Vector space retrieval model for e-commerce
Sep 04, 2020 • 11 min read
Sep 04, 2020 • 11 min read
Smart autocomplete is an essential feature for improving online catalog navigation and boosting product discovery for users. This is especially true for mobile search, where the smaller screen and keyboard, limit the use of more traditional faceted search selectors. Smart autocomplete, also known as autosuggest, does more than merely forcecast words or phrases the user is typing. Smart autocomplete goes a step further, and anticipates the user’s intentions in order to make helpful suggestions. These suggestions improve the user’s search experience, increasing both online conversion rates and average online cart value.
More than a simple autocomplete or typeahead tool, smart autocomplete can improve the user’s navigation and product discovery experience. Because of its subtlety, the user perceives query suggestions as impartial. This raises its value as a product discovery tool for both users and retailers. Overall, smart autocomplete improves both the customer's experience, as well as helping the retailers merchandisers and bottom line. Here are some of the specific things it can do:
If a user types in a generic subject, say “dresses”, smart autocomplete can use this opportunity to suggest dress brands, like “Calvin Klein dresses”. Smart autocomplete can also suggest dress types, like “prom dresses”, which give users a helpful suggestion, and also applies a merchandiser's business logic of promoting a brand or specific type of dress.
Let’s say a user is searching for jackhammers. This is an opportunity to inform them via the autocomplete list, that jackhammer rentals are also available. This result may appeal to the user who had not even considered this, or was even aware of this option.
Smart autocomplete can be used to explain or promote important product features to customers. Interestingly, these suggestions are made by other experienced users via their previous search history. For example, users who typed in “dress shirts” might be looking for “slim fit dress shirts”, a suggestion generated by previous users’ searches.
If a user begins a query by typing “dresses”, autosuggest, supported by the search engine, can suggest “dresses under $100”. In this case, the phrase appears because autosuggest is able to anticipate phrases like "<phrase> under $<price>".
Autosuggest can be used for implicitly promoting particular brands without annoying customers with ads and banners. For example, if a user types in “shoes”, and a retailer is promoting the UGG brand, this brand can be placed on the top of the auto suggest list rather than using a banner ad to attract attention.
The quality of the displayed listings covered above are highly dependant on building a quality corpus, the fuel for the autosuggest engine. We will cover how to build the corpus in the next section.
Smart autocomplete relies on a phrase corpus, a large and structured set of text phrases, to generate suggestions for user queries. When building a smart autocomplete system, you need to consider two major approaches for building the autocomplete phrase corpus. The first approach, or the customer phase approach, uses customer query data logs to populate the phrase corpus. The product data approach relies on phrases and keywords used in the product catalog. Both approaches have the following pros and cons:
|Customer search||● Natural language
● High recall (incl. synonyms and non-product search) *
● Ranks *
(* with high enough volume of search history )
● Semantic similarity
● Irrelevant results
● Zero results
● New product types and brands
|Product data||● Correct spelling *
● Relevant results *
● Semantic uniqueness *
● Full coverage of products data
(* with correct and consistent products data)
|● Natural language
● Recall issues
● Underused suggestions
It is interesting to note that the pros and cons of each approach mirror each other. The advantages of one approach often becomes a challenge for the other. However, these two approaches are not mutually exclusive. Therefore, the best approach may be a smart mix of both.
The product data approach can be a good option if the retailer’s catalog attributes are well-defined, and not too diverse. The logic here is that often users are attempting to enter a phrase similar to the one included in the product data corpus, which is based on the product catalog. The product data corpus may also be used simply because data from past customer searches is not available. This may be because policy issues forbid the capturing of consumer data.
The alternative approach uses data from previous customer searches to power helpful suggestions. Given the high volume of searches on heavily trafficked retail sites, this method allows you to leverage popular searches for autocomplete.
When starting to develop a corpus of autocomplete phrases, we recommend initially focusing on customer search phrases. In most cases, the advantages outweigh the challenges. The remainder of this article will focus on the customer data approach.
Building autocomplete phrases from customer searches starts with capturing clickstream data. To build a useful corpus, we will also need:
Once raw clickstream data is captured, filtering will be required to extract useful information for potential autocomplete display results. Any phrase displayed by autocomplete should be relevant to the user. If irrelevant information (false promises) appear in the autocomplete window too often, the user’s confidence in the results will diminish, as will engagement.
Partially matched phrases which yield a false promise or zero results, should be excluded from the autocomplete display.For example, the query “long dresses” provides the suggestion “long prom dresses”, which has no results but provides a partial match of “prom dresses”. The issue here is that the keyword long has been omitted from the partial match, substantially changing the intent of the original query.
