Automated search recipe generation
Techniques are disclosed for dynamically generating a data set representative of search results in response to a query and using the data set to accurately rank search results in response to a domain specific search query. Upon receiving the search query, features of the query and features of each search result are extracted. A relevance ranking may be assigned to each search result based on a comparison of the features of the query and each search result. The relevance ranking of each search result may be adjusted based on metrics related to user interactions. A data set may be created which includes the query, search results, extracted features, and metrics. The data set may be used to train a machine learning model to accurately determine a ranking of search results in response to a subsequent search query.
1. A computer-implemented method for dynamically generating a data set representative of search results responsive to a search query, the method comprising:
identifying features of the search query and features of a plurality of search results;
comparing the features of the search query and the features of the plurality of search results;
determining a relevance score for each search result in the plurality of search results based, at least in part, on the comparing and one or more metrics related to prior interactions with the plurality of search results by a plurality of users, wherein:
the one or more metrics related to the prior interactions comprise, for each respective search result of the plurality of search results, an assisted support deflection rate based on a respective number of assisted support sessions invoked by a respective user of the plurality of users after the respective user selected the respective search result,
the relevance score of a given search result is higher if the assisted support deflection rate of the given search result is higher, and
the relevance score of the given search result is lower if the assisted support deflection rate of the given search result is lower;
generating an ordered list of the plurality of search results based on the relevance score for each search result; and
returning the ordered list in response to the search query.
2. The computer-implemented method of claim 1 , wherein determining the relevance score for each search result comprises:
determining an initial relevance score for each search result in the plurality of search results based on the comparison;
assigning a weight to each of the features of the search query, the features of the search results, and the one or more metrics to generate weighted features; and
adjusting the initial relevance score for each search result based in the ordered list based on the weighted features and metrics.
3. The computer-implemented method of claim 1 , further comprising:
generating a data set including the plurality of search results, the relevance scores for each search result, and the metrics related to the prior interactions with the plurality of search results; and
training a machine learning model with the data set.
4. The computer-implemented method of claim 1 , wherein the metrics related to the prior interactions further comprise at least one of:
a click rate; or
an average voted.
5. The computer-implemented method of claim 4 , wherein the click rate indicates a number of times each search result is selected from the ordered list of the plurality of search results and is normalized by at least one of: dividing the number of times each result is selected by a static value and assigning a weight to each selection of a given search result, wherein the weight is inversely related to the relevance score of the given search result.
6. The computer-implemented method of claim 4 , wherein the average vote indicates a level of relevance of each search result provided by a user.
7. The method of claim 1 , wherein the metrics related to the prior interactions further comprise, for each respective search result of the plurality of search results, implicit feedback based on a length of time that a support document indicated by the respective search result was open after a respective user of the plurality of users selected the respective search result.
8. A non-transitory computer-readable storage medium storing instructions, which, when executed on a processor, perform an operation to dynamically generate a data set representative of search results responsive to a search query, the operation comprising:
identifying features of the search query and features of a plurality of search results;
comparing the features of the search query and the features of the plurality of search results;
determining a relevance score for each search result in the plurality of search results based, at least in part, on the comparing and one or more metrics related to prior interactions with the plurality of search results by a plurality of users, wherein;
the one or more metrics related to the prior interactions comprise, for each respective search result of the plurality of search results, an assisted support deflection rate based on a respective number of assisted support sessions invoked by a respective user of the plurality of users after the respective user selected the respective search result,
the relevance score of a given search result is higher if the assisted support deflection rate of the given search result is higher, and
the relevance score of the given search result is lower if the assisted support deflection rate of the given search result is lower;
generating an ordered list of the plurality of search results based on the relevance score for each search result; and
returning the ordered list in response to the search query.
9. The non-transitory computer-readable storage medium of claim 8 , wherein determining the relevance score for each search result comprises:
determining an initial relevance score for each search result in the plurality of search results based on the comparison;
assigning a weight to each of the features of the search query, the features of the search results, and the one or more metrics to generate weighted features; and
adjusting the initial relevance score for each search result based in the ordered list based on the weighted features and metrics.
10. The non-transitory computer-readable storage medium of claim 8 , the operation further comprising:
generating a data set including the plurality of search results, the relevance score for each search result, and the metrics related to the prior interactions with the plurality of search results; and
training a machine learning model with the data set.
11. The non-transitory computer-readable storage medium of claim 8 , wherein the metrics related to the prior interactions further comprise at least one of:
a click rate; or
an average vote.
12. The non-transitory computer-readable storage medium of claim 11 , wherein the click rate indicates a number of times each search result is selected from the ordered list of the plurality of search results and is normalized by at least one of: dividing the number of times each result is selected by a static value and assigning a weight to each selection of a given search result, wherein the weight is inversely related to the relevance score of the given search result.
13. The non-transitory computer-readable storage medium of claim 11 , wherein the average vote indicates a level of relevance of each search result provided by the respective user.
14. The non-transitory computer-readable storage medium of claim 8 , wherein the metrics related to the prior interactions further comprise, for each respective search result of the plurality of search results, implicit feedback based on a length of time that a support document indicated by the respective search result was open after a respective user of the plurality of users selected the respective search result.
15. A system, comprising:
a processor; and
a memory hosting an application, which, when executed on the processor, performs an operation to dynamically generate a data set representative of search results responsive to a search query, the operation comprising:
identifying features of the search query and features of a plurality of search results;
comparing the features of the search query and the features of the plurality of search results;
determining a relevance score for each search result in the plurality of search results based, at least in part, on the comparing and one or more metrics related to prior interactions with the plurality of search results by a plurality of users; wherein,
the one or more metrics related to the prior interactions comprise, for each respective search result of the plurality of search results, an assisted support deflection rate based on a respective number of assisted support sessions invoked by a respective user of the plurality of users after the respective user selected the respective search result,
the relevance score of a given search result is higher if the assisted support deflection rate of the given search result is higher, and
the relevance score of the given search result is lower if the assisted support deflection rate of the given search result is lower;
generating an ordered list of the plurality of search results based on the relevance score for each search result; and
returning the ordered list in response to the search query.
16. The system of claim 15 , wherein determining the relevance score for each search result comprises:
determining an initial relevance score for each search result in the plurality of search results based on the comparison;
assigning a weight to each of the features of the search query, the features of the search results, and the one or more metrics to generate weighted features; and
adjusting the initial relevance score for each search result based in the ordered list based on the weighted features and metrics.
17. The system of claim 15 , the operation further comprising:
generating a data set including the plurality of search results, the relevance scores for each search result, and the metrics related to the prior interactions with the plurality of search results; and
training a machine learning model with the data set.
18. The system of claim 15 , wherein the metrics related to the prior interactions further comprise at least one of:
a click rate; or
an average vote.
19. The system of claim 18 , wherein the click rate indicates a number of times each search result is selected from the ordered list of the plurality of search results and is normalized by at least one of: dividing the number of times each result is selected by a static value and assigning a weight to each selection of a given search result, wherein the weight is inversely related to the relevance score of the given search result.
20. The system of claim 15 , wherein the metrics related to the prior interactions further comprise, for each respective search result of the plurality of search results, implicit feedback based on a length of time that a support document indicated by the respective search result was open after a respective user of the plurality of users selected the respective search result.