Another disappointment with the 2017 Google MQ for Insight Engines is the fixation of evaluating search applications by the quality of their auto-suggest functionality. As a caution against Sinequa Gartner comment that “reference customers did not indicate that they chose Sinequa for its autosuggest tool functionality.”  I’m having great difficulty in understanding the import of that comment. The fundamental problem here is that ‘autosuggest’ is a concatenation of a number of query modification approaches which do different things in different ways depending on the search software and how it has been configured. Let me deal quickly with auto-correct, in which a thesaurus comes into play as the query is typed and from the index either suggests an alternate spelling, or (for example an acronym) offers more than one possible search query. As with all these approaches one of the decisions that has to be made is how many characters the user should type before being offered options. This is not a simple problem as it may vary by subject and by language.

To move on to query completion, this is a very complex topic. Indeed A Survey of Query Auto Completion in Information Retrieval is the title of a book published in 2016 running to almost 100 pages. The information needed to offer an expanded query can come from a number of different sources, including the index, search logs, prior searches and even the profile of the user. Enterprise search is getting very close to Google in this respect. Another aspect is whether the query auto completion (usually referred to as QAC) is repository dependent. In other words if a user has decided to search within the intranet will the auto completion only offer terms relevant to the intranet? These are all decisions that need to be made and yet which cannot easily be specified because of the range of options. Another factor to be taken into account is the security trimming. If someone types in [Bristol] and gets offered [Bristol redundancies] the news might spread quite quickly.

User interface issues also need to be taken into account. Studies have shown that users who touch type and are looking at the query box will make different decisions to those who are looking at the keyboard, especially if undue weight is given to the second term. The user may well look at the resultant suggestion and have no idea how they arrived at it. Users tend to be quite concerned if they type in a few letters and get offered a four word query phrase. “How did I get there from here?”. Some search applications (and Google is a good example) may offer more than one expansion offer. This looks very effective until the user realises that they can only click one. Having clicked it, been disappointed with the results and tried the original query it may be that the options presented are different because of the discarded query sequence. A balance has to be struck between being responsive to the user and yet providing a degree of stability in what expansion is being offered. Users new to the organisation may have a different level of requirement to users who have been in the organisation for many years. A further frustration can be that the query options are not presented with the number of hits in the way that facets (usually!) are. So the user might select what seems to be a very appropriate expansion only to find that there are a small number of results. That could be fine if the search is for a known item but not for exploratory search. An additional complication is when a federated search application is in use, when query expansion options can get very interesting indeed.

As with every element of search it makes no difference what ‘good practice’ is but whether users feel it is useful to them. It requires some well-managed user testing against personas and use cases to work out exactly what works best. Moreover queries change with time, and at certain times of the year “job” expanded to “job openings” is less useful than “job” expanded to “job performance review”  This argues for query expansion functionality that is easy to customise rather than a very powerful set of routines that match none of the use cases.

To say that this has been a concise summary of QAC is a complete understatement. Run a search in Google Scholar and you will find thousands of research papers on this topic along with a large number of Google patents. The logic behind query completion is a very sound one given the number of users that only use 1.5 query terms. Selecting a search application on the basis of the Gartner assessment of its autosuggest capabilities is probably less sound.

Martin White