We recently listened to Google’s 6th annual I/O keynote address; however it was a year earlier at their 5th annual keynote address that Google announced an advancement in technology and programming that signaled “the end of search as we know it”, the title of these year’s search section of their keynote address. That advancement was the launch of Google’s Knowledge Graph.
Before knowledge graph, all search queries were interpreted by Google’s search engines as a literal string of words, however, human beings understand that strings of words, phrases and sentences can have meanings that are much richer than the literal understanding of those words alone.
For example, the phrase [Khao San Road] can be interpreted by humans as either, a real world destination in Bangkok, Thailand, or as the name of a Thai restaurant for example. Whereas, pre knowledge graph, Google’s search engines wouldn’t “interpret” [Khao San Road] at all and would instead simply match that keyword phrase, to content on the World Wide Web, with no idea which definition of [Khao San Road] you intended when making the query, unless you refined your search by adding additional words to the search string.
Knowledge graph is the database the serves to make the “interpretation” of search queries possible, by working to allow Google to understand real world entities and moreover, the relationships between these entities, rather than only understanding search strings. It does this by tapping into the collective intelligence of the web, rather than just matching keywords to content arbitrarily, moreover, by being tuned according to what people search for and what they find on the web, Google will begin to ‘understand’ the world more like humans do.
At this stage in its development, back in 2012, knowledge graph had already provided the potential to enhance search in 3 main ways:
1.) To allow the search engines to understand features of language like ambiguity – do you mean Khao San Road the place or Khao San Road the restaurant? By organising search results in to separate streams relevant to each meaning of the search query, it allows you to select the results relevant to the definition of the query you intended. This means search can find the right thing more efficiently that it did prior to knowledge graph.
2.) By allowing Google to better understand a search query, knowledge graph enables Google to display a better, more relevant summary of the information related to that query. So again, if we were searching for Khao San Road the summary may include its location: Bangkok, Thailand, and establishments that can be found on it or near it such as places to eat, stay and visit. Knowledge graph finds the content for this summary by aggregating what users have been searching for in Google, alongside which other real world entities Google understands have a connection to the query in question.
3.) Finally, by enabling Google to make these summaries of facts relevant to your search query, and understand the inter-relationships between real world entities, it opens up the opportunity for making unexpected discoveries. Essentially this means the search experience provided by knowledge graph will allow you to access both deeper and broader results, opening up new potential lines of inquiry.
These three factors mean Google can now answer your second search query before you’ve even asked it because the results they display are informed, thanks to knowledge graph, by what others have searched for.
So that was 2012, now in 2013 Google has expanded, grown and developed this knowledge graph to such an extent that at this years I/O Keynote address they announced their first successes in the field of Conversational Search with “OK Google”.
Due to Knowledge Graph having already being able to address the problem of ambiguity in human language by allowing Google to interpret the different meanings of a single search query; a major doorway into conversational search had been opened.
The ambiguity of natural language is a fundamental trip wire for attempts at natural language processing, by being able to break through the layers of possible ambiguities on certain search terms; Google has established a solid platform from which to begin tackling the other problems natural languages and natural language processing poses for conversational search.
Moreover, another aspect of natural language Google has seemingly overcome is that of reference. Often, most of what is said during a conversation is reliant of referencing, most commonly anaphoric referencing. An anaphoric reference is a word that refers back to an idea, entity or object previously mentioned, most commonly it takes the form of ‘it’.
Google exemplifies this in their keynote address (if you skip to 1.50 minutes into the keynote you’ll be at the beginning of the search section of the presentation) with the demonstration by Joanna Wright of how ‘OK Google’ responds when Google is asked the question:
“How far is it from here?”
Google is able to interpret ‘it’ as an anaphoric reference to Santa Cruz which was the topic of the previous few search queries and to also interpret ‘here’ as the users current destination. It is this latter accomplishment that is more noteworthy as it offers an insight into how Google is aggregating data for Knowledge graph and where Google is searching in order to find answers and interpretations for the search queries it handles.
This is because at no point in the demonstration is Joanna’s destination mentioned, meaning Google is interpreting ‘here’ externally of the search queries and filtering the results from the data stored on the device the query is made from. This is an aspect of Google Now, enabling you to answer questions about your own world.
In combination, these too abilities, the ability to handle ambiguity and the ability to handle referencing, for the most part allow Google to interpret the displacement feature of natural languages. Displacement refers to natural languages’ ability to speak about not only what is happening at the time of conversation, but also to discuss, past, future, real, unreal and other situations.
Voice Recognition & Natural Language Processing
At this stage it seems Google has been able to effectively tackle certain aspects of natural language processing that have proved tricky before, however there is still a wealth of language related problems they will need to solve before conversational search can become the norm; a fact which Amit Singhal himself acknowledges in the 2013 keynote.
For example, while the voice recognition technology Google has developed so far appeared to work well in the keynote address, they still have to contend with making the technology sensitive to accents, dialects and languages.
This in itself is a vast task, but when you consider the diversity of accent and dialect within a single language the possibility of making voice recognition effective across the whole seems to grow smaller. Take English for example, there is American English, British English, Singlish – a version of English spoken in Singapore and then within those variations of the language there are dialectal variations such as Liverpudlian, Yorkshire and Texan.
Equally voice recognition technology may also run in to difficulties for conversational search when it comes to delivering the service to tone languages like Chinese, where meaning is attributed to a word using intonation alone. It is similar to the example in French of ‘ca va’, with a falling intonation meaning ‘yes/I’m okay’ whereas ‘ca va’ with a rising intonation informally poses the question ‘how’s it going?’
Beyond this there is the problem that natural languages are infinite in their ability to create new utterances, words and meanings and given the fast paced nature of today’s world, it will be interesting to see how and indeed if Google can a.) Programme software to replicate the creativity of natural language using mathematical formulae founded in syntax or b.) Whether they can effectively keep pace with the rate of language variation and change. Already this year the noun Vine came to have a new meaning thanks to the work over at Twitter, and twitter itself incidentally is another word with another new meaning.
Moreover, prevarication is another feature of human language conversational search may need to anticipate though it is less integral to the ultimate functionality of the software. Prevarication is the ability of natural languages to construct utterances or sentences while knowing they are false, with the intention of purposeful misleading the receiver of the information.
I’m sure given greater thought there are other problems and pitfalls Google may encounter in their quest to bring conversational search to the fore and make it the norm be they mathematical, linguistic or technological. And indeed if you can think of any I’d be interested to hear them. Nonetheless I conclude that given the progress Google has made so far, conversational search is a definite possibility rather than an absolute uncertainty and I look forward to it entering the mainstream.
Introducing Knowledge Graph: Posted by Amit Singhal, SVP, Engineering
Yule, G. (2006). The Study of Language. Cambridge: Cambridge University Press. In Encyclopaedia of Language and Linguistics – 2nd Edition (Keith Brown, Editior). Oxforsd: Elsevier, 2005. [Accessed Online]
Jessica Lee, May 16th, Ok Google.