<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sofwathullah Mohamed on Medium]]></title>
        <description><![CDATA[Stories by Sofwathullah Mohamed on Medium]]></description>
        <link>https://medium.com/@sofwathullah.mohamed?source=rss-895c235b6d47------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*OLFk2Co3uxTzOLqr</url>
            <title>Stories by Sofwathullah Mohamed on Medium</title>
            <link>https://medium.com/@sofwathullah.mohamed?source=rss-895c235b6d47------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 21:41:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sofwathullah.mohamed/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[how are you ← ހާލު ކިހިނެއް]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/how-are-you-%DE%80%DE%A7%DE%8D%DE%AA-%DE%86%DE%A8%DE%80%DE%A8%DE%82%DE%AC%DE%87%DE%B0-173d0349650e?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/173d0349650e</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Sun, 09 Dec 2018 14:24:07 GMT</pubDate>
            <atom:updated>2018-12-10T12:20:02.607Z</atom:updated>
            <content:encoded><![CDATA[<h4>Translation from Dhivehi to English using Attention</h4><p>This is an experiment to train a sequence to sequence model for English to Dhivehi translation.</p><p>We use a “seq2seq” model that has encoder, decoder and intermediate step as its main components. Embedding is used, so we have to first compile a “vocabulary” list containing all the words we want our model to be able to use or read. The model inputs will have to be tensors containing the IDs of the words in the sequence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*m4pqJh3GwXcReULBM5Bdpw.png" /></figure><p>The dataset used can be downloaded from <a href="https://drive.google.com/file/d/1QTjo0iUKlAWeolkN29sHUZCvHkaOoGzR/view?usp=sharing">here</a></p><h4>Sample Results</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*TEW-ukTtfCQWIMpNvTCdng.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/652/1*w7l9DmAqnNJs3XtcV3g2vw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/652/1*w7l9DmAqnNJs3XtcV3g2vw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*TEW-ukTtfCQWIMpNvTCdng.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/431/1*0t9CnqAeuSCy7vpiBcYyxA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/505/1*3vUKC5CrGmhZCZ-4NLFgbg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/658/1*RBMDz75xZY99eQ-73LPjCw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/544/1*UQQxZcKwINyRCSu-hA1VPg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/533/1*HVD0PX_k2LP2EFOcFPa0Jw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/511/1*wgwD2ysZBPSw6KpJ5hgyvw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/623/1*Hy9A5UqtBUubbYEYVd4-vQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/1*1HBB-bOsVuTofQWlNQOBjg.png" /></figure><h4>Code for the Model</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/8af1b356dceadb45816943daf87054f8/href">https://medium.com/media/8af1b356dceadb45816943daf87054f8/href</a></iframe><p>All the code is based on Tensorflow and it was adopted from the tutorial provided on the official documentation.</p><h4>Conclusion</h4><p>The translation quality is reasonable for a toy example using a small dataset. However more research is needed on how we can produce better translations for Dhivehi. The amount/quality of data would be one major factor.</p><p>Training NMT models requires text that has been paired with reference translations, known as <em>bilingual data</em>. Unfortunately, bilingual data for Dhivehi is very hard to find.</p><p>Another factor is training of this type of a models requires several days on a single machine, which is too slow for experimentation. The FREE options are very limited on the resources and MNT models are usually very resource hungry.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=173d0349650e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ބަލާފަ .( ވިސްނާފަ) . ބުނޭ]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/%DE%84%DE%A6%DE%8D%DE%A7%DE%8A%DE%A6-%DE%88%DE%A8%DE%90%DE%B0%DE%82%DE%A7%DE%8A%DE%A6-%DE%84%DE%AA%DE%82%DE%AD-b6caf932a813?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6caf932a813</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Thu, 06 Dec 2018 11:10:07 GMT</pubDate>
