<?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 Shubhashis Sengupta on Medium]]></title>
        <description><![CDATA[Stories by Shubhashis Sengupta on Medium]]></description>
        <link>https://medium.com/@shubhashis.sengupta?source=rss-9e6fab1a6c5e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*d4IrO8u-ufv6xk3E.</url>
            <title>Stories by Shubhashis Sengupta on Medium</title>
            <link>https://medium.com/@shubhashis.sengupta?source=rss-9e6fab1a6c5e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 19:59:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shubhashis.sengupta/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[Does your Chatbot really understand you? Part II]]></title>
            <link>https://medium.com/@shubhashis.sengupta/does-your-chatbot-really-understand-you-part-ii-143d639e745d?source=rss-9e6fab1a6c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/143d639e745d</guid>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <dc:creator><![CDATA[Shubhashis Sengupta]]></dc:creator>
            <pubDate>Sun, 08 Sep 2019 06:38:08 GMT</pubDate>
            <atom:updated>2019-09-08T06:41:12.343Z</atom:updated>
            <content:encoded><![CDATA[<h3>Does your chatbot really understand you? Part II</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*f0Io_AOSJAqgf9Uw" /><figcaption>Photo by <a href="https://unsplash.com/@franckinjapan?utm_source=medium&amp;utm_medium=referral">Franck V.</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In the <a href="https://medium.com/@shubhashis.sengupta/does-your-chatbot-really-undertands-you-part-1-9043f1a11275">previous blog</a> of the series, we discussed some of the current challenges being encountered to develop a truly smart BOT. In essence, among many others, there are three fundamental issues — (a) today’s BOTs do not comprehend the user utterances deeply and meaningfully, (b) today’s BOTs can’t reason over domain data and axioms and they also do not even possess a modicum of common sense; and therefore, (c) today’s BOTs can’t generate a truly relevant, contextual and empathetic answer.</p><p>In other words, today’s BOTs are not humanized or anthropomorphic. As a result, while BOTs are being used for performing tasks and commands — such as playing a song or fetching your bank balance or even suggesting a recipe; you wouldn’t yet have a BOT as a companion to discuss current affairs or marital discord.</p><p>Let us look at each of the issues stated above. In the previous blog, we have given anecdotal examples for each type. Now, we will delve a little deeper technically and try to find where research is heading and what we are doing at Accenture Labs to answer some of the issues.</p><p>First up, the BOT should understand not just what you say (lexical understanding) but also what you mean (semantic understanding). It should distinguish between first <em>leg</em> of a flight itinerary, first <em>leg</em> of a football tournament and front <em>leg</em> of an animal. It should distinguish between the actions of <em>release</em> of a movie and <em>release</em> of a prisoner. Such varying interpretation of the same word depending on context (also known as word polysemy) is an absolute minimum. Interpretation or representation of meaning of text is an exciting research area in AI and techniques from SynSet lookup to neural text embeddings are widely being used in this context. Additionally, new research in Abstract Meaning Representation (AMR) of an utterance through semantic parsing techniques to learn categorical grammar and computation will help to create a computed algebraic meaning abstraction on the utterance that is fully machine interpretable (machine can now capture the full semantic meaning of words, actions and dependencies on any utterance about a domain). The second problem in language understanding is pragmatics — the liberal use of references and contexts in our conversations. Finding out previous references for a mention, linking to a correct mention and resolving ambiguity not only on the entities but also on the property of the entities is a complex challenge.</p><p>In our previous blog, a fictional user asks the airline BOT: <em>My violin case is kind of big. I am flying coach class. Do you think I can place it below the seats there? </em>There are two tricky ambiguous tasks lurking in this apparently simple set of statements. The first task is called co-reference resolution in pragmatics. The AI system must understand that “it” refers to the violin case and “there” refers to the space below as seat in an economy class of the airplane. The second task falls under the general scheme of what we call the Winograd Schema Challenge. The AI system has to understand that the violin case is “bigger” and it may not fit under a “smaller” seat. While part of the second task (and therefore, the answer) of reasoning that a big thing may not fit in a small space can be tackled separately, the mere task of linking the violin case to something big and area under seat to something small and creating a logical comparative question is a tough task by itself. Good news is that considerable progress to understand pragmatics have been made already.