<?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 Vishwanath Jha on Medium]]></title>
        <description><![CDATA[Stories by Vishwanath Jha on Medium]]></description>
        <link>https://medium.com/@ahamvishwanath?source=rss-ca19bdf7aa01------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*9TvCVTDYptY7HNeaz8aFdg.png</url>
            <title>Stories by Vishwanath Jha on Medium</title>
            <link>https://medium.com/@ahamvishwanath?source=rss-ca19bdf7aa01------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 10:42:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ahamvishwanath/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[Impact of COVID-19 — Is it a boon or a bane for Indian e-Commerce?]]></title>
            <link>https://medium.com/@ahamvishwanath/impact-of-covid-19-is-it-a-boon-or-a-bane-for-indian-e-commerce-6597167ea22d?source=rss-ca19bdf7aa01------2</link>
            <guid isPermaLink="false">https://medium.com/p/6597167ea22d</guid>
            <category><![CDATA[customer-service]]></category>
            <category><![CDATA[covid19]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[india]]></category>
            <category><![CDATA[ecommerce]]></category>
            <dc:creator><![CDATA[Vishwanath Jha]]></dc:creator>
            <pubDate>Tue, 07 Jul 2020 14:19:21 GMT</pubDate>
            <atom:updated>2020-07-07T19:22:36.939Z</atom:updated>
            <content:encoded><![CDATA[<h4>e-Commerce penetration in India will triple in the next 4 years to 300 Mn. At 3.4% today, businesses find Customer Service unmanageable. COVID-19 has made this multi-headed problem evident. Here we discuss the next big barrier to scale for the Indian e-Commerce, to open up a discussion with business leaders who care for that growing market.</h4><h3>Today, e-Commerce penetration in India stands only at 3.4% roughly amounting to 100 million online customers. And, it’s expected to reach 350 Million by 2025.</h3><p>The market is big enough for 2 decades, and growth will be astronomical. It is estimated that the new shopping behavior in e-Commerce across the country will facilitate three-fold revenue growth for business to $85 Billion by 2024.</p><p><strong><em>The story of e-Commerce in India hasn’t even crossed the quarter mark, and is revealing it’s own diverse challenges for forward-thinking businesses to solve.</em></strong></p><h3></h3><p>Read the most insightful report on the future of Indian e-com by @BainAlerts and @Flipkart Here is a thread with key learnings: 1. E-com penetration in India is 3.4% vs. 15-25% in US/China. 100M shoppers in India today https://t.co/vmZyqoQ0Z5</p><blockquote><strong>This mammoth growth has been largely fueled by the first-time online shoppers of the new India where every 9 out 10 new internet users are Indian language internet users.</strong></blockquote><blockquote>If you go further deeper into the order volume distribution, you’d see that <a href="https://economictimes.indiatimes.com/small-biz/startups/newsbuzz/bharat-to-haul-ecomm-gmv-past-the-100-billion-mark-by-2025/articleshow/76396865.cms?from=mdr">“every 3 out 5 orders are coming from tier 2 cities or below. Also, the price is only marginally low in tier 2 cities and below, when compared to tier 1 cities.”</a></blockquote><h3>The State of e-Commerce during Coronavirus</h3><p>The current pandemic has impacted the supply chain globally. There were only a handful of business categories which were running uninterruptedly during the government-imposed lockdown due to COVID-19.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Xw5Z-sgWqxsTuMz3" /><figcaption><a href="https://www.forbes.com/sites/neilstern/2020/04/27/e-commerce-and-grocery-this-time-its-real/#46526da65d65">https://www.forbes.com/sites/neilstern/2020/04/27/e-commerce-and-grocery-this-time-its-real/#46526da65d65</a></figcaption></figure><p><strong>Coronavirus has driven technology averse customers to go online for shopping of essential products irrespective of the cities, towns or villages they dwell in</strong>. Consumers now are indoors, more than ever, and are looking for safe and hygienic shopping experiences and e-Commerce clearly fits the bill.</p><p>Another impact of the coronavirus pandemic was a clear demarcation between essentials and non-essential which resulted in many businesses going out of favor with the administration for a long time, and many to discover alternate streams of revenue.</p><p>Above all, the fight to break into the “magic ring” during the first half of the lockdown was an interesting trend, as the regulators were only prepared to deploy their resources for emergency facilities that preserve safety and sanity of the masses, during the lockdown. <strong>This also signaled the nimbleness today’s businesses need to overhaul or augment categories they operate in, for non-uniform periods of time.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CHFgYvhthuCQpbaX" /></figure><blockquote>This rare natural adversity is slowly proving to be a bloom for e-Commerce and to expand to under-served areas like the tier 2 cities or below of India. This is largely a blessing for e-Commerce companies.</blockquote><p>Several e-Commerce frontrunners and category leaders are engaged in making the most of this business opportunity. To cope up with impacts of COVID-19, Amazon India is on a mission to enable local shops all across the country sell online, supporting them with the required technology and training among others.