chinese, Hindi, Arabic, German, Russian, Ukrainian, and Amharic 🤗 https://lnkd.in/d6Wy4R3Q More details: Identification of toxicity in user texts is an active area of research. Today, social networks such as Facebook, Instagram are trying to address the problem of toxicity. However, they usually simply block such kinds of texts. We suggest a proactive reaction to toxicity from the user. Namely, we aim at presenting a neutral version of a user message which preserves meaningful content. We denote this task as text detoxification. In this competition, we suggest you create detoxification systems for 9 languages from several linguistic families. However, the availability of training corpora will differ between the languages. For English and Russian, the parallel corpora of several thousand toxic-detoxified pairs (as presented above) are available. So, you can fine-tune text generation models on them. For other languages, for the dev phase, no such corpora will be provided. The main challenge of this competition will be to perform both supervised and unsupervised cross-lingual detoxification. You are very welcome to test all modern LLMs on text detoxification and safety with our data as well as experiment with different unsupervised approaches based on MLMs or other paraphrasing methods! The final leaderboard will be built on a manual evaluation of a test set subset performed via crowdsourcing at Toloka.ai platform. In the end, you will have an opportunity to write and then present a paper at CLEF 2024 (https://clef2024.imag.fr/) which will take place in Grenoble, France! Important Dates February 1, 2024: First data available and run submission opens. April 22, 2024: Registration closes. May 6, 2024: Run submission deadline and results out. May 31, 2024: Participants paper submission. July 8, 2024: Camera-ready participant papers submission. September 9-12, 2024: CLEF Conference in Grenoble and Touché Workshop.

I’d love to see more work like this, (also see work by Jacaranda Health on a #Swahili Llama2 model). Uday T Kumar could you share more research or findings on fine-tuning Llama2 or other 7b-level models for low resource languages.. also how does this compare with Aya from Cohere For AI? (Which is closer to a 70b model?) I’m curious to how much you can push a 7b variant on languages with different language features like agglutinations and other morphological features and scripts (Arabic Cyrillic) how that affects the tokenizer and methods/best practices around that? Have you explored mixture of experts models to? Maxime Labonne .. and finally how you go about generating datasets for low-resource languages? Sara Hooker Shivalika Singh

AI and Languages I just watched a video where the new Chat GPT4-0 seamlessly translated a conversation between an Italian and English speaker. I know this isn’t the first tool to do this, but it’s the first time I’ve seen an example where I thought about how useful this is. It gave me the realization that instant language translation will revitalize diversity of language In my travels, I’ve noticed that English is a language that is becoming more and more widespread. Not everyone knows English, but recently in both France and Spain I had far less challenges communicating compared to my travels to France 12 years earlier. I think this stems from a move towards everyone desiring to speak a common language. Want to be able to talk to people in most parts of the world? Learn English. But maybe that desire will diminish now. If I get to speak in my mother tongue and someone who speaks English can hear a seamless translation, do I really need to learn English? Maybe in the future people will be less likely to pick up a new language? Will we see a slowdown in the acquisition of the English language? While I think we’ll see this shift, it won’t be drastic. Yet I can see both positives and benefits to this. A positive is that people will be more likely to hold on to the language of their heritage. A negative could be that in countries with high immigration the effort to learn the country’s home language might be less desirable. While this won’t necessarily cause an issue communicating since these AI tools can help, it can potentially undermine the social fabric of the country. And maybe that’s not as big a concern as I’m making it out to be? Still, I’m excited about the ease with which I’ll be able to travel to countries where the primary language isn’t English. I look forward to having conversations I could not have previously had. Tools like this make almost every person in the entire world a possible acquaintance, colleague, and friend. That’s a pretty exciting thing to think about. ~~~ As an aside, a lot of AI image creators still have issues with text, as the image accompanying this post demonstrates. This was my prompt: An English, Spanish, and French person sitting at a table, each saying “Good Morning” in their own language, in a speech bubble. _____ Shared on my Daily-Ink blog: https://lnkd.in/dx9Wszwc

