Apparatus and method for optimized search results using document and passage-level integration
View Patent ↗A method is presented for enhancing search results by segmenting documents into smaller passages and utilizing those passages as the search unit. This method integrates the passage rankings from two search models to produce a new document ranking and arranges the documents accordingly. The ranking of document is also rearranged based on the proportion of passages taken from the same document versus the total number of passages in that document. The final search ranking system combines document-level and passage-level search rankings to rank documents. This method ensures that when conducting a passage search, the search results consider the general relevance of the entire document, which leads to better performance compared to searching only by document-level or passage-level searches.
1 . A computer-implemented method for providing a user with search results corresponding to a query entered by the user based on a document corpus including a plurality of documents to be searched and a passage corpus including a plurality of passages extracted from each document of the document corpus, the method comprising:
(a) extracting and arranging, by a document-level search model, D documents from the document corpus corresponding to the query;
(b) extracting and arranging, by a passage-level search model, a globally top-N set of passages across the passage corpus corresponding to the query, the globally top-N set being determined as top-N with respect to passage relevance to the query;
(c) arranging M documents containing the N passages in a rank corresponding to the rank in which the N passages are arranged, wherein M is less than or equal to N;
(d) arranging the M documents based on a relationship between the number of passages extracted from a single document among said N passages and the total number of passages in the single document; and
(e) determining a final ranking of documents by integrating the results of arranging for the D documents in step (a) and the results of the arranging for the M documents in step (d), by performing rank fusion to integrate rankings, the rank fusion being configured to operate even when the result lists are non-overlapping and/or have different sizes (D≠M);
wherein, in step (d), the M documents are arranged either in an order from documents with a smaller value given by the relation (Np−np)/Np to documents with a larger value, or in an order from documents with a larger value given by the relation np/Np to documents with a smaller value, where np represents a number of passages extracted from a particular document among said N passages, and Np represents a total number of passages in the particular document.
2 . The method of claim 1 ,
Wherein the passage-level search module of step (b) includes a first search model having a relatively high recall and a relatively fast processing speed and a second search model having a relatively high precision and a relatively slow processing speed, and
wherein step (b) includes:
(b1) extracting and arranging, by the first search model, N passages from the passage corpus in correspondence with the query;
(b2) re-ranking, by the second search model, the N passages extracted in step (b1) based on query; and
(b3) integrating the results from step (b1) and step (b2) for the N passages to generate a final ranking for the N passages.
3 . The method of claim 2 , wherein the integration in step (b3) is performed by Reciprocal Rank Fusion (RRF).
4 . The method of claim 1 , wherein the integration in step (d) is made by Reciprocal Rank Fusion (RRF).
5 . An apparatus for providing a user with search results corresponding to a query entered by the user from a passage corpus comprising a plurality of passages extracted from each document of a document corpus, comprising:
at least one processor; and
at least one memory for storing computer-executable instructions,
wherein the computer-executable instructions stored in the at least one memory make the at least one processor to perform the following steps:
(a) extracting and arranging, by a document-level search model, D documents corresponding to the query from the document corpus;
(b) extracting and arranging, by a passage-level search model a globally top-N set of passages across the passage corpus corresponding to the query, the globally top-N set being determined as top-N with respect to passage relevance to the query;
(c) arranging M documents containing the N passages in a rank corresponding to the rank in which the N passages are arranged, wherein M is less than or equal to N;
(d) arranging the M documents based on a relationship between the number of passages extracted from a single document among said N passages and the total number of passages in the single document; and
(e) determining a final ranking of documents by integrating the results of arranging for the D documents in step (a) and the results of the arranging for the M documents in step (d), by performing rank fusion to integrate rankings, the rank fusion being configured to operate even when the result lists are non-overlapping and/or have different sizes (D≠M);
wherein, in step (d), the M documents are arranged either in an order from documents with a smaller value given by the relation (Np−np)/Np to documents with a larger value, or in an order from documents with a larger value given by the relation np/Np to documents with a smaller value, where np represents a number of passages extracted from a particular document among said N passages, and Np represents a total number of passages in the particular document.
6 . The apparatus of claim 5 , wherein the passage-level search model includes a first search model having a relatively high recall and a relatively fast processing speed and a second search model having a relatively high precision and a relatively slow processing speed.
7 . The apparatus of claim 6 , wherein at least one of the first search model and the second search model is an artificial intelligence based search model.
8 . The apparatus of claim 5 ,
wherein at least part of the documents in the document corpus has titles, and
wherein each passage of the passage corpus includes the title of a document of which said each passage is a part.