Amazon Technologies, Inc.

Jianhui Wu

Srikanth Thirumalai

Debprakash Patnaik

Krishnamurthy Koduvayur Viswanathan

Rongzhou Shen

Rui Wang

Gabriel Parent

Wenke Zhang

Jiwei Cao

Nicholas R Boyd

Taibai Xu

Techniques for extracting unstructured quantitative data may be provided. For example, a process may attempt to extract unstructured quantitative data to form structured data. This quantitative data may be used for searchable ordering of items in an electronic marketplace. For example, a process may attempt to find common attributes amongst several item descriptions. The common attributes may be recognized in the unstructured data, stored as structured data, and incorporated with a network page to allow the user to search for a particular item with a particular attribute. The desired attribute can help narrow a set of results from a search query.

Extracting data from a catalog

Systems and methods involve filling missing attribute values from unstructured text. A computing device may provide a plurality of items, such as an item catalog for an electronic marketplace. When an item is found to have a missing attribute value, a plurality of existing values for that attribute is compiled by mining other items. Text associated with the item is parsed to determine possible values for the attribute. From those possible values, the most likely value is identified and the missing attribute value is populated with that value.

Attribute fill using text extraction

Systems and methods are disclosed for generating recommendation rules based on the attributes of items that are purchased together at a threshold rate. The attributes of the items may be extracted from item-detail content associated with the items. Using a count of the frequency with which pairs of items include pairs of attributes, a recommendation rule can be created that recommends items with particular attributes to users who access other items with particular attributes. Further, using the recommendation rules, items may be selected for recommendation to users who access an item that lacks historical access data from which to generate recommendations solving the 'cold-start' problem. Moreover, negative rules may be generated based on historical access data and attributes of items purchased and/or not purchased together at a threshold rate to prevent the recommendation of particular items to users who access items associated with the negative rules.

Recommendation generation for infrequently accessed items

Embodiments may include, for each item in a subset of items from a larger group of items, evaluating item description information about that item to identify a respective set of candidate phrases to be evaluated. Embodiments may also include, for each phrase in the sets of candidate phrases, generating multiple component scores based on one or more of the frequency with which that phrase occurs in the item description information for the subset of items and/or the frequency with which that phrase occurs in a corpus of item description information for the overall group of items. Embodiments may also include, for each phrase in the sets of candidate phrases, generating a respective phrase score based on the component scores generated for that phrase. Embodiments may include, based on phrase scores, selecting a subset of phrases from the sets of candidate phrases as being feature phrases for the subset of items.

System and method for identifying feature phrases in item description information

Embodiments may generate a population of candidate rules including multiple rule conditions for detecting duplicates, each duplicate representing different sets of item description information that describe a common item. For each candidate rule of the population, embodiments may apply that rule to a reference data set including known duplicates and non-duplicates. Embodiments may assign each candidate rule a fitness score generated with a fitness function based on the performance of that candidate rule. Embodiments may, based on the fitness scores, select a subset of the population of candidate rules as parents for the new generation of candidate rules. Embodiments may perform crossover and/or mutation operations on the parent candidate rules to generate the new generation of candidate rules. Embodiments may select from the new generation of candidate rules (or from subsequent generations of candidate rules), rules for inclusion within a rule set for detecting duplicates within item description information.

Growing community of inventors

Bellevue, WA, United States of America

Jianhui Wu