            <atom:updated>2018-12-06T18:07:55.023Z</atom:updated>
            <content:encoded><![CDATA[<h3>ބަލާފަ .( ވިސްނާފަ) . ބުނޭ</h3><p>Image captioning is generating automatic text descriptions for a given image. What if we can get the results in Thaana, that would be even more fun. For this a Machine Learning model is trained with Thaana text descriptions and images as input.</p><p>The technical details are published in a paper titled “<a href="https://arxiv.org/pdf/1502.03044.pdf">Show, Attend and Tell: Neural Image Caption Generation with Visual Attention</a>”</p><p>The basic idea here is that you give the model an image and it will output(generate) a label/caption for the given image. It is done by using computer vision to understand the content of the image and a language model (natural language processing) to turn the understanding of the image into words in the right order.</p><h4>Sample Outputs</h4><p>Input Image:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*bWuhvfdSDSDakoZwT4KDpw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/731/1*9sK0XD6iHt-tQ8NrcSVYPQ.png" /></figure><p>Output Generated : ‘ބްރިޖްގެ ފައުންޑޭޝަނަށް ޕްލެޓް ފޯމް އަޅާފައި’</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/765/1*rDefd8S3ZM9Mb3O1Vzg6yA.png" /></figure><p>Output Generated: ‘އުރީދޫ ދިވެހި ޕްރިމިއާ ލީގުގައި މިރޭ ކައިރިން ބޯޅަ ހިފައިގެންދަނީ ‘</p><p>Input Image:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*L2kVHT_diW7qlJlScemoow.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/752/1*FU4Eb7tmaRFqjKduSG9I8g.png" /></figure><p>Output Generated: &#39;ރައްޔިތުންގެ މަޖިލީހުގެ ޖަލްސާއެއްގައި މެމްބަރުން ވާހަކަދައްކަވަނީ &#39;</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/610/1*KlLHjM8KPrQIRTRIOLzwKQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/782/1*eDSrBpgdmRJu0QW-xfPR7Q.png" /></figure><h4>Model</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WCR0h_oMYCCf-GWG-Fbo_Q.jpeg" /><figcaption>The model is trained in a sequence to sequence manner. A feature a_t, at time t, from location s_t is sampled and fed to a LSTM to generates a label. This process is repeated K times to generate K-words image caption.</figcaption></figure><h4>Code</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/6e993c17eeb9ae0f6450cbc314fbf729/href">https://medium.com/media/6e993c17eeb9ae0f6450cbc314fbf729/href</a></iframe><p>Most of the code is from Tensorflow example.</p><h4>Data</h4><p>The data was used was taken from mihaaru news site. The dataset can be found <a href="https://drive.google.com/file/d/1lqGJjuym2EatIG9zzn_xkQ7ktrF-rEy6/view?usp=sharing">here</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6caf932a813" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ދެން ބުނާނީ ކީކޭބާ؟]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/%DE%8B%DE%AC%DE%82%DE%B0-%DE%84%DE%AA%DE%82%DE%A7%DE%82%DE%A9-%DE%86%DE%A9%DE%86%DE%AD%DE%84%DE%A7-5f7206a49507?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f7206a49507</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Thu, 29 Nov 2018 19:39:33 GMT</pubDate>
            <atom:updated>2018-11-29T19:39:33.828Z</atom:updated>
            <content:encoded><![CDATA[<p>Predicting the next word in a sentence can be a bit of a challenge. One common way of doing this was using a text corpus and identify the most common n-grams/trigrams. The same can be achieved and can produce better results if a Machine Learning model is used, in this case a RNN.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/591/1*JmqD0YvKlMJQEjmu74mhyw.png" /><figcaption>source: <a href="https://www.superdatascience.com/the-ultimate-guide-to-recurrent-neural-networks-rnn/">https://www.superdatascience.com/the-ultimate-guide-to-recurrent-neural-networks-rnn/</a></figcaption></figure><p>Hochreiter and Schmiduber introduced the Long Short Term Memory networks, usually called “LSTMs” . This is a Recurrent Neural Network (RNN) that is trained using backpropagation through time and overcomes the vanishing gradient problem. LSTMs are widely used for speech recognition, language modelling, sentiment analysis and text prediction.