Today's e-commerce platforms provide a huge amount of products attached with a rich set of properties. Modern online shops support users in the product search process by o ering facet retrieval, i.e., limiting search results by using several lters. However, many systems support retrieval of items ful lling Boolean conditions so that only exact matching leads to the inclusion in the result list. On the other side, users often have vague conditions in mind, e.g, "high resolution laptop\, and cannot describe their information need precisely. Users are forced to specify crisp conditions to get any results but this leads to the problem that only a few items are returned by the system. Often, there exist other products that closely satisfy the user's information need. In this work, we present the obvious search problems with vague conditions and potential research questions w.r.t product search. We argue that leveraging probabilistic modeling techniques and especially user-driven UI development can optimize the overall retrieval quality.
One important application in Information Retrieval (IR) is product search on ecommerce platforms. In recent years online shopping gained more popularity and the online retail worldwide increased rapidly1. This shows that online shopping plays a signi cant role in today's goods acquisition and overtakes traditional shopping services.
Users frequently use online-shops either for known item search or for search based on product properties description. While the former one delivers satisfactory results, the latter often yields insu cient answer sets. This problem occurs especially in case of vague search conditions which cannot be handled by existing systems. For example, someone is searching for a "powerful and lightweight laptop with a long battery runtime for less than $800\. For such query examples, many IR systems o er full-text search, which is based on lexical comparisons and ignore the query semantics. Phrases like \long battery runtime" would hardly lead to appropriate result items and the asking price would be interpreted as a search for the number 800 in the product description. Faceted lters are provided to assist the user in the search process by restricting search results according to speci c conditions, e.g., price or display size, price value range. But in this way the user is forced to concrete her information need and to enter crisp conditions for facet lters. In some cases this is di cult; for example when a user searches for \lightweight laptops", one has to determine the threshold for \lightweight". For some product properties there are no lters available, e.g., \powerful laptops". Another issue occurs when some criteria are in con ict with each other, e.g., "powerful\ vs. "cheap\. By default the query conditions are strictly interpreted and must all be ful lled like in Boolean retrieval systems. These systems reach their limits when the query becomes more complex and results in phenomena like the empty answer problem. Users have to re-formulate the query by relaxing certain conditions. But instead of this heavy-going re-formulation process, the system could help the user, if query conditions were understood as vague, and a discrimination between relevant and less-relevant criteria would be possible. In the interest of the user, it could be useful to retrieve laptops that slightly miss the query conditions. By incorporating user's query constraints, i.e., conditions that must be met, and query preferences, i.e., optional and less-relevant conditions, the IR system can distinguish among the di erent criteria and return more potentially relevant items.
The focus of our research lies in developing interactive IR methods that are capable of nding objects that also closely match the query conditions in the use case of product search. The concept of vagueness will be investigated with regard to text and fact conditions. We are working on the implementation of a search engine that addresses the challenges in the aforementioned problems. 2
Probabilistic IR models are based on the probabilistic ranking principle (PRP)
claiming that documents should be ranked according to their probabilistic
relevance w.r.t. query [
In the recent past, a quantum logic based approach (QQL) was introduced
by [
In the following we describe the challenges and research questions according to vague text and fact conditions in product search by using the laptop purchase as on-going example.
Vague Fact Conditions A vague query fact condition describes the extent of ful llment according to an attribute value of one laptop. Instead of binary weights, the associated score is probabilistic and ranges from 0 to 1. There are di erent types of vague fact conditions to distinguish: { Value Equality refers to a concrete value of an attribute, e.g., "display size = 14 inch\ { Threshold refers to a concrete value of an attribute which should not be below or above, e.g., "price < $500\ { Tendency refers to a preference direction, e.g., "lightweight\
Attributes can have di erent scales, e.g., nominal, ordinal or metric. Depending on the vague fact condition type and the attribute scale, di erent approaches have to be investigated for computing a respective vagueness score. For our use case vague fact conditions can be applied to all technical product properties, e.g., CPU, memory, hard drive, display size, battery runtime, color, operating system, etc., as well as to the product price. Another obstacle is to nd laptops that most closely meet all technical requirements, but for the lowest price. For each laptop an overall probabilistic relevance score w.r.t. the user query conditions can be computed. In this way, the result list includes also items with relative high probabilistic scores and not only items perfectly tting to the user query.
Vague Text Conditions One problem in full-text search arises with
mismatching vocabularies between query and document. The same concepts can be
described by di erent words and language styles, thus making the matching
procedure more complicated. When a user searches for \high performance laptops",
matching on the lexical level would fail in case of laptops containing synonyms
of the query phrase like \high quality notebooks" or \powerful computers".
Instead of matching both sides by word-by-word comparisons which is prevalent in
many search engines, we aim to develop a semantic approach capable in
measuring the similarity between query and document. Previous work [
Faceted retrieval conjunctively combines query conditions speci ed by the user. In order to receive any results, users often specify conditions carefully, otherwise the description of the optimal result would yield an empty result list. An alternative solution would be to introduce optional query conditions. It should be possible to weigh selected query criteria di erently. Especially for binary laptop attributes it would be important to allow for optional condition selection. Moreover, the relative importance among single conditions enables a better treatment of con icting criteria, e.g., price vs. performance. As search queries are extended by vague text and fact conditions, it is important to develop ranking functions which take vagueness as well as explicit user weighting into account.
The association between vagueness and weighting can be explained by an easy example. A user searches for a laptop with 16 00 display size and a SSD with 256 GB for a low price. If the user prioritizes the price over technical requirements, then the top ranked laptop in the result list could be a laptop with only 15 00 display size and a SSD in the desired size for $750. A laptop ful lling all technical criteria but being $100 more expensive would be ranked on the second position. Only due to the query expansion by the vague fact condition w.r.t. display size, the user receives the rst result.
Some phrases in the user query refer to property values of one laptop, e.g., "high performance notebook\ refers to the attribute CPU speed. Nowadays, many IR systems do not di erentiate text from fact conditions and treat the previous example query phrases as text conditions, i.e., the system retrieves only products that contain these terms in the product title or description. But a mapping from text to fact conditions is not performed.
One possible solution to this problem is based on the analysis of query logs. Frequent phrases can be detected and mapped to their respective vague fact condition, e.g., "cheap\ ! "low price\. Another approach is to measure the correlation between phrases in laptop descriptions or titles with attribute values, i.e., generate rules by regarding the association of a term-attribute pair. The nal performance of the transformation rules will be evaluated in real search sessions by leveraging user's implicit feedback.
Another interesting research topic is result set visualization. One challenging task is the determination of the ordering of result objects including the snippet and overview information. Furthermore, the user should be able to interpret and understand the results returned by the system. Hence, it is necessary to implement a "White-Box\ system that contains for each result item additional information about which query conditions are met and which not. The transparent result set visualization could help the user in her decision-making process. Additionally, it is useful to investigate on alternative result list representations where the user can easily read and compare con icting conditions, e.g., "low price vs. high performance\, and to select a reasonable trade-o . The interaction with the result list enables the collection of implicit feedback, e.g., relevance feedback by accepting and denying laptops. If a user accepts one laptop, the corresponding attribute values contribute to adjusting parameters of the vagueness and ranking methods.
As experimental data we crawled more than 20,000 laptops from the Amazon
platform. On the one hand, we can use this dataset to model di erent vagueness
functions by utilizing probabilistic logic like proposed in [
The development of the UI considers user interaction. We want to investigate how a user can a ect the parameters of the di erent concepts, e.g., vagueness and weighting functions of fact conditions or term selection for query expansion for treatment of vague text conditions. This user feedback is relevant to judge the quality of the internal functions.