</p><blockquote><strong>“Myntra’s End of Reason Sale</strong> concluded in the last week of June 2020, and the key highlight of the event was integration and collaboration with <strong>400 offline stores of more than 60 brands, </strong>enabling them to go live with their inventory on the Myntra platform. It’s very interesting to know that <strong>about 3.5 million customers shopped during the sale, with over 700,000 first-timers.”</strong></blockquote><blockquote><strong>56% of them were from tier-2 cities and beyond. The top tier-2 and 3 cities include Guwahati, Bhubaneswar, Dehradun, Imphal, Aizawl and Panchkula.</strong></blockquote><blockquote>Big Basket recently hit annualized gross sale runrate of $1 billion and similarly, other e-commerce players like Lenskart, Ninjakart, Grofers, Pharmeasy, 1mg, etc. are expected to witness a rapid lift in their order volumes as well.</blockquote><h3>Problems, Innovations and Further Challenges</h3><p><strong>The growth story is always music to the ears. But, the devil’s in the details. What remains a barrier to e-Commerce, is accessibility of these platforms in Indian languages for the evolving India.</strong></p><p>Keeping this in mind, Flipkart recently launched a voice AI assistant to help increase the discoverability of products, shopping cart management and also enabled three new Indic languages on its platform — Tamil, Telugu and Kannada as part of its vision to make online commerce more inclusive and accessible for Indian language users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/580/0*So5cOPs3l8ZRjTnu" /></figure><p>Despite all these efforts, the <strong>first-time customers are facing many frequent challenges </strong>during their maiden online shopping journey. And consistent with old habits, <strong>they are preferring to call the support center</strong> for assistance during entire order fulfillment lifecycle.</p><p><strong>However, due to executive orders and necessary health compliances, the contact centers don’t have adequate number of skilled support agents to handle this sudden surge in volumes of support request. Besides the spike, the diversity of languages is fast becoming a major barrier to communication.</strong></p><blockquote>The demand for multilingual call support is natural and the entire value chain in Contact Centers from hiring to quality support is impacted by this demand. This is proving to be the curse of growth and is becoming more evident as the coronavirus pandemic runs through its course.</blockquote><blockquote>In their future of CX report, PwC surveyed 15,000 consumers and found that 1 in 3 customers will leave a brand they love after just one bad experience, while 92% would completely abandon a company after two or three negative interactions.</blockquote><blockquote>This poses the grand challenge for e-Commerce enterprises, “How to ensure consistent and remarkable customer experience as they scale in the most opportune time ever for e-Commerce?</blockquote><p>In this moment of natural business urgency, enterprises are exploring automation for query resolutions on various support channels such as website, mobile apps, WhatsApp &amp; calls to scale contact center operations.</p><p>It is important to note here that virtual assistants or chatbots as we know it in common parlance are not something new or unheard of. We have seen emergence of digital assistants or chatbots as a powerful medium to standardize the customer support function in last couple of years, at least in Indian ecosystem. Swiggy launched customer support chatbot to deal with high volume of interactions with customers, ICICI Bank has recently started WhatsApp banking to engage and assist users on their preferred channel and similarly there are multiple such examples.</p><p><strong>In order to cater to these business requests, there are several hundreds of chatbot vendors offering chatbot solution to their customers. </strong>You can estimate the fierceness of the competition in market by the mere fact that in India alone there are hundreds of vendors offering chatbot solutions built on top of existing low end chatbot development platforms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3GfBhHXz7POEjpVZ" /><figcaption>India itself must have an army of chatbots now owing to the mushrooming of similar vendors building on top of open-source or off-the-shelf chatbot platforms</figcaption></figure><p>Ideally these chatbot vendors should’ve pounced on this “once in a lifetime” opportunity and gone for the kill. Unfortunately, we have a huge “but” here, and it musters strength from the fact that despite market being inundated with various chatbot providers, chatbots are not widely adopted by enterprises as expected, even if enterprises themselves are going digital-first.</p><p>Again, few interesting questions pop up, i.e. <strong>“Why are chatbots not widely adopted by enterprises? Is it because consumers are not happy with the substandard experience of low end chatbots? What’re the factors that we need address to get this flywheel moving?”</strong>. Well in my honest opinion, there are a plenty of technological challenges which warrants urgent attention to enable a semantically coherent dialog experience over multiple turns during the conversation and <a href="https://medium.com/saarthi-ai/state-of-conversational-ai-assistants-limitations-challenges-and-solutions-38f6336a5a00?source=---------2------------------"><strong><em>I had discussed them in detail in my previous blog.