The more things change, the more they stay the same. Back in the early days of the Internet, there was a lot of discussion about the digital divide. Often cited in international development circles was how the dominance of English on the Internet was slowing economic growth in Tier 1, Tier 2, and Tier 3 countries. The commercial software industry was so focused on the English-speaking market, that is did not see much point in developing for the non-English speaking world. As I moved into technology development, the challenge was not just accessing content in various languages, it was as fundamental as getting characters and scripts to render properly. This was critical to the DoD Mission I was supporting. We had to train our students to function digitally in foreign languages. So we had to develop our own unicode scripts to display Chinese, Arabic, Cyrillic, etc. This was the case well into the mid-2000s. Our mission still requires us to teach many languages that are less common and therefore underrepresented in AI training data. They have smaller populations of speakers and in some cases, are largely aural. Don't get me started on the challenges of right to left languages, something that many software tools and applications- including AI- still struggle with. As more people gained internet access, more content was created, and there was more demand for scripts. The industry started to see the money-making proposition. However, English still dominates the internet which means most of the AI training data is in English. The DoD has as one of its 5 Core Principles for AI "Equitable" which is focused on identifying and avoiding data bias. The main concern is that AI augmented tools will do harm or will act in unintended ways. But what of cultural and language marginalization? LLMs are great at predicting the next word. . . in English, French, Spanish, but what about the meaning and implications of the words in Arabic or Turkish? What happens when you have to label data for training AI systems and you have linguistic barriers in the labelling? Is a tomato always a tomato? While AI can do amazing things in many foreign languages, the digital divide is real and compounding quickly. This article from Stanford University's Human-Centered Artificial Intelligence explains well with a case study of AI and Vietnamese. https://lnkd.in/gcrjGmrS Chuck Casey Jeremy Tomaino Adam Frost Greg Lewis Habib S. Sandy Wagner Jesse High Michael Kanaan Dr. Cara Aghajanian

I admit my bias. I'm English, and therefore when it comes to linguistics, despite working for a localization company, I'm just not so great when it comes to other languages. This would have irritated the hell out of my Greek Cypriot family, but fortunately, at least my little brother can speak the language. So, why admit my bias? Well, it turns out that I'm not the only one who could be accused of bias. When it comes to LLMs, it shouldn't be a surprise to anyone that the vast majority of the data that they can utilise is in English. After all, the Internet and World Wide Web as we know it has a pretty substantial English base. Why is this important? Let's put it this way: I can think of at least one recent project where a client asked us to evaluate translations based upon utilising an AI tool. When we sent the samples off to human linguists, the universal opinion was that the translations were technically correct, but no one who actually spoke the language would use the chosen phrases. As noted in the article attached, that's something that has ramifications beyond simple translation. Does anyone really, as a business, want to rely on a tool that has a heavy Anglo bias when trying to develop relationships in other linguistic spheres? No-one really wants a hovercraft full of eels, no matter how technically correct that might be. #localisationservices #translationservices #largelanguagemodels #ChatGPT #websitetranslation #humantranslation

ChatGPT tip: Go deeper with translations and learning languages. Now that the NTU teaching semester has ended, I have more time to get back to my Chinese language studies. I was reading the Zaobao media interview with outgoing PM Lee Hsien Loong and came across this sentence: “当前纷扰的世界局势正扰乱多边贸易的开放常态，也引发各种区域安全隐患.” 1. I asked ChatGPT to “translate text into English”: ChatGPT: “The current tumultuous global situation is disrupting the normal state of open multilateral trade and is also causing various regional security risks.” 2. There are many different phrases in the sentence, so I told ChatGPT 4: “translate each term”. It dutifully did so, like this: 当前: current 纷扰: tumultuous 世界局势: global situation 引发: causing 3. Now, I’m not familiar with how to use the phrase 引发, so I asked ChatGPT: “when do i use 引发”.  It then gave an in-depth explanation of how to use 引发 in different scenarios. In 2021 BC (Before ChatGPT), I had to use Chinese websites like Baidu Translate or Baidu Encyclopedia to find the same explanations, but the text would be in Chinese. Now I can get similar guidance in English, with very clear examples using both Chinese and English sentences. This is a great way to learn languages, and do also try it out for your own translation needs. 

