Search as a Conversation

Most search applications assume a query-response paradigm: the searcher submits a search query, and the search engine responds with results. The query-response paradigm works well for simple search needs that the search engine understands.

The query-response paradigm breaks down, however, when searchers have more complex needs, or when the search engine struggles with query understanding. In those cases, it’s better to model search as a conversation.

Communication vs. Relevance

When we evaluate the performance of search engines, we usually think in terms of relevance — measuring it as a combination of precision and recall. In the query-response paradigm, we optimize for relevance: we have only one chance to satisfy the searcher’s needs, and we want to make the most of it.

When we model search as a conversation, however, we need to consider a bigger picture. A search session is a series of exchanges between the searcher and the search engine. An effective session is more than a sequence of isolated query-response pairs: it’s a learning process for the searcher. The effectiveness of that learning process depends on the quality of communication between the searcher and the search engine.

This bigger picture means good news and bad news for search application developers.

The good news: a conversational approach takes some of the pressure off of the search engine. Query understanding is never perfect, and a conversational model recognizes this imperfection. A conversational model encourages searchers to clarify or refine their queries when the search engine fails to understand them. Moreover, it offers searchers the opportunity to evolve their own intent based on the results the search engine returns to them.

The bad news: an effective conversation requires more from the search engine. It’s not enough for the search engine to make a best effort to return relevant results. The search engine also has to provide transparency, control, and guidance to searchers:

• Transparency, to know why they’re seeing the results the search engine returns to them — especially if those aren’t the results they want.
• Control, to take over manually when the search engine fails to automatically understand their queries.
• Guidance, to navigate the otherwise overwhelming space of ways in which they could reformulate their queries.

Transparency

The perfect search engine reads a searcher’s mind. But no search engine is perfect, so it’s important for a searcher to be able to recover from the search engine’s mistakes. In order to do so, the searcher needs to know how the search engine understood — or misunderstood — the query. In other words, the search engine has to provide transparency.

Transparency is difficult to define. A search engine can provide a literal and exhaustive description of how it processed a search query, but the searcher is unlikely to read all of it, let alone understand it. What the searcher wants and needs from the search engine is a clear, concise explanation of how it processed the query.

At the very least, a search engine should expose any spelling correction, query expansion, query relaxation, query segmentation, or query scoping that might not be obvious to the searcher. The searcher won’t care about the details when things go right; but when things go wrong, it’s important for the searcher to be able to figure out why.

Control

Transparency is necessary to let the searcher know what went wrong, but it isn’t sufficient. The searcher also needs control, in order to fix the problem.

Control means that the searcher can do — or undo — the same things the search engine does automatically. Searchers should be able to use quotation marks to specify exact words (overriding stemming or any other query rewriting) or phrases, as well as Boolean operators like AND, OR, and NOT to specify their query with arbitrary precision.

Providing control is not an excuse for the search engine not to make its own best effort. When the searcher’s intent is obvious, the search engine should be able to automatically derive that intent from the query, without requiring the searcher’s assistance.

Guidance

Transparency and control are necessary for a conversational model, but they’re not sufficient. The search engine also needs to offer guidance so that searchers can navigate the otherwise overwhelming space of ways they can formulate and reformulate their queries.

There are many ways that search engines can provide guidance to searchers. Autocomplete, when well-implemented, guides searchers to queries that the search engine understands. Spelling correction automatically or interactively (through “did you mean”) guides searchers to the queries that the searchers intened. Faceted search provides guidance to searchers by offering them ways to narrow their queries. Search engines can also suggest related searches: variants of the searcher’s query that the search engine hopefully understands.

No amount of guidance can address every possible communication breakdown between the searcher and the search engine. But something is far better than nothing, and guidance is often the difference between a successful course correction and a frustrating failure.

Summary

Modeling search as a conversation is a powerful way to address complex search needs. The conversational model recognizes that search engines sometimes struggle with query understanding. It allows searchers to clarify or refine their queries, as well as to evolve their own intent based on results the search engine returns to them.

But a conversational model requires more from the search engine. The search engine has to provide transparency, control, and guidance to searchers. Indeed, building a conversational search engine is harder than just supporting the query-response paradigm. But it’s worth the effort to address the wide range of searchers’ needs.

Previous: Personalization

Next: Clarification Dialogues