</em></strong></a></p><p><strong>Having said that, I’d like to mention here two key challenges/opportunities that could be attributed to lower adoption of virtual assistants in customer support function from the consumer’s standpoint.</strong></p><p>Majority of Indians are at least bilingual speakers, thanks to our education system. However, we’ve certain amount of affinity towards conversing in our mother tongue. We frequently borrow vocabulary from different languages to frame a sentence in conversation. This might either be a limitation or a preference for words of different vocabularies.</p><p>So, voice AI assistants that understand these finer nuances, and offer truly lossless multilingual support sans transcription and translation, are the need of hour. It’s important here to underscore the fact that such quality voice AI assistants take ample data and development time as well as it comes at a relatively higher cost but so do all good things in life. You want the best for your customer, you pay the price for it.</p><p>Poor dialog experiences are evident with low end chatbots due to obvious challenges such as poor quality of transcription &amp; translation for low resource languages, lack of contextual awareness, lack of empathetic responses, inability to handle multiple intents in a single utterance, etc. Essentially, it’s all about the performance of voice AI assistants while making real conversations with customers. There’re many chatbot vendors in market which try to whitewash these challenges by quoting that it’s cheap to use translation and there’s not much significant loss due to it. However, these vendors often forget that in a generation of customer obsessed enterprises, only those who walk the extra mile to provide uniform and delightful digital experiences to the customers will win.</p><blockquote>Now let’s address the elephant in the room, “What should the customer experience team of leading e-Commerce companies do to get past the situation, and instead leverage it for their own growth?”</blockquote><p>Taking into account several factors including current market situation, I believe there’s a very strong need for call center AI agents, especially voice AI assistants, that speaks multiple languages, understands real customer problems, empathizes with them and resolves their issues. If there’s an out of the box query or maybe a very critical query, these AI assistants should simply transfer call to the limited number call center executives with smart routing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zDqrFW7cltqObG16NavOGg.png" /><figcaption><a href="https://saarthi.ai/">Enterprises seek hassle free integrations with staff and systems, and like the flexibility to run experiments.</a></figcaption></figure><p>For enterprises, it should be quick to setup, easy to utilize their existing data to train such AI assistants, and seamless to integrate them with existing call center staff and systems.</p><p>The key challenges being resolved here are the diverse array of Indic languages, unfamiliarity with digital interfaces and an assisted experience at any time or on any platform that the customer will always be grateful for. There’s also a need for robust omnichannel strategies that take into account the offline stores as well.</p><blockquote>One of the first issues exposed by the growth of e-Commerce in India was a fragmented logistics system that is unable to cope with demand. The next arm that would be falling short is the communications system.</blockquote><p>There’ll be multiple issues which might seem trivial for digital savvy consumers but may not be the same case for new Indian internet users. This is where enterprises need to walk that extra mile and build trust with the first-time online shoppers. In order to be able to do that, domain specific ontology of queries and related cognitive signals will play a massive role and it’ll take time. If you’ve to reach a certain destination tomorrow, you’ll have to start today.</p><p><strong>Irrespective of COVID19, these multilingual call center AI assistants will prove beneficial to enterprises in the long run as well. If your intention is to grab new under-penetrated markets and you’re making everything sure that you get maximum orders, you’d never want to scale your contact centers linearly as your user base grows.</strong></p><p>If you’ve multilingual call center AI assistants, you can filter a large volume of queries and manage operations with a<strong> lean contact center.</strong></p><p>COVID-19 is also a great time for companies to leverage data at deeper levels to predict the changing needs of customers and the internal workforce.</p><p><strong>Combining AI Assistants and Analytics with an agile and flexible approach to managing workforce will enable contact centers to meet demand in the future effectively even during such unprecedented volatility.</strong></p><p>On this note, I’d like to introduce you to <a href="https://saarthi.ai/e-Commerce%20Call%20Center%20AI/MAIA.html">MAIA</a> — a multilingual call center AI assistant for order fulfillment assistance. It makes semantically coherent multi turn dialogs with customers in 6 languages, i.e. English, Hindi, Kannada, Marathi, Telugu &amp; Bangla.