Skip to main content

Semantic Search and Future of Search Engines



Search engines are used very differently from how they were just ten years ago. Search products have had a rich and vibrant diversity in terms of features, use cases, evolution and audience. There is more data to be sorted through, and greater expectation to get to the most relevant data points faster. User behaviour is gradually shifting away from sorting through several search results manually to get to the information they need, to getting answers up front along with references.

Evolution

A very good summary of search engines and their evolution is presented by Vaibhav Saini, Co Founder at Signy: https://hackernoon.com/evolution-of-search-engines-how-will-search-engines-look-like-after-10-years-8996321e93d. We see how we started with search engines focused on index based text searching to search engines recently incorporating more advanced search metaphors, including image search, relevance measures such as back links and popularity, voice input and even AR.

Learning user intent

The latest buzz in search evolution is semantic search. Pandu Nayak, VP, Search at Google, aptly describes Search as being about understanding language: https://www.blog.google/products/search/search-language-understanding-bert/. The same post mentions the technique that Google uses to achieve this - BERT. The term semantic search encapsulates the understanding in the search tech world today, of the need to understand user intent; going beyond simplistic keyword matching, as was the norm the past. 



Knowledge graphs

A cursory glance at SERPs (search engine result pages) of most search engines would be a powerful evidence of this, with a knowledge graph being displayed by popular search engines on the right side of the screen. Knowledge graphs are another popular technique that big companies such as Google, Amazon and Microsoft have employed for serving user intent without significant user effort. In fact most users would be unaware of having interacted with knowledge graphs. Alexa, for example, performs very well at answering trivia questions from around the world, with knowledge graphs playing the key role behind the scenes.











Conversation

With the predominance of semantic search today, the product's shape and form has changed for the user as well. We see increasing user interaction with conversational interfaces, including chatbots, for making transactional search queries, such as checking weather, getting data point information and generally any query that has a unique high relevance result. From a user experience perspective, the question and answer paradigm feels more natural and intuitive for the user if implemented intelligently. 

Accessibility

Voice input, multi lingual interfaces and data, screen readers and haptic feedback, while useful innovations, still fall short of making tech, including search, accessible and inclusive as much as we'd like. But the problems go beyond technology. We need to also evolve business models locally and create ecosystems for them, that would give incentive to both enterprises and users to adopt these products locally and diversely. More on this in a future post.

Future

While we have seen breathtaking advancements in search tech, we're still far away from a near human semantically proficient AI. One reason for this is the massive entry barrier in this field which limits independent innovation. Most of the tech mentioned here has been developed by big shops which can afford to run TPU and GPU clusters en masse on large amounts of data without breaking sweat. The "BERT large" implementation for example, learns 340 million parameters, would take 4 days to learn on 16 TPUs if not optimised. 

Open sourcing of BERT as well as release of pre-trained models by Google, is an example of efforts by large companies to drive independent innovation in application of the underlying science. There are also efforts to be more prudent while machine learning models, to be more efficient and cater them to specific use cases as far as possible. 

There are several areas where both businesses and customers have need and expectation from semantic search to get to the most relevant information faster, ideally in a question and answer format. Use cases include e-commerce, academic research and organisational data mining to aid decision making. We can expect applications in these areas to proliferate in the coming years along with exploration of new and hybrid use cases where the journey begins at search, but doesn't end there.

Comments

Popular posts from this blog

JavaScript Closure Gotcha with Vuex

Closures in Javascript are not just an additional feature, they're inherent to the language. Every time a function is created, it's lexical scope is automatically captured in it's current state along with the function.  It can quickly get complex when using reactive paradigms, such as with frameworks like Vue , as happened with me today. Here's the code: async setSwaggerClient({ state, commit, rootState}) { ... let authToken = rootState.user.key; Swagger( '/api/swagger.json', { requestInterceptor: (req) => { req.headers["X-CSRFToken"] = g_csrftoken; if(authToken !== null) { req.headers["Authorization"] = `Token ${authToken}`; } return req; } } ) ... } The function setSwaggerClient is a Vuex action that creates and sets the swagger client in a Vuex store. During creation I pass a functi...

The autoform-select2 + AJAX data adventure

While developing a Gig creation form for Zoomout, using meteor-autoform and select2, I realized that search and add feature for adding artists to a Gig would be a problem. Select2 supports AJAX datasource for generating options based on user input, but till version 3.5 it depended on the use of a hidden input field for adding new options on the fly. The autoform-select2 package, although a great tool, did not yet offer support for hidden input fields.  #14  from the autoform-select2 issues describes this problem. Select2 recently released  4.0.0-rc.1 which thankfully removes the need for a hidden input field, but the meteor-select2 package, on which the autoform-select2 package depends, still uses 3.5.1 of select2. So till meteor-select2 releases with an upgraded version of select2, which may happen when select2 makes a stable 4.0.0 release, follow these steps to make use of an AJAX datasource in your meter autofroms: git clone https://github.com/abhishekbatra/mete...