Search analysis and retrieval via machine learning embeddings
Various embodiments of the present disclosure provide methods, apparatus, systems, computing devices, computing entities, and/or the like for retrieving relevant items for user queries by generating, using a search engine machine learning model, a prediction-based action for the query input wherein query input embeddings of the query input are generated. For each query input embedding, a k-Nearest-Neighbor (KNN) search is performed with respect to search engine repository item embeddings to generate initial search results, and for each initial set result, performing N hops within a semantic graph starting from nodes associated with the initial search result to generate related search results. The search engine machine learning model is trained by generating a search engine repository item embeddings according to embedding techniques for respective content categories and generating the semantic graph based at least in part on a measure of similarity for pairs of search engine repository item embeddings.
1. A computer-implemented method comprising:
receiving, by one or more processors, a query input; and
generating, by the one or more processors and using a search engine machine learning model, a prediction-based action for the query input by:
(i) generating a plurality of query input embeddings of the query input that respectively correspond to a plurality of content categories. wherein the plurality of content categories comprises a geospatial content category and the plurality of query input embeddings comprises a query geohash embedding corresponding to the query input,
(ii) identifying a plurality of initial search result sets, using a k-Nearest-Neighbor (KNN) search, respectively corresponding to the plurality of content categories based at least in part on a comparison between the plurality of query input embeddings and a plurality of search engine repository item embeddings,
(iii) generating, via N hops within a semantic graph starting from a plurality of nodes respectively associated with the plurality of initial search result sets, a plurality of related search results, and
(iv) generating the prediction-based action for the query input based at least in part on the plurality of related search results.
2. The computer-implemented method of claim 1 , wherein a search engine repository item of a plurality of search engine repository items respectively corresponding to the plurality of search engine repository item embeddings comprises metadata corresponding to a content category label, and a search engine repository item embedding is generated for the search engine repository item using an embedding technique corresponding to the content category label.
3. The computer-implemented method of claim 2 , wherein the embedding technique comprises generating a syntactic embedding by:
generating tokenized content by tokenizing content of the search engine repository item with subword n-grams;
training a term-frequency inverse document frequency (TF-IDF) model on the tokenized content; and
removing one or more bottom tokens with low TF-IDF values from the tokenized content.
4. The computer-implemented method of claim 2 , wherein the embedding technique comprises generating a semantic embedding by:
retrieving a domain corpus; and
tuning a universal sentence encoding model on the domain corpus.
5. The computer-implemented method of claim 2 , wherein the embedding technique comprises generating a geohash embedding by:
converting geospatial information into one or more geocodes according to a geohashing algorithm;
identifying, from a data source, procedure geocode pairs based at least in part on procedures associated with respective ones of the one or more geocodes; and
training a skip-gram model based at least in part on the procedure geocode pairs.
6. The computer-implemented method of claim 1 , wherein the plurality of content categories comprises one or more of: a syntactic content category, a semantic content category, and the geospatial content category.
7. The computer-implemented method of claim 1 , further comprising:
generating the semantic graph based at least in part on a measure of similarity for pairs of the plurality of search engine repository item embeddings respectively corresponding to a plurality of search engine repository items, wherein generating the semantic graph comprises:
retrieving a plurality of semantic embeddings respectively corresponding to the plurality of search engine repository items;
computing the measure of similarity between pairs of the plurality of semantic embeddings;
generating edges between the pairs of the plurality of search engine repository item embeddings in the semantic graph based at least in part on the measure of similarity; and
removing given ones of the edges according to a threshold value.
8. The computer-implemented method of claim 7 , wherein the measure of similarity comprises pairwise cosine similarity.
9. The computer-implemented method of claim 1 , wherein the plurality of query input embeddings further comprises a user embedding that is based at least in part on historical information and activities data associated with a user that provides the query input.
10. The computer-implemented method of claim 1 , further comprising: ranking one or more search results of the plurality of initial search result sets and the plurality of related search results based on a user relevance score.