</p><p><a href="https://saarthi.ai/e-Commerce%20Call%20Center%20AI/MAIA.html">MAIA</a> is built with domain ontology of more than five hundred unique support queries and service requests related to order fulfillment assistance, over our Conversational AI Platform. Besides, <a href="https://saarthi.ai/e-Commerce%20Call%20Center%20AI/MAIA.html">MAIA</a> comes with Analytics to understand customers better and make the contact centers more efficient.</p><p>Saarthi.ai intends to make contact centers in e-Commerce nimbler and responsive to the multi-headed demands of the Indian e-Commerce buyer.</p><p>Though the spread of coronavirus (COVID-19) cases in India had some impact on the digital ride-hailing and food delivery segments, it is more of a blessing for e-commerce companies as more users prefer purchasing goods, especially grocery, and other daily needs from online platforms.</p><p>Flipkart, BigBasket and Grofers have witnessed about a 20–30 % uptick in orders as customers are avoiding visiting crowded places like malls and supermarkets, and staying indoors.</p><blockquote>“It is partly driven by new customers coming online because they don’t want to go to the (physical) stores, and there are also existing customers who are stocking up much more products over fears of things running out,” said Vipul Parekh, co-founder of BigBasket, one of the largest online and food grocery stores.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W54nTaNAj7OjvRth" /><figcaption>Photo by <a href="https://unsplash.com/@edwinhooper?utm_source=medium&amp;utm_medium=referral">Edwin Hooper</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><strong>However, customer service is increasingly becoming the curse of such scale, as business disruptions expose weak call center systems and challenge the work habits of customer support agents.</strong></p><p>The primary reasons for India’s underperformance have traditionally been market inefficiencies and underdevelopment. This is most prominently manifested in hight return to origin (RTO) rates across India which hovered at more than 30% in the pre-COVID world, especially when cash on delivery (COD) orders are involved.</p><p>And now, the fast changing demography of online shoppers is confounding many Business Leaders of the e-Commerce and Contact Center industry. The numbers signals a wide efficiency gap in the Indian e-Commerce industry, especially in the hinterlands.</p><p>Also, there are various unpredictable circumstances that may impact the supply chain in a way that’s unimaginable. Very recently, local restrictions had cut off first-mile and last-mile logistics for e-commerce firms and effectively choked them. As per Clickpost, there was a 40% drop in the number of e-commerce deliveries across India and return to origin or RTO orders saw a massive increase by 330% on March 23, 2020 due to this.</p><p>All this sends a stark warning for the future where Indian e-Commerce companies might find themselves juggling between a rapidly growing demand, a broken logistics system, and an inefficient expensive customer communications center.</p><p>We have extensively studied the market and worked at the grass-roots level with Customer Support agents dealing with e-Commerce specific queries to design solutions to such problems.</p><p><a href="https://saarthi.ai/e-Commerce%20Call%20Center%20AI/MAIA.html">MAIA, our omnichannel virtual assistant </a>for e-Commerce, was born out of this approach. We combined domain expertise with hands-on experiences of the Contact Center environment and the different processes involved, to ensure all stakeholders are seamlessly served.</p><p>If your business is receiving large volume of queries on different support channels and you wish to provide your diverse user-base with a truly unified omnichannel experience, we can collaborate with you to provide a solution. In less than 2 weeks, we’ll integrate a solution for 15% of your queries and present a roadmap to establishing a lean multilingual contact center. Take action now, scale at pace and reimagine the future.</p><p>Please send in your comments and questions below or write to us at <a href="/service/mailto:hello@saarthi.ai">hello@saarthi.ai</a>, and <strong>I’ll be glad to have a conversation.</strong></p><blockquote><em>I regularly write here and tweet at </em><a href="https://twitter.com/AhamVishwanath"><em>@</em>AhamVishwanath</a> <em>on advancements in Conversational AI, Contact Center Technology and e-Commerce, and would appreciate hearing your thoughts.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6597167ea22d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[State of Conversational AI Assistants: Limitations, Challenges and Solutions]]></title>
            <link>https://medium.com/saarthi-ai/state-of-conversational-ai-assistants-limitations-challenges-and-solutions-38f6336a5a00?source=rss-ca19bdf7aa01------2</link>
            <guid isPermaLink="false">https://medium.com/p/38f6336a5a00</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[contact-center]]></category>
            <category><![CDATA[virtual-assistant]]></category>
            <category><![CDATA[conversational-ai]]></category>
            <category><![CDATA[chatbots]]></category>
            <dc:creator><![CDATA[Vishwanath Jha]]></dc:creator>
            <pubDate>Tue, 31 Mar 2020 10:34:33 GMT</pubDate>
