Relevance, Like Beauty Lies In The Eyes Of The Beholder

Consider this: I want to cook green curry for the first time. Hence, I have a need, a need for information to help me prepare the dish. I first get onto a recipe search site and express my need using the keywords “green curry”. The search engine returns about 1,200 results and I interact with them and revise the keywords until I get what I was looking for. In this case, I save and print the seventh, tenth and twelfth recipes from the results. Based on this example, one might look at a recipe and say that it is relevant to the need if it contains the input keywords or words that are closely related to them. But does relevance amount to just that; the words we can express in a search box? For many small to medium sized companies which offer search on the periphery or are overly reliant on off-the-shelf search solutions for their product offerings, the answer is often yes, relevance is simply a what-you-give-is-what-you-get, one-size-fits-all concept.

Photo by Kat J on Unsplash

Relevance is contextual

The topic of relevance, which predates modern search engine by at least 60 years, remains today an area of rigorous study for a good reason. Relevance was and still is not a well understood concept. One thing which is clear, however, is that there is more to relevance in search than just keywords. To illustrate this, consider the “green curry” example again. Assuming that all the retrieved results contain the keywords “green” and “curry, then what made the seventh, tenth and twelfth recipes stand out for me? The answer is quite simple; context. In this particular scenario, in addition to just my keywords “green” and “curry”, there were other things in my mind when I was going through the results which were either too inconvenient or impossible to express to the search engine explicitly. For one, I have a preconceived preference of what makes good recipes; they have to be simple recipes with as few ingredients and cooking steps as possible. The same result may not appeal to you if all you are after are recipes from three hatted chefs.

The reality is relevance has always been contextual, and topicality is just a small part of the whole. Location is another common type of context across Web and vertical search. If you look for “green curry restaurant in Melbourne”, any results about restaurants which are not in the Melbourne location are then immediately not relevant. Other types of context include popularity and freshness of documents, and preferences, social connections and circumstances surrounding the users. In our earlier “green curry” example, my preference as a user for simple recipes became the final deciding factor of what was relevant. In this case, the characteristics of the recipes which I have downloaded, saved or printed from my past searches can be extracted and used as context in relevance. However, the absence of pre-canned search solutions that would work with the wide variety of contexts across markets and verticals makes the topic potentially expensive to explore. For companies where search is at the core of their business, investment is necessary to customise search solutions that are highly tuned for their marketplaces.

What are the types of context?

The leading players across the Web as well as the employment search industry are increasingly experimenting with beyond topical context to offer more personalised results in the hope of improving relevance and subsequently ROI. Amit Singhal, a Senior VP at Google acknowledged that there is “…tremendous potential to make search better by understanding what you care about”. Generally, the types of context can be grouped into query-dependent, user-dependent and global. Query-dependent contexts are based on what we can do with explicit user inputs, and they can appear in the form of keywords, drop-down selections or even clicks on maps. User-dependent and global contexts are the more implicit types of context. User-dependent contexts pertain to specific users, and can range from preferences discovered from individual user behaviours onsite and circumstances uncovered from candidate profiles, to the current locations of users from GPS-enabled devices. Global contexts, on the other hand, are independent of individual queries and users. They are often to do with the information that search is serving up such as the freshness of documents, and the popularity based on the number of times a document has been viewed.

Relevance in job and talent search

In the employment marketplace, it is not difficult to imagine how these different types of context can equally apply. Imagine that you are looking at a ranked list of job ads for the query “project manager”. A job ad that a candidate finds relevant may not appeal to others. For one, if you specialise in managing projects in the ICT industry, the results that you are interested in will be different to the relevant results for another candidate in the building and construction industry. All is well if you specify this explicitly as query-dependent context using some kind of industry facet. Otherwise, instead of showing anyone searching for “project manager” everything, these user-dependent contexts can be extracted from the candidate profiles and past search behaviours to remove irrelevant results depending on who is performing the search. If I were the one looking at the results, job ads from companies that share the same characteristics as the hirers from my past searches can be ranked higher. Hence, I will see jobs from my preferred companies first.


All in all, there is more to relevance than just what the users can explicitly provide to the search engine. In other words, relevance is contextual and something that is regarded as relevant by one person may not be relevant to another, despite the same keywords used by both. Many of the contexts can be mined from user profiles and click and query logs, and used to better tailor the search results.

I’m a professional data + product leader trained in comp + info science. I code, write, take photos for fun.