</p><p>RNNs are capable of a number of different types of input / output combinations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Kje9TpsSvMGeBF6vIv6k5Q.jpeg" /><figcaption>source: <a href="https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/recurrent_neural_networks.html">https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/recurrent_neural_networks.html</a></figcaption></figure><p>You can read more on RNN <a href="https://www.deeplearningbook.org/contents/rnn.html">here</a> and <a href="https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/recurrent_neural_networks.html">here</a>.</p><h4>A Quick Implementation of Keras LSTM TimeDistributed for Thaana Next Word Prediction.</h4><p>For this I used 10,000 lines of Dhivehi news headlines as the training dataset.</p><p>If you want to get better accuracy on the output, you can use a bigger dataset(<a href="https://drive.google.com/file/d/1a_ZRZFM8j3ir2HeUYomloMLgTYn77z93/view?usp=sharing">data.txt</a>). Since I don’t have that much computing power on my MacBook Pro, I used <a href="https://colab.research.google.com/">Google Colab</a> to train the model using <a href="https://cloud.google.com/tpu/">TPUs</a>. However since there is a restriction on the memory limit, I was not able to train on the full dataset.</p><p>The other challenge was Colab only gives free 12 hours of processing per session, so I had to make sure the model was a trained in that time frame or less. Meaning the dataset had to be small and model needed to be optimised. If I ran the training locally (with the full dataset) on the MacBook Pro (on CPU), I guess it would take me a few months or more to get it trained.</p><h4>The <strong>Model Architecture</strong></h4><p>What we have is basically a sequence learning problem. The model setup is a multi layer LSTM in Keras with 4 LSTM layers. The input to the LSTM is the last 3 words and the target for LSTM is the next word. The next 3 layers are TimeDistributed on the output layer to wrap a fully connected Dense layer. This layer allows us to build the model that can do <em>one-to-many</em> and <em>many-to-many</em> architectures. This is because the output function for each of the “many” outputs is the same function applied to each timestep.</p><p>The final layer of the model is a softmax layer. This predicts the likelihood of each word.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/560/1*kw7nIb2i7lp3EuggfASYUQ.png" /></figure><h4>Output</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/234/1*0YixbJX9y_GrK6or4Smkrg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/227/1*_9AOT9iDjksy_o-NjpmU6Q.png" /></figure><h4>Code — TPU Version</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/c98d4d2f94eb890868d0f92039cec9a7/href">https://medium.com/media/c98d4d2f94eb890868d0f92039cec9a7/href</a></iframe><h4>Next Step</h4><p>Explore alternate model architecture that allow training on a much larger dataset to produce better results. You can look at some of these strategies in the paper “Strategies for Training Large Vocabulary Neural Language Models”- <a href="https://arxiv.org/abs/1609.07843">https://arxiv.org/abs/1512.04906</a></p><p>Converting the model to tensorflow.js for web deployment.</p><h4>Further Reading</h4><p><a href="https://keras.io/layers/wrappers/#timedistributed">TimeDistributed</a> in the Keras API</p><p><a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">Understanding LSTM Networks</a></p><p><a href="https://skymind.ai/wiki/lstm">A Beginner’s Guide to LSTMs and Recurrent Neural Networks</a></p><p><a href="https://blog.kovalevskyi.com/howto-start-using-tpus-from-google-colab-in-few-simple-steps-fff0fd2cb361">HowTo Start Using TPUs From Google Colab in Few Simple Steps</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f7206a49507" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Should Dhivehi/Thaana Matter in the Digital Age?]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/should-dhivehi-thaana-matter-in-the-digital-age-ca865cd48768?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca865cd48768</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Tue, 20 Nov 2018 10:15:08 GMT</pubDate>
            <atom:updated>2018-11-20T10:15:08.693Z</atom:updated>