            <atom:updated>2020-04-01T08:43:52.410Z</atom:updated>
            <content:encoded><![CDATA[<h3>State of Conversational AI Assistants: Challenges and Solutions</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RZRwqE5n35mbMbeK" /><figcaption>State of Conversational AI in 2020: Limitations, Challenges and Solutions</figcaption></figure><p>The likes of Alexa, Google Assistants, Cortana, etc. of the world have brought first-hand exposure to Conversational AI Assistants. Increasing adoption in contact centre and customer support automation has provided a certain level of awareness among the masses.</p><p>It’ll not be preposterous to say Conversational AI(CAI) Assistants have pervaded in our everyday lives and user adoption is increasing as we speak.</p><h3>A Brief on Conversational AI Assistants (CAI)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*T8xAHnIfPjyjvag8" /><figcaption>Photo by <a href="https://unsplash.com/@rocknrollmonkey?utm_source=medium&amp;utm_medium=referral">Rock’n Roll Monkey</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Today I am sharing with you all a brief introduction to CAI (Conversational AI Assistants), as in what are they, how are they different from one another, what are the major challenges, and how researchers of NLP technology are approaching these challenges.</p><h4>Types of Dialogue or Conversational Agents:</h4><p>Conversational AI assistants and AI Chatbots, are also known as Dialogue Agents or Dialogue Systems can be broken down into three major categories:</p><blockquote>i. Knowledge Base QA(KBQA) systems,</blockquote><blockquote>ii. Goal-Oriented Dialogue Systems(GODS) and</blockquote><blockquote>iii. Social Chatbots.</blockquote><p><strong>KBQA — KBQA: </strong>These systems are built to answer the user queries such as, “What’s the temperature”, “When did India get independence?”, wherein they fetch knowledge from a prebuilt knowledge base and render the same in an output format understandable by users.</p><p><strong>GODS — GODS:</strong> They are designed to accomplish specific goals or tasks such as movie booking, restaurant reservation or customer support agents, etc.</p><p><strong>Social ChatBots: </strong>Objective of social Chatbots is to converse inexorably about almost anything under the sun and beyond, such as Mitsuku.</p><h3>What is the Challenge for AI Assistants and AI Chatbots?</h3><p>Today, what remains a challenge for the AI Assistants is their inability to make good conversations over multiple turns across languages. It is frustrating for a user to repeat himself, even in English, and engage in a long, repetitive flow that otiose in nature.</p><p>Do you wanna know why it happens?</p><h4>Failure in Designing Goal-Oriented Conversations</h4><p>We have failed to understand that making goal-oriented conversations is an optimal decision-making process.</p><p>In an ideal situation, dialogue agents would always have access to previous state information which get updated at every turn and make the most optimal decision.</p><p>However, many AI chatbots we see today lack these well-defined belief states and thus, they treat each user queries independently (instead of contextually), or approximate some rule engines.</p><p>Before I get to the current challenges, let me briefly talk about <strong>how we measure the performance of different CAI Assistants.</strong></p><blockquote><strong>Conversation-turns per session(CPS) </strong>is one metric that we observe to evaluate and measure the efficiency of GODS &amp; Social Chatbots.</blockquote><blockquote>However, objectives are different in GODS &amp; Social Bots — the former strives to minimise CPS, i.e complete the goal via shortest path &amp; the latter tries to maximise CPS for ever-increasing social engagement.</blockquote><blockquote>We also measure the responses generated by CAI Assistants in terms of their relevance &amp; sensibility with respect to user’s input.</blockquote><h3>The 6 Roadblocks Conversational AI needs to Overcome</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4zht8pwqPRb1qUdH" /><figcaption>Photo by <a href="https://unsplash.com/@pietrozj?utm_source=medium&amp;utm_medium=referral">Pietro Jeng</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Coming to the challenges, there are six major roadblocks that need to be overcome to build a really good Conversational AI Assistants</p><h4>1. Speech to text transcriptions:</h4><p>Transcription in low resource languages is a well-known research problem which had been actively pursued over the decades. Furthermore, in a multilingual environment where queries are weaved using vocabulary from multiple low resource languages pose bigger challenges.</p><p>Dialects and colloquialism add more dimensions to this problem. Repetition and stutter are also challenges that remain unsolved in most of the languages. Due to these inefficiencies, we incur a significant loss of information which gets further propagated to the NLU.</p><h4>2. Machine Translation Systems</h4><p>We have made really good advances over the last decade in machine translation but languages are evolving much faster than technology. We’re yet to achieve a state wherein we can rely on machine translations in Conversational AI.</p><p>The information loss incurred in translating to any high resource languages for better natural language understanding worsened the situation of NLU (Natural Language Understanding) engines.