It is possible that an exact match could still yield non-usable items in the autocomplete box.
If the search engine supports spell correction at autocomplete query-time, we recommend discarding misspelled candidates for suggested phrases. When a customer misspells a phrase in the search box, autocomplete identifies misspelling, fixes it on the fly and displays the correctly spelled suggestions instead.
Some words, however, are often misspelled, such as “Tommy Hilfiger” or “refrigerator”. These misspelled phrases can represent a significant amount of traffic. In these cases, discarding misspelled phrases will bias the ranking towards the terms with easier spelling. To avoid this, business metrics of misspelled phrases should be inherited by phrases with the correct spelling.
Additional difficulties can be introduced by misspellings in the retailer’s catalog data. For example, it is possible to find difficult words such as “fuchsia” or “fluorescent” spelled incorrectly, even in product catalogs. Unfortunately, this issue affects not only autocomplete, but also the core spell correction functionality. The best solution is to carefully QA the product catalog data.
Phrases which are semantically similar but contain different spellings should only be displayed once. For example, the word “dress” and “dresses” are semantically similar, but the difference is, one is plural. Only one form should appear in the autocomplete window.
Autocomplete must not show semantically identical phrases with close spellings, for example:
Semantic deduplication can be achieved by normalizing phrases to exclude all “semanticless” features. There can be exceptions for this approach (like “shirt dress” vs. “dress shirt”), which have closed spelling, but different semantics. This normalization can only be used for deduplication, not for a display list.
There can be semantic similarities between phrases with different spellings. This type of duplication, however, may not be noticeable, except to users with a deeper knowledge of the business domain.
For example, phrases with different spelling, but with the same semantics due to synonyms, could be treated as the same item. A full phrase could be treated the same as an abbreviated phrase. “Calvin Klein dress” is the same as “ck dress”. Product type synonyms can also be treated the same. An example of this rule is “living room sofas” verses “living room couches”.
Phrases with different search terms, but yielding the same results may also not be noticed except by a user familiar with the product. For example, “Lancome” and “Lancome Cosmetics” could be considered a duplicate.
Phrases from a semantic group may satisfy all other criteria of filtering, we recommend keeping them all in the corpus. This will ensure good recall even for those queries which match only one phrase in a group. However, the phrases should be marked by the same label within a semantic group. Therefore, if a query matches any phrase in that group, autocomplete will display only one instance in the display list.
It is possible that some poorly spelled phrases will pass all through all filters, therefore it may be required to add additional rule-based filters to block unwanted phrases from the user:
The process of filtering and grouping customer phrases for use with autosuggest is similar to other processes such as generating business reports on customer search habits. In fact, this reporting can become useful in many other business cases going beyond autocomplete. It can help retailers get a deeper understanding of how customers interact with the search engine.
Grouping similar semantic concepts is useful for extracting sales and conversion metrics. In the example below, separating “Michael Kors handbags”, “Michael Kors bags” and “mk bags” is not useful for reporting purposes. Autosuggest filtering can collapse these entries into one entry with a similar semantic concept. This revised list produces a much more useful report.
Partial matches may contain a relevant key match mixed with unrecognized terms. This type of report capture and emphasize this type of information. In the example below, “Pandora” is a brand name. Grouping all phrases containing "Pandora" gives insight to the number of times a user search for this brand, regardless of additional information they provided in the search bar.
Common misspelled terms should analysed, as it may contain useful information for potential sales conversions. For example, a user might type “Ridgid”, a brand of tool this retailer does not carry, but which have close spellings to words existing in the index, like “rigid”. This would create a false positive, and should be blocked by the spell corrector module configuration if it occurs too often.
Common misspellings should be captured along with properly spelled phrases for reporting purposes.
As with autosuggest, exact match phrases producing false promises may produce a dissatisfied customer. This information can be useful to companies looking to raise customer satisfaction.
Building a high quality text corpus is important, but only part of the autosuggest equation. Query-time matching and ranking must also be considered to produce the desired results. We will cover query-time operations next.
It can be tempting to use typical normalization techniques from traditional product search with autocorrect search, but this may have undesirable results. An autocomplete input query attempts to match as the user is typing. The incomplete phrase is often small, and we need to consider every aspect of it. This can be lost if significant normalization is applied. Therefore, we will apply what we call “consertative normalization”.
There are instances when preserving stop words, words a search engine will typically ignore, can be useful. Stop words can be the part of the brand name, such as “The North Face”. In this case, when the query is “the no” , “the” should be preserved since it can help to identify customer interest in The North Face products sooner.