            <content:encoded><![CDATA[<p>There are mixed views on the subject of investing on Dhivehi/Thaana in the digital domain. There are some who feel that we should get rid of Thaana and others who feel we need to do more work on this area.</p><p>The reality is Dhivehi/Thaana still is the official language of the Maldives. Officially all government communication needs to be in Dhivehi/Thaana. After the fonts and a Thaana keyboard for the computers and the mobile phones (input methods) we seem to have stopped there. This too is mostly work of individuals who have contributed to the fonts and input methods. We are able to read Thaana text and do basic input. We are happy now?</p><p>My view is if we don’t work more on Thaana it will die soon and it will be something from the past. Kids today talk in English more than they do in Dhivehi , while there are some kids who can’t even read or write in Thaana. Most of us are left with no choice as most of the news is in Dhivehi. When it comes to social media and messaging mostly we use Latin.</p><p>The world is changing and fast. Keyboards won’t be the only input methods soon. We are at the dawn of AI and other exciting changes. From our phones to devices like Amazon’s Alexa, we live in a world surrounded by machine learning. Without investing on Natural Language Processing for Dhivehi the language will die. Technology will kill the language and technology can also save the language. It’s up to us to decide which direction to go.</p><h3><strong>Fonts</strong></h3><p>Thanks to Google (and a team of local Maldivians who helped with Thaana font), today we have the “Noto” fonts which includes <a href="https://medium.com/kodefly/is-noto-sans-thaana-relevant-29a8307fb6cb">Thaana</a>. This enabled us to render Thaana without having to install the Thaana fonts every time (as the “Noto” fonts is now bundled with iOS/MacOS and Android). There is still more work to be done on the Thaana fonts side. We take this for granted. You are able to read and type in Thaana, but this is happening because some people worked for it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/513/1*1cYvj6qFbpL5M9Fmq33XRA.png" /></figure><h3><strong>Input Methods.</strong></h3><p>We do have input methods for Thaana. Windows, MacOS, iOS and Android. Thanks to companies like Microsoft and Google. With machine learning we are now even able to do word suggestions based on what you type and what you’re currently typing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/277/1*t3Ixad_CgQzIw3XVNv7Fzw.png" /></figure><p>This is an example of “natural language processing” in action. This makes your life much easier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/632/1*Jt8gIr5SDGT3HU-fJygjIQ.png" /></figure><h3><strong>Thaana OCR</strong></h3><p>Thaana OCR is something BIG! Thanks to Machine Learning, we are now able to do optical character recognition for Thaana images and get the text.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*f4k2WgNPoI50ZGi3ZwDvPA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/471/1*uQIf2VdqOtuy9DnKK7WVUg.png" /></figure><p>It’s not perfect yet, but we are now able to use <a href="https://github.com/Sofwath/GoogleCloudVisionOCR">Google Vision API</a> and get Thaana OCR done.</p><p><a href="https://github.com/tesseract-ocr/tesseract">Tesseract</a> OCR (Tesseract 4) too now has support for Thaana.</p><h3>Whats Next?</h3><p>There is a lot more that is missing. We need to do more work on Thaana NLP. The whole idea is to make things easy so people will find it easy to adopt Thaana on the digital domain. Example is the input method, where word suggestions help in typing. This encourages more people to use Thaana.</p><h4>Latin to Thaana Transliteration</h4><p>Though not given much priority this is something which cannot be ignored anymore. Most of the user generated content from social media and other platforms such as messaging platforms is in Latin.</p><p>We need better methods to handle this data. We do have a few rule based transliterating tools but these tools won’t do the job. The biggest problem is that we can’t force users to follow standards. Different individuals use different standards for Latin and there is a lot of mistakes while typing. We need tools which should be able to handle all this. Adding to the complexity we also use English words with latin. These needs to be handled differently.