</p><p>Translation of code mixed inputs to one particular language has also been a great challenge for researchers from across the globe.</p><h4>3. Goal and relevant belief state identifications</h4><p>Defining the dialogue space of an agent and identifying the shortest path for each goal along with its intermediary states have been ignored for long. This requires a deep understanding of the business process and most probable comebacks from the users.</p><p>It’s important to note that NLU/SLU (Natural or Spoken Language Understanding) alone can’t give a multi-turn conversational experience unless the dialogue manager is well thought of, and can explore and exploit novel journey through states from historical episodes.</p><h4>4. Lack of Contextual Awareness</h4><p>This has to be the most common pitfall amongst even the best “AI” chatbots. Let’s say if I ask,</p><blockquote>“Show me tickets for the movies running near me”</blockquote><p>and bot successfully fetches a list of options and displays in front of me. Later, if I say,</p><blockquote>“Book the first one, this looks interesting”,</blockquote><p>there is no reference to any movie or ticket in my second query. If the bot never knew in what context the second query had been spoken, it’ll never be able to complete the task.</p><h4>5. Handling non-goal &amp; chitchat related states</h4><p>It’s obvious and very natural for users to ask out of domain queries or start small-talk in the middle of a task. Here, handling the happy path is fine but what happens when you digress? The bot breaks!</p><p>Let’s say I’m in a conversation with a chatbot to book a movie ticket and have navigated in the conversation to the seat-selection step. Here, if I ask, “Which is your favourite colour?”, ideally bot should say, “Apologies, I can only help you with movie booking. Can we please get back to your seat selection for “XYZ” movie?”.</p><p><strong>Knowledge about which questions should be answered and which ones are to be left is very important and comes through contextual awareness.</strong></p><h4>6. Cognitive Intelligence (Emotion, intonation, sarcasm, hate speech, etc. )</h4><p>Today most of the chatbots have a flat and bland response. They lack empathy and emotion and are not able to pick sarcasm which makes them conversationally less friendly. No matter if a customer is angry or pleasantly surprised, the bot will throw the same templated response.</p><h3>So now, what’s the solution or the near solution?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MiazDATyF-09lifz" /><figcaption>Photo by <a href="https://unsplash.com/@neonbrand?utm_source=medium&amp;utm_medium=referral">NeONBRAND</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4><strong>1. Using Speech Signals and Transcriptions to identify Intents &amp; Slots —</strong></h4><p>I must admit it’s a very interesting problem and is one of my research areas. We’ve found some success with Spoken Natural Language Understanding of the knowledge base question answering agents. However, this success needs to be replicated in multi-domain &amp; multi-turn conversations with GODS. I strongly believe that 2020 will show us the light and push us forward.</p><h4><strong>2. Eliminating Translations — Using Native Language Understanding</strong></h4><p>Having worked on this problem for a couple of years, I can say that this has been solved by a few, and is ready to be disseminated further.</p><p>When we observed this information loss and corruption of input data upon translation, we tried to remove this unnecessary step and built native language understanding systems in 23 languages, which didn’t require any translation to high resource languages. We have, since then, established state of the art results in each of these languages.</p><h4><strong>3. Treating goal-oriented conversations as Markov Decision Processes(MDP) and Partially Observable MDP (POMDP)</strong></h4><p>This is an important area of research which has seen massive traction over the last couple of years. MDP can be thought of like the shortest path with fixed belief states for any specific goal completion. Also, this helps in defining the search space at every state post-goal identification and thus, improves natural language understanding. In conversations where user deviates from fixed path, i.e. ask out of domain question or give random input, dialogue can be viewed as a Partially Observable Markov Decision Process.</p><h4>4.<strong> Encoding contexts derived from the episodic conversations and sharing it with next state to facilitate accurate intent, slots and state identification.</strong></h4><p>If we observe the nature of conversations, we would find a context around which dialogue happens, and this context remains latent at many turns. Identifying this latent context and sharing it with the next state helps CAI Assistants take more informed decision and thereby reduces errors due to coreference resolution and entity linking problems. This also helps in handling journeys where the user switches context and jumps to another goal.</p><h4><strong>5. Dialogue State Trackers (DST) &amp; Intelligent Dialogue Policies</strong></h4><p>Another interesting area where reinforcement learning has a massive role to play. Dialogue agents designed to explore new paths and to exploit the knowledge in a continuous learning environment have seemed to perform better. I’ve been personally very curious to study this problem. We’ve another area of research in our SALT lab and seen some really good delta over the last few months in reducing CPS for GODS.