</p><p>The second thing is about knowledge and reasoning. As we indulge in our day-to-day conversations with our colleagues, friends, tax planner, cabbie or dentist — we rely a lot on our shared background, domain, past interactions and current context. In a machine, such context may be stored in a Knowledge Graph that contains domain details in an entity-relationship associative memory, past interactions in a temporal or episodic memory and current context in a working memory. Apart from fetching answers to your questions (like your dentist) or giving recommendations on tax savings (like your tax planner), and prepping you on New York state motorways (like your cabbie); the BOT should also have enough common sense to solve a basic level of New York Times Crossword puzzle. Or it should at least be able to reason that you can carry grain in a cotton sack but not water.</p><p>The furious pace of advancement in knowledge graph holds out hope for us. Large graph databases are already becoming systems of choice for storing large relationship graphs where entities are nodes and relationships between them are arcs. Distant- and semi-supervised learning techniques are increasingly being used to acquire and curate large domain ontology instances from diverse textual data. Knowledge Graph Embedding (KGE) techniques for both entity-relationships and temporal episodes are being used for knowledge representation. There has also been a conscious move towards agglomeration of knowledge graphs and social network analysis — wherein the entity-relationships and social connections can be combined to produce powerful insights. Sometimes, simply walking through graph will be a nice idea to see if a relationship link exists between any two entity nodes.</p><p>For example, to answer the question <em>Can we take a stroller on-board for my 2 year old daughter?</em> the AI can do a graph-walk to see if a 2 year old (or passenger below a particular age group) is associated with a special baggage and if a stroller <em>is</em> a special baggage type. A system based on logic propagation (also called forward or backward chaining system) can traverse across the nodes in a graph to establish the link and answer the question.</p><p>To tackle the third challenge, the BOT must have domain and common-sense reasoning capability. Domain reasoning techniques on knowledge are typically being done through symbolic logic, statistical relational learning and various deep representational learning techniques. On the other hand, common sense reasoning is a tough ask from AI perspective; and, as of now, researchers are using Upper Ontology based approach with Natural Language Inference and Textual Entailment. An upper or general-purpose ontology, such as OpenCyc, Sumo, Freebase or YAGO capture a very large number of taxonomical and ontological relationships among common real-world concepts and entities — something a conversation BOT can make use of for answering a general knowledge question or for reasoning over Known Facts. Natural Language Inference and Entailment techniques work over Premise-Hypothesis pairs, where the machine try to predict a plausible hypothesis given a premise.</p><p>In one of the examples given in our previous blog, the user asks the airlines BOT — <em>Can I bring a metal crutch with me? </em>To which, if the user is allowed, a BOT must respond in a humane way — <em>Do you need any assistance while boarding?</em> Underneath, the AI system matches the user question to a Premise — <em>A invalid or aged person will require a walking prop</em> (the taxonomy graph tells the AI reasoner that metal crutch is kind of a walking prop and a PNR check will tell AI whether the person is aged or has known physical deformity) and arrives at the Hypothesis <em>The passenger may be invalid or aged.</em> The AI system can then simply look up the related business rules to find the rule: <em>An invalid or aged passenger may be offered assisted boarding </em>and generate appropriate question through a natural language generation system. You can see various underlying systems at play here!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=143d639e745d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Does your chatbot really undertands you? Part 1]]></title>
            <link>https://medium.com/@shubhashis.sengupta/does-your-chatbot-really-undertands-you-part-1-9043f1a11275?source=rss-9e6fab1a6c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/9043f1a11275</guid>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Shubhashis Sengupta]]></dc:creator>
            <pubDate>Sat, 31 Aug 2019 04:17:16 GMT</pubDate>
            <atom:updated>2019-09-08T06:40:50.199Z</atom:updated>
            <content:encoded><![CDATA[<h3>Does your chatbot <em>really understand you? Part 1</em></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6S0KbEpGhcZQ2giL" /><figcaption>Photo by <a href="https://unsplash.com/@thebeardbe?utm_source=medium&amp;utm_medium=referral">Filip Bunkens</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>We live in the era of Virtual Assistants and, according to some pundits, human interactions with VAs in some shape and forms are set to eclipse human-human interactions in not so distant future [Ref: <a href="https://www.monster.com/career-advice/article/siri-or-spouse-bots-1116">https://www.monster.com/career-advice/article/siri-or-spouse-bots-1116</a>]. As a savvy user of the personal assistant in your latest smartphone, you may get instant replies, tips and suggestions about the places around you, reminders of your chores and calendars, switch on appliances, get recommendations on things to do, see and eat and many other common tasks. But is your personal virtual assistant really smart? Does it really understand the hidden meaning and implications of your utterances? Can it provide a reasonably satisfying or intelligent reply to a query that would require basic commonsense and domain understanding? Alas, the answer is a resounding “No”!