</p><p>The only possible solution is using Machine Learning to solve this problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/412/1*WW2N9fG5ZtkHGKSjuuwWlw.png" /></figure><p>Example output generated using a <a href="https://medium.com/@sofwathullah.mohamed/latin-to-thaana-transliteration-with-a-bidirectional-lstm-6538655afafd">Machine Learning Model.</a> Notice how words like JSC, court, supreme and MDP are transliterated correctly. The accuracy of the transliteration depends on the quality and quantity of the dataset used for training the ML model.</p><h4>Thaana Text Autocorrecting</h4><p>Example of output generated using a Machine Learning model for Thaana Text autocorrecting</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/839/1*XNPWvjwYgSP7HQdxpcZSpA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/827/1*nWDQAK0s9L469uGZ7wkeJA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*A_JW6bVIGmKhZWcsdmlFhQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/1*JlWBcpBgezH2zURS8NFQGw.jpeg" /></figure><h4>English to Thaana / Thaana to English Machine Translation</h4><p>This is something that is technically easily possible but needs a lot of data for us to get this done. What is missing is a good text corpus. We need to collect and build a dataset that will have English to Dhivehi translations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/665/1*TIuZtC5t2A63dm_M3LGs7A.jpeg" /><figcaption>Results generated from a English to Dhivehi Translation Model (2017)</figcaption></figure><p>We won’t be able to get good results unless we work on building a good dataset with lots of data.</p><h4>Dhivehi Text Classification</h4><p>This is another area of NLP which might prove to be useful for a lot of applications. Text classification and sentiment analysis on Dhivehi/Thaana text is something that will boot productivity and save a lot of time for a lot of applications.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fn_7-a0sonLA%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dn_7-a0sonLA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fn_7-a0sonLA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f67a31b2db18db781b0ef65fcbd6685f/href">https://medium.com/media/f67a31b2db18db781b0ef65fcbd6685f/href</a></iframe><p>Example use case of classification can be automatically classifying comments posted on news sites for profanity classification.</p><h4>Dhivehi Text to Speech and Dhivehi Speech Recognition</h4><p>This again like language translation needs a lot of data. If we can build big enough datasets this is something that can be easy done as well.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca865cd48768" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Maldives a Destination for the Digital Nomads ?]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/maldives-a-destination-for-the-digital-nomads-97d30a41428a?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/97d30a41428a</guid>
            <category><![CDATA[maldives]]></category>
            <category><![CDATA[digital-nomads]]></category>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Fri, 16 Nov 2018 07:34:04 GMT</pubDate>
            <atom:updated>2018-11-16T07:52:22.442Z</atom:updated>
            <content:encoded><![CDATA[<h3>Maldives a Destination for the Digital Nomads ?</h3><p>Before we talk more into this. The simple answer is NO!</p><p>Moving on who are these Digital Nomads?</p><blockquote>“A Digital Nomad is someone who is working (remotely) online. Another word for it is ‘Location Independent Entrepreneur’.”</blockquote><p>These are people who use the internet to earn a living and, more generally , conduct their life in a nomadic manner. Thus the name a “digital nomad”.</p><p>Presently some of the most popular destinations for this group of people (Digital Nomads) are places like Bali (Indonesia) and Chiang Mai (Thailand: 6,000 to 7,000 is a rough estimate of their numbers). Why is this ? Maybe it’s because of the following reasons:</p><ul><li>Perfect weather and beautiful beaches</li><li>Great and cheap service</li><li>English spoken almost everywhere</li><li>Low cost of travel</li><li>Good digital infrastructures (fast, reliable internet at cafes, restaurants and hotels)</li><li>Cheap accommodation</li><li>Nightlife</li><li>Friendly locals</li><li>Safe and peaceful environment</li></ul><p>How does the digital nomads help the locals ? Before we go into that lets gauge how much an average digital nomad makes. Its estimated that they earn somewhere around USD 1,600 to USD 20,000 per month. Some work full-time and others 15 hours a week. It too varies depending on the type of work they do and how much they earn.