Stop words also can add more meaning to the sentence phrase structure. This additional information may make the typeahead prediction more accurate.
Stemming, the process of removing a word stem suffix to reduce it to its root form, may alter the intent of the customer query. For example, a word like “batteries” is typically used as a noun in its plural form. If it is stemmed to its singular form, “battery”, this could introduce unwanted matches to its adjectives form, such as “battery charger”.
Query-time spelling correction is considered a good practice, because the user can immediately see feedback of the correct spelling. Spelling correction and matching should be attempted, as the user is typing, rather than waiting until the user completes the potentially misspelled word, as shown in the example below.
Multi-match is used in product searches to allow matching of different tokens of a phrase on the same product attribute or value. For example, it can help query tokens with different kinds of spelling to match the same product value through synonyms, such as "black oxford shoes".
We don’t use synonyms for autocomplete during query, and we do not need multi-match at query time. Every token of the query has to match its own unique token in an autocomplete suggest phrase. This helps identify cases when a repeated word changes semantics. An example of this is Michael Michael Kors, which is a sub-brand of Michael Kors.
The absence of multi-match is also helpful when the last incomplete query token (for example, in the figure below, “pro”) and a complete query token (for example, “protein”) are accidentally able to match the same suggest phrase token.
After building the phrase corpus, semantic deduplication should be applied at query-time. If semantically similar phrases have been marked with a shared label during the corpus building process as suggested, these labels can be used to deduplicate during query time. Only one phrase with the same label should be displayed.
If smart autocomplete is only used as a typeahead feature, it can be tempting to allow the match only from the beginning of a suggestion phrase, like matching “shoe rack” for “shoe”, but not “women shoes”, but this simplistic approach should not be used. Users increasingly use the search bar for interactive suggestions, which can be accommodated by smart autocomplete:
As a starting point, a “bag of words” approach is recommended. Both the query and suggestion phrase tokens are handled irrespective of their order.
This approach is good enough for the majority of cases, but can be improved using advanced ranking techniques discussed in the next section.
The most obvious approach is to rank suggested phrases by the number of search events, which works fairly well. There are, however, a few issues which must be addressed to improve it’s base behaviour:
The “bag of words” matching model works reasonably well, but it has issues in some particular cases
Middle tokens of a of multi-token named entity might be erroneously ranked too high. For example, “deck” is more likely to indicate an interest in “deck furniture”, even if “black & decker” is more popular.
As another example, “for” more likely indicates interest in “formal dresses”, even if “dress for women” is more popular.
It is possible to match such multi-word phrases where other semantics exists for a different order of the same words. For example, “dress shi” is more likely to indicate an interest in “dress shirt”, even if “shirt dress” is more popular.
Resolution of such issues may require advanced techniques such as the Learn-to-rank approach or entity tagger, used to tag entity phrases. These entity phrases can then have rank prioritized based on word placement within the phrase. These concepts are beyond the scope of this blog post, and may be discussed in a future blog post.
There are other ranking techniques that can make smart autocomplete ranking more accurate.
Geographic segmentation, or considering the users geographic location can improve ranking results. A user in Florida typing “snow” is more likely to be interested in “snow cone” than “snow blower”.
Taking into account the user’s location on the website, can affect phrase ranking. For example, a user typing in “Calvin Klein” while in the menswear section. is more likely to be interested in Calvin Klein shirts, rather than Calvin Klein skincare.
Implement highlighting of the displayed suggested phrases indicates which part of the phrase was actually input by the user. The display should highlight all of the matches of the “bag of words” model.
Allow your customers to interact with the suggestions list using a keyboard:
While a user navigates through the suggestion list using a keyboard, instant search can show the top few products for a currently selected suggestion. This can be used to make a shortcut from product discovery directly to checkout. It also gives the opportunity to explain meanings of suggested phrases. For example, if a user is searching “refrigerators”, autosuggest can list “french door refrigerators” - which may also include a drawing of a french door refrigerator, so the user understands its meaning.
Smart autocomplete, or autosuggest, can be a powerful discovery tool when implemented correctly. This blog post introduces smart autocomplete, and how it can improve your search experience. It also explains how to build a phrase corpus from a customer search log approach, and then demonstrates smart autocomplete best practices in displaying and ranking suggestions for a user query.
Overall, retailers need a smart autocomplete solution to anticipate a customer’s search intent, and provide useful suggestions. These suggestions will help guide the customer through the product discovery experience, and remove barriers to finding new products online. Investing in these features will consistently improve online conversion rates, and the size of online shopping carts, especially on mobile devices. Given the high impact from this feature, retailers with a large online catalog are essentially leaving money on the table without a powerful smart autocomplete solution.