</p><p>Let’s take the case of a chatbot that handles customer queries on baggage handling for an airline. (By the way, inquiries on baggage allowance, routing, lost or damaged baggage are the second most common queries in aviation) to illustrate our points.</p><p>Consider the following query where a customer asks –</p><p><em>(1)</em> <em>Can we take a stroller on-board for my little one?</em></p><p>This is a sort of logical query with implicit dependencies where the BOT has to understand that the accompanying passenger is a child (little one) — most possibly an infant or toddler; confirm the age of the child; check for the business rule if a stroller can be allowed for a passenger of that age; and optionally, may have to check on the flight leg (many times the flight leg determines the type of aircraft in service) and class of the passenger (business or economy) and then arrive at an answer.</p><p>Clearly therefore, a lot of common sense and domain and process knowledge get into answering an apparently simple (to human) question and today’s technology of handling conversation through dialog flows does not help at all [Ref: <a href="https://docs.microsoft.com/en-us/azure/bot-service/bot-service-design-conversation-flow?view=azure-bot-service-4.0">https://docs.microsoft.com/en-us/azure/bot-service/bot-service-design-conversation-flow?view=azure-bot-service-4.0</a>]. If one has to explicitly write branches of dialog responses for all possible query combinations of classes and passenger ages — it may soon run into hundreds of them and BOT developers will loathe at the prospect of handling something like that!</p><p>Let us now look at example of another query: answering which would require intuition and empathy on part of the airline BOT to create a positive experience for the customer.</p><p><em>(2)</em> <em>Can I bring my metal crutch along with me?</em></p><p>The expected answer is a “no”, due to obvious security restriction; or an “Yes” — depending on whether the passenger has a medical condition or infirmity or age-related issue which would allow an exception. This may be complex enough to model in today’s dialog management systems. However, it will be really nice if the answer is “Yes” and the BOT adds a thoughtful rejoinder –</p><p><em>Do you need any assistance during boarding?</em></p><p>For a human service agent, it is almost intuitive to do so, but how will you train a BOT to say something like that for a question that is linguistically so different?</p><p>To an AI practioner, the above is a classic case of Natural Language Inference [Ref: <a href="http://nlpprogress.com/english/natural_language_inference.html">http://nlpprogress.com/english/natural_language_inference.html</a>]; where, if, the answer to utterance (2) is true then the rejoinder must logically follow. This is something today’s BOTs fail to provide.</p><p>Let’s look at one final example.</p><p><em>(3)</em> <em>My violin case is kind of big. I am flying coach class. Do you think I can place it below the seats there?</em></p><p>We will wager a bet that most of the Natural Language Understanding systems in today’s BOTs will not be able to understand the query. There are two tricky co-referential connections to be resolved here — “it” should refer to the violin case and “there” should refer to the coach class seats. This is a classic example of another type of machine intelligence problem commonly known as Winograd Schema Challenge [Ref: <a href="https://en.wikipedia.org/wiki/Winograd_Schema_Challenge">https://en.wikipedia.org/wiki/Winograd_Schema_Challenge</a>]. Hard as it may be to the machine, in our daily conversations (we have a name for it — it is called pragmatics), we resolve such (and much more complex) references pretty easily.</p><p>We can go on giving such “smartness” tests to the BOTs and these need not just be around language understanding and query answering. In case the airline misplaces your luggage, is the BOT smart enough to look at your literary and deliver the luggage to the final destination? If your flight gets delayed, can your personal assistant go ahead and reschedule that crucial client meeting?</p><p>The problem lies in the way today’s chatbots are designed. These are designed as flows of dialogs or scripts. But life is not just a script. Many tasks that we accomplish today with relative ease are done with implicit domain knowledge, contextual / situational awareness and commonsense reasoning. Sadly, we do not have such knowledge representation and reasoning capabilities in today’s conversational AI systems. Deep Learning and other systems are good in predicting response sequences, but the link to the semantics is often missing.</p><p>All these have prompted many researchers to start conjuring a system that merges domain knowledge captured in a Knowledge Graph and reasoning thereof with conversational flow. They will need to work on mining enterprise and linked data to curate enterprise knowledge graphs to understand key entities and their linkages; and develop complex query and reasoning systems on the knowledge graphs and weaving this domain knowledge seamlessly in the context of dialogs. We can take care of many of the situational complex queries as mentioned above.</p><p>Exciting days are ahead. Stay tuned to the <a href="https://medium.com/@shubhashis.sengupta/does-your-chatbot-really-undertands-you-part-1-9043f1a11275">next installment</a> of this blog where we discuss some of the details!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9043f1a11275" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>