</p><p>So back to the question of how do they help the city/island/country. Although they are called digital nomads not all of them are constantly moving locations. A lot of them stay in one host country for months or even a year or two. Most likely they don’t even pay any tax to the host country as they are just staying as tourists. They don’t have a working visa. The pattern of spending for these digital nomads also resembles that of a foreign expatriate. They spend on accommodation, food, transport, entertainment and even gym and so on. The major differentiator between a expat and a digital nomad is the digital nomads work from cafe’s or co-working spaces.</p><p>Obviously the digital nomads don’t spend as much as the short-term tourists, but they contribute significantly to the local economy of the host countries.</p><p>In theory the digital nomads are working without a work permit and its also hard to monitor the activities (as they are not registered to work) they are almost never taxed. However there have been some cases where the digital nomads offer jobs to locals and this creates a small economy-within-an-economy.</p><p>How can the Maldives take advantage of this? Maldives lacks experienced people in most of todays digital world/economy. If we can find a way for the digital nomads to work alongside the Maldivians, this might be a good opportunity share experience and knowledge-transfer for our younger generation.</p><p>Mostly interaction between the digital nomads and the locals are fairly limited. This is due to the fact that the digital nomads work online on the internet and when needed can hire on the internet or amongst each other.</p><p>If we are able to encourage and develop ways to attract the digital nomads to Maldives this will benefit the local economy. This can be achieved by offering cheaper accommodation options for long term stay , introduce co-working spaces and develop infrastructures for such.</p><p>In addition to the above if we offer the digital nomads a long-term visa, this will be more attractive. As a way to attract this group of people (at the initial stage) maybe we should consider a tax free option for the digital nomads. Maybe give tax exempt for the digital nomads who employe local Maldivians.</p><p>On the long run this might help to create more Foreign Direct Investment (FDI) via these digital nomads, employment and most importantly knowledge transfer.</p><p>At present Maldives is <strong>NOT</strong> a preferred destination for the digital nomads, though a lot of digital nomads might fancy the idea. This maybe mainly due to the high cost of accommodation and travel. Though we have a lot of coffee shops we don’t have proper public space to work, like co-working spaces.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97d30a41428a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Latin to Thaana Transliteration with a Bidirectional LSTM]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/latin-to-thaana-transliteration-with-a-bidirectional-lstm-6538655afafd?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/6538655afafd</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Fri, 09 Nov 2018 11:16:53 GMT</pubDate>
            <atom:updated>2018-11-09T13:36:11.677Z</atom:updated>
            <content:encoded><![CDATA[<p>We already have a few good latin to Thaana/Dhivehi transliteration tools, however there are no implementations done using a machine learning model.</p><p>I think if we use a good ML model to do the transliterations, we might get better results than the present rule based transliteration tools. It is hard to get people to follow or enforce strict rules when it comes to transliteration. Though it is not given much priority; Latin is now widely used on social media and with text messaging services.