11. The computer-implemented method of claim 10 , wherein the user relevance score is computed based at least in part on semantic relevance, syntactic relevance, geospatial relevance, and user relevance.
12. A system comprising one or more processors and at least one memory storing processor-executable instructions that, when executed by the one or more processors, configure the system to perform operations comprising:
receiving a query input; and
generating, using a search engine machine learning model, a prediction-based action for the query input by:
(i) generating a plurality of query input embeddings of the query input that respectively correspond to a plurality of content categories, wherein the plurality of content categories comprises a geospatial content category and the plurality of query input embeddings comprises a query geohash embedding corresponding to the query input,
(ii) identifying a plurality of initial search result sets, using a k-Nearest-Neighbor (KNN) search, respectively corresponding to the plurality of content categories based at least in part on a comparison between the plurality of query input embeddings and a plurality of search engine repository item embeddings,
(iii) generating, via N hops within a semantic graph starting from a plurality of nodes respectively associated with the plurality of initial search result sets, a plurality of related search results, and
(iv) generating the prediction-based action for the query input based at least in part on the plurality of related search results.
13. The system of claim 12 , wherein a search engine repository item of a plurality of search engine repository items respectively corresponding to the plurality of search engine repository item embeddings comprises metadata corresponding to a content category label, and a search engine repository item embedding is generated for the search engine repository item using an embedding technique corresponding to the content category label.
14. The system of claim 13 , wherein the embedding technique comprises generating a syntactic embedding by:
generating tokenized content by tokenizing content of the search engine repository item with subword n-grams;
training a term-frequency inverse document frequency (TF-IDF) model on the tokenized content; and
removing one or more bottom tokens with low TF-IDF values from the tokenized content.
15. The system of claim 13 , wherein the embedding technique comprises generating a semantic embedding by:
retrieving a domain corpus; and
tuning a universal sentence encoding model on the domain corpus.
16. The system of claim 13 , wherein the embedding technique comprises generating a geohash embedding by:
converting geospatial information into one or more geocodes according to a geohashing algorithm;
identifying, from a data source, procedure geocode pairs based at least in part on procedures associated with respective ones of the one or more geocodes; and
training a skip-gram model based at least in part on the procedure geocode pairs.
17. The system of claim 12 , wherein the plurality of content categories comprises at least one of: a syntactic content category, a semantic content category, and the geospatial content category.
18. The system of claim 12 , wherein the operations further comprise:
generating the semantic graph based at least in part on a measure of similarity for pairs of the plurality of search engine repository item embeddings respectively corresponding to a plurality of search engine repository items by:
retrieving a plurality of semantic embeddings respectively corresponding to the plurality of search engine repository items;
computing the measure of similarity between pairs of the plurality of semantic embeddings;
generating edges between the pairs of the plurality of search engine repository item embeddings in the semantic graph based at least in part on the measure of similarity; and
removing given ones of the edges according to a threshold value.
19. The system of claim 12 , wherein the plurality of query input embeddings further comprises a user embedding that is based at least in part on historical information and activities data associated with a user that provides the query input.
20. One or more non-transitory computer-readable storage media comprising instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
receiving a query input; and
generating, using a search engine machine learning model, a prediction-based action for the query input by:
(i) generating a plurality of query input embeddings of the query input that respectively correspond to a plurality of content categories, wherein the plurality of content categories comprises a geospatial content category and the plurality of query input embeddings comprises a query geohash embedding corresponding to the query input;
(ii) identifying a plurality of initial search result sets, using a k-Nearest-Neighbor (KNN) search, respectively corresponding to the plurality of content categories based at least in part on a comparison between the plurality of query input embeddings and a plurality of search engine repository item embeddings,
(iii) generating, via N hops within a semantic graph starting from a plurality of nodes respectively associated with the plurality of initial search result sets, a plurality of related search results, and
(iv) generating the prediction-based action for the query input based at least in part on the plurality of related search results.