A paper titled "Evaluating LLMs’ Effectiveness in Culturally Nuanced, Low-Resource Real-World Scenarios" by Millicent Ochieng, Varun Gumma, Sunayana Sitaram, Jindong Wang, Vishrav Chaudhary, Keshet Ronen, Kalika Bali, and Jacki O'Neill investigates the performance of seven Large Language Models (LLMs) in sentiment analysis using a dataset derived from WhatsApp chats in English, Swahili, and Sheng (a mix of English and Swahili). Key findings include: 1. Performance Variation: GPT-4 and GPT-4-Turbo excelled in understanding diverse linguistic inputs and context, showing high consistency and transparency, whereas other models like Mistral-7b and Mixtral-8x7b achieved high F1 scores but struggled with linguistic nuances and transparency. 2. Cultural Nuance Challenges: All LLMs had difficulties incorporating cultural nuances, especially in non-English settings, indicating a need for continuous improvement in handling low-resource languages and culturally diverse data. 3. Evaluation Methodology: The study combined quantitative analysis (F1 scores) with qualitative assessment of LLMs' explanations, highlighting the importance of qualitative Human-Computer Interaction (HCI) methods in understanding model performance in real-world, multilingual scenarios.

What will happen with small languages in a world of large language models? A few days ago, Alf B. Prøven asked Inga Strümke, Morten Goodwin, and me a question here on LinkedIn. I paraphrase: "Should Norway spend substantial resources on building large language models?"  The answer to that is, in my opinion, yes. All nations should consider prioritising this. 1: Existing LLMs don’t support other languages than English well Generic large language models (LLMs) built by vacuuming the internet are overly English-biased. Various analyses find that over 60% of the world's web pages are in English. All other languages drown out, including French and German, each with around 2,5 % of the corpus.   Generally, the larger the input, the better the model. If you want good results, you will use English.   2: Languages not used in the digital world will not survive Over time, languages that aren't supported by dominant digital platforms are likely to be forgotten or transformed. 3: Technology shapes our culture Technology, in general, and LLMs, particularly, define the way we look at life, the universe, and everything. The texts on the internet, and therefore the LLMs, reflect cultural norms, values, history, ethics, and ways of communicating, debating politics, conducting business, raising children, and everything you can think of. That's a reason the Japanese government decided that the Japanese would not have copyright for use in LLMs. Maintaining the copyright of Japanese texts would make the Japanese language, culture, norms, and ethics, less present and reflected in the language models. 4: LLMs are platforms LLMs are foundations, or platforms, on top of which our future innovations, industry, defence, science, engineering – anything you can imagine - will be built.  Hence, we should not consider foundation models as just another technology. Rather, they are an enabling technology that will be used as a platform for building our future societies. 5: LLMs are critical infrastructure LLMs should be viewed as critical infrastructure and cultural corpus because of their ubiquitousness, transformative power, and cultural bias. We would then discuss their role in society differently—and thus, whether we should build our own. 6: We need to build national LLMs A couple of months ago, the Expert Group on Ethical AI (of which I was a member) to The Nordic Council of Ministers and The Nordic Council delivered its recommendations to the Nordic governments. One recommendation was that the Nordic countries develop LLMs that reflect their unique linguistic and cultural diversity, ensuring digital sovereignty and alignment with democratic values. 7: Foundation models are too important only to be developed and controlled by a few commercial companies. #CLAIRE NORA – The Norwegian Artificial Intelligence Research Consortium Adra - AI-Data-Robotics-Association Peter Møllgaard Søren Hvidkjær Trine Johansen Meza Christen Krogh Digital Norway Digital Dogme