</p><p>To generate good results we need a lot of data to train such a model. This applies to all ML based problems. When it comes to Thaana/Dhivehi datasets its hard to find any. To test out the solution I had to write a crawler that collected the data from a few local news websites. Most of the sites had the news headline in Latin in the &lt;title&gt; tags that matched 80% of the time.</p><p>Code for the model and dataset used is <a href="https://github.com/Sofwath/div-transliteration">here</a></p><p>Results of the test run :</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/823a97d12e160a7560549c73eac866b1/href">https://medium.com/media/823a97d12e160a7560549c73eac866b1/href</a></iframe><p>Basic code for the data scraper:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b2a85b20b00906036c00d2951e9318dc/href">https://medium.com/media/b2a85b20b00906036c00d2951e9318dc/href</a></iframe><p>Process the text:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/1d2cce160b7930a473d53e36d192b976/href">https://medium.com/media/1d2cce160b7930a473d53e36d192b976/href</a></iframe><p>Code to test the model:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/489ba44db553b53f24f2880210e64d20/href">https://medium.com/media/489ba44db553b53f24f2880210e64d20/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6538655afafd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dhivehi News Classification with Facebook’s fastText]]></title>
            <link>https://medium.com/@sofwathullah.mohamed/dhivehi-news-classification-with-facebooks-fasttext-289ee2f5064?source=rss-895c235b6d47------2</link>
            <guid isPermaLink="false">https://medium.com/p/289ee2f5064</guid>
            <dc:creator><![CDATA[Sofwathullah Mohamed]]></dc:creator>
            <pubDate>Tue, 06 Nov 2018 04:36:10 GMT</pubDate>
            <atom:updated>2018-11-06T04:38:16.016Z</atom:updated>
            <content:encoded><![CDATA[<p>fastText is a library for efficient learning of word representations and sentence classification.</p><p>It uses RNN for it’s classification task so I wanted to try and see if fastText can be used for a simple Dhivehi news sentiment analysis.</p><p>Here are some of the inputs (not included in the training data) and results:</p><blockquote>ރައީސް ނަޝީދު މިއަދު އައްޑޫ އަށް ވަޑައިގަންނަވަނީ __label__positive</blockquote><blockquote>މަރުތަކަށް އިންސާފު ލިބުން ލަސްވަނީ ކޮރަޕްޝަނާއި ނުފޫޒު ވަންނާތީ — ޑިފެންސް މިނިސްޓަރ __label__negative</blockquote><blockquote>އައްޑޫ ސިޓީއަށް ގޮނޑި އިތުރު ނުކުރަން ނިންމައިފި __label__positive</blockquote><blockquote>ހިތްފަސޭހަ ތަނަކަށް މާލެ ހެދޭނީ ލާމަރުކަޒީ ނިޒާމުގެ ތެރެއިން __label__positive</blockquote><blockquote>ފްލެޓު ލިސްޓާ އެއްގޮތަށް އާ ސަރުކާރުން ފްލެޓް ނުދޭ ނަމަ ކޯޓަށް ދެވޭނެ __label__negative</blockquote><blockquote>އަފްރާޝީމް، ރިލްވާން އަދި ޔާމީންގެ މައްސަލަ އަލުން ބަލަނީ __label__negative</blockquote><blockquote>ޕްރައިމަރީ އަކާއި ނުލައި 12 މެމްބަރުންނަށް ޓިކެޓު ދޭން ނިންމައިފި __label__positive</blockquote><blockquote>ޕީޕީއެމް ކޮންގްރެސްގައި ޢުޞޫލުތަކާ ޚިލާފަށް އެތައް ޕާޓީއެއްގެ މެންބަރުން ވޯޓް ލާފައިވޭ __label__negative</blockquote><blockquote>ރައްޔިތުންނަށްޓަކައި ނަޝީދު ވެވަޑައިގެންފައިވަނީ އާދަޔާ ހިލާފު ގުރުބާނީއެއް __label__positive</blockquote><blockquote>ސާފް ތާހިރުކަން ކުޑަވެގެން ސީގަލް ކެފޭ ބަންދުކޮށްފި __label__negative</blockquote><blockquote>ދުނިޔޭގެ އެންމެ ފުރަތަމަ ކަނޑުއަޑީގެ ވިލާ ކޮންރަޑް ރާއްޖޭގައި، އެންމެ ދުވަހަކު ހުރެލަން 7 ލައްކައަށް ވުރެއް ގިނަ ރުފިޔާ __label__positive</blockquote><blockquote>އިންތިޚާބީ ރައީސް ސޯލިހް ސައުދީއަށް ވަޑައިގަންނަވަނީ __label__positive</blockquote><blockquote>ރާޖަޕަކްސައަކީ ބޮޑުވަޒީރުކަމަށް ބަލައިގަތުމަށް ލަންކާ ޕާލަމެންޓްގެ ރައީސް ދެކޮޅު __label__negative</blockquote><blockquote>މައުމޫނާ އެކު އުހަށް އެރި ޕީޕީއެމް، ޔާމީނާ އެކު ފުލަށް __label__negative</blockquote><blockquote>ތޮއްޑޫ ވާވު މަގު ރުކުން ޖަރީވެފައި __label__positive</blockquote><p>The results are not bad but I think it can be improved a lot if we had more data.</p><p>The data is generated by scraping the mihaaru.com news site. <a href="https://drive.google.com/open?id=1NsLmqF3H8sOPUXVqwc8Yq7zRN9Sgf_5V">Here</a> is the link to dataset used.</p><p><a href="https://fasttext.cc/docs/en/supervised-tutorial.html">Here</a> is the link to fastText tutorial on how to do classification.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=289ee2f5064" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>