This is a guest post by Nitin Karandikar, author of the Software Abstractions blog.
Recently I was looking at the log files for my blog, as I regularly do, and I was suddenly struck by the variety of search queries in Google from which users were being referred to my posts. I write often about the different varieties of search – including vertical search, parametric search, semantic search, and so on – so users with queries about search often land on my blog. But do they always find what they’re looking for?
All the major search engines currently rely on the proximity of keywords and search terms to match results. But that approach can be misleading, causing the search engine to systematically produce incorrect results under certain conditions.
To demonstrate, let us take a look at three general use cases.
[Note: The examples given below are all drawn from Google. To be fair, all the major search engines use similar algorithms, and all suffer from similar problems. For its part, Google handles billions of queries every day, usually very competently. As the reigning market leader, though, Google is the obvious target – it goes with the territory!]
1. Difficulty in Finding Long Tail Results
Take Britney Spears. Given the current popularity of articles, news, pictures, and videos of the superstar singer, the results for practically any query with the word “spears” in it will be loaded with matches about her – especially if the search involves television or entertainment in any way.
Let’s say you’re watching the movie Zulu and you start wondering what material the large spears that all the extras are waving about are made of. So, you go to Google and type in “movie spears material” – this is an obviously insufficient description, as the screen shot below shows.
What happens if you expand on the query further – say: “what are movie spears made out of?” – does it help?
The general issue here is that articles about very popular subjects accumulate high levels of PageRank and then totally overwhelm long tail results. This makes it very difficult for a user to find information about unusual topics that happen to lie near these subjects (at least based on keywords).
2. Keyword Ordering
Since the major search engines focus only on the proximity of keywords without context, a user search that’s similar to a popular concept gets swamped with those results, even if the order of keywords in the query has been reversed. For example, a tragic occurrence that’s common in modern life is that of a bicycle getting hit by a car. Much less common is the possibility of a car getting hit by a bicycle, although it does happen. How would you search for the latter? Try typing “car hit by bicycle” into Google; here’s a screen shot of what you get. [Note the third result, which is actually relevant to this search!]
3. Keyword Relationships
Since the major search engines focus only on the keywords in the search phrase, all sense of the relationship between the search terms is lost. For example, users commonly change the meaning of search terms by using negations and prepositions; it is also fairly common to look for the less common members of a set.
This takes us into the realm of natural language processing (NLP). Without NLP, the nuances of these query modifications are totally invisible to the search algorithms.
For example, a query such as “Famous science fiction writers other than Isaac Asimov” is doomed to failure. A screen shot of this search in Google is presented below. Most of the returned results are about Isaac Asimov, even when the user is explicitly trying to exclude him from the list of authors found.
All of the searches shown above look like gimmicks – queries designed intentionally to mislead Google’s search algorithms. And in a sense, they are; these specific queries can be easily fixed by tweaking the search engine. Nevertheless, they do point to a real need: the value of understanding the meaning behind both the query and the content indexed.
That’s where the concept of semantic search comes in. I attended a media event earlier this year at stealth search startup Powerset (see: Powerset is Not a Google-killer!), at which they showcased a live demo of their search engine, currently in closed alpha, that highlighted solutions to exactly this type of issue.
For example, type “What was said about Jesus” into a major search engine, and you usually get a whole list of results that consist of the teachings of Jesus; this means that the search engine entirely missed the concepts of passive voice and “about.” The Powerset results, on the other hand, were consistently on target (for the demo, anyway!).
In other words, when you look at just the keywords in the query, you don’t really understand what the user is looking for; by looking at them within context, by taking into account the qualifiers, the prepositions, the negatives, and other such nuances, you can create a semantic graph of the query. The same case can be made for semantic parsing of the content indexed. Put the two together, as Powerset does, and you can get a much better feel for relevance of results.
What about Google? I’m sure the smart folks in Google’s search-quality team are busily working on this problem as well. I look forward to the time when the major search engines handle long tail queries more accurately and make search a better experience for all of us.
Update: for an expanded version of this article with real-life user queries, see my blog.