</p><h4>6 <strong>Encoding cognitive signals from previous turns and share them with next state (</strong>Same as 4)</h4><p>This will augment CAI Assistants in adapting their existing responses into far more nuanced responses based on cognitive signals such as emotions, intonations, abuse, etc.</p><p>I hope this blog helped you get an overview of Conversational AI Assistants in its current state and where it is headed towards in future.</p><p>If you have questions on Conversational AI Assistants or you want to automate your customer service, please don’t hesitate to share your queries. I’m sure I’ll have a few cents to offer you ;-)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=38f6336a5a00" width="1" height="1" alt=""><hr><p><a href="https://medium.com/saarthi-ai/state-of-conversational-ai-assistants-limitations-challenges-and-solutions-38f6336a5a00">State of Conversational AI Assistants: Limitations, Challenges and Solutions</a> was originally published in <a href="https://medium.com/saarthi-ai">Saarthi.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Startup Ecosystem of Indian Languages & Speech Technologies — Challenges & Opportunities]]></title>
            <link>https://medium.com/saarthi-ai/startup-ecosystem-of-indian-languages-speech-technologies-challenges-opportunities-998327408de6?source=rss-ca19bdf7aa01------2</link>
            <guid isPermaLink="false">https://medium.com/p/998327408de6</guid>
            <category><![CDATA[narendra-modi]]></category>
            <category><![CDATA[language]]></category>
            <category><![CDATA[speech-recognition]]></category>
            <category><![CDATA[startup-india]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Vishwanath Jha]]></dc:creator>
            <pubDate>Tue, 16 Jul 2019 16:54:48 GMT</pubDate>
            <atom:updated>2020-05-01T08:51:32.046Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Challenges and Opportunities in Startup Ecosystem of Indian Languages &amp; Speech Technologies</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JBO09qNo3XgARGA-" /><figcaption>Photo by <a href="https://unsplash.com/@dylandgillis?utm_source=medium&amp;utm_medium=referral">Dylan Gillis</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Over the last decade, India’s cities have been transformed into startup hubs, with Bangalore at the centre of this transformation. The president of India, Shri. Ram Nath Kovind, recently mentioned in a joint sitting of both parliamentary houses, that the government has a goal of establishing 50,000 startups by 2024.</p><p>Cognate with this transformation is a shift in digital demography with the rise of vernacular users, and the growing popularity of speech. It brings a pressing need for evolution of Indian language technologies and speech, and with it, to accept both the challenges and opportunities on our journey towards it.</p><p>In support of this objective, Ministry of Electronics &amp; Information Technology, FICCI-Indian Language Internet Alliance, and NITI Aayog have been working tirelessly since October last year. Numerous workshops, conferences &amp; dialogues have been organized to understand the challenges &amp; opportunities for both industry &amp; academia alike across multiple research problems focusing vernacular languages &amp; speech technology.</p><p>Having been an active part of these exhorting dialogues, particularly as an industry stakeholder, I’m bullish about the growth of startups in language and speech technologies.</p><p>So, what are the big challenges for this ecosystem, in order to make the most out of this opportunity?</p><p>Let me lay down it for you one by one.</p><h4><strong>Lack of Large-Scale Gold Standard Research Corpora</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Td6BA24S00B36POk" /><figcaption>Photo by <a href="https://unsplash.com/@alfonsmc10?utm_source=medium&amp;utm_medium=referral">Alfons Morales</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Languages and speech technology, by its very nature, require a huge amount of data. To put things into perspective, developing an industry-standard speech to text engine requires at least 50,000 hours of audio data, as well as its transcription.</p><p>To make things further difficult, annotating audio file incurs a cost of 200 USD per hour approximately. Young startups generally can’t afford to collect a humongous amount of data and transcriptions on their own.</p><p>The government should step in by making Gold standard annotated research corpora available to startups, and help them solve novel tasks that are shared, for the greater good of language and speech technology.</p><h4><strong>The high Cost of Computing Infrastructure</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vOl68-lJLWpZGvdz" /><figcaption>Photo by <a href="https://unsplash.com/@alexkixa?utm_source=medium&amp;utm_medium=referral">Alexandre Debiève</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>On top of the challenge in data collection and annotation, computing infrastructure required to work with these humongous datasets, pose another roadblock for startups working on language and speech.</p><p>Specialized high-performance GPU’s are one of the basic necessities of computation, however, these chips are profusely expensive. Startups can’t win this uphill battle on their own. Government, here, should take this cost-lid off the cauldron, and facilitate opportunities for startups to excel.</p><p>Startups should be given a monthly credit of 500,000 INR, to use AWS or Google Cloud , and have a long-term goal of setting up a government’s own computing clusters.</p><h4><strong>Siloed Resources</strong></h4><p>There are many institutions like C-DAC, TDIL, research labs across institutes of eminence like IITM, IITB, IISc, AU, etc. where significant work has been done in languages and speech technology over the years.</p><p>However, accessing those resources is extremely tedious, or should I say, impossible due to lack of canonical interfaces, and red tape involved in getting access permissions. The Government should consolidate all such relevant resources into one single platform, and make it open for the startups to play with and build upon.</p><h4><strong>Lack of Standard Industry Benchmark</strong></h4><p>Language and speech are emerging fields and hence, there is an absolute necessity to establish the credibility of the products offered by different startups. Here, quality assurance becomes a significant duty of the government to protect market infiltration with substandard services.</p><p>This can be achieved by setting up a public leader-board for multiple shared research tasks. This will also spur active research in vernacular languages and speech, thereby generating further demands for a native research conference.</p><p>Financial grants should be given to startups pro-bono. All this will ensue into customer enterprises developing a greater trust in a self-sustaining native ecosystem.</p><h4><strong>Open Innovation Projects &amp; Grants</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*szCguZ1Zeufpy4rM" /><figcaption>Photo by <a href="https://unsplash.com/@jdiegoph?utm_source=medium&amp;utm_medium=referral">Diego PH</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>To further up the ante, the government should start engaging startups through open innovation projects and work as distribution channels for the products which startup develops in public sector enterprises. This step will eventually solve the biggest problem for startups, i.e. “Distribution”, thereby, giving birth to many promising enterprises in the long run.</p><p>The government can avail the services rendered by startups at discounted rates as quid pro quo. This will spur competitiveness in pricing, leading to a standard pricing strategy gradually.</p><h4><strong>Lack of More Degree Programs in AI &amp; Computational Linguistics</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*66z8xDKPE_D7DozI" /><figcaption>Photo by <a href="https://unsplash.com/@aaronburden?utm_source=medium&amp;utm_medium=referral">Aaron Burden</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Another hurdle is a lack of skilled resources for this industry at large. More academic programs should be started countrywide, offering courses and degrees in the area of Artificial Intelligence with a deep focus on language &amp; speech technologies.</p><p>Right now, we as entrepreneurs, have to incur the extra cost and precious time in grooming the resources, to equip them with the right skills. These programs should be research projects are driven rather than GPA driven. This academic and pedagogical shift is much needed if we’ve to race ahead of giants like China &amp; USA.</p><h4><strong>National Emerging Technology Accelerator</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vLUwiWh7KbT-B8RU" /><figcaption>Photo by <a href="https://unsplash.com/@ninjason?utm_source=medium&amp;utm_medium=referral">Jason Leung</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Despite the government’s push to imbue an entrepreneurial mindset among youth, the services and policies are largely inaccessible and simply insufficient.</p><p>If we’re to believe our president’s ambitious goal of having 50,000 Startups by 2024, it has to be backed by substance and not by humbug. It should be the government’s foremost duty to identify the startups and facilitate basic infrastructure such as office spaces and other peripherals, to begin with.</p><p>And as the current government is very much in love with acronyms, I propose an Accelerator, NETA — <strong>N</strong>ational <strong>E</strong>merging <strong>T</strong>echnologies <strong>A</strong>ccelerator, for the very job headed by the amalgamated body of industry, academia &amp; government.</p><p>I acknowledge that this list might not be exhaustive, and therefore, I urge other entrepreneurs to voice their opinions on this piece, so we’re heard out loud and clear, and are subsequently involved in designing the blueprints of the exciting future that awaits on the other side.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=998327408de6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/saarthi-ai/startup-ecosystem-of-indian-languages-speech-technologies-challenges-opportunities-998327408de6">Startup Ecosystem of Indian Languages &amp; Speech Technologies — Challenges &amp; Opportunities</a> was originally published in <a href="https://medium.com/saarthi-ai">Saarthi.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>