Summarizing collections of documents
Provided is a process including: obtaining a corpus having a plurality of document collections, each of which is associated with features; for a given document collection, computing a pertinence score for each feature; ranking the features based on the features' pertinence scores; selecting a first set of features based on a first coverage score thereof and a threshold; re-ranking the first set of features based on the features' relevance to the document collection; and selecting a second set of features from the first set of features based on a second coverage score thereof and the threshold, the second set of features being used for summarizing the document collection.
1. A computer-implemented method of summarizing a collection of documents in a corpus, the method comprising:
obtaining a corpus comprising a plurality of document collections, wherein each of the document collections is associated with a plurality of features obtained from documents in the respective document collection, each of the plurality of document collection having more than 100 documents, the plurality of features comprising n-grams, topic keywords, document classification codes, or a combination thereof;
for a given document collection among the plurality of document collections, computing, by one or more processors, a pertinence score for each of the plurality of features of the given document collection, wherein the pertinence score is indicative of pertinence of the respective feature to the given document collection relative to other document collections in the corpus;
ranking, by one or more processors, the plurality of features of the given document collection based on the features' pertinence scores;
selecting, by one or more processors, a plurality of highest ranked features from the ranked plurality of features of the given document collection to form a first set of selected features, wherein the first set of selected features are selected, at least in part, by determining a first coverage score for the first set of selected features is above a threshold, and wherein the first coverage score is indicative of coverage of documents in which the first set of selected features occur among the documents in the given document collection, wherein selecting a plurality of highest ranked features comprises:
(i) for a given set of features with n consecutive highest pertinence scores, where n is a natural number, computing an amount of the number of documents in which the given set of features occur relative to the number of the documents in the given document collection; and
(ii) in response to determining the amount is above the threshold, selecting the given set of features; and in response to determining the amount is not above the threshold, updating the given set of features by adding another feature with the (n+1)th highest pertinence score and repeating the computing and determining operations (i) and (ii) at least once;
re-ranking, by one or more processors, the first set of selected features based on the features' relevance to the given document collection; and
selecting, by one or more processors, a plurality of highest re-ranked features from the first set of selected features to form a second set of selected features, wherein the second set of selected features are selected, at least in part, by determining a second coverage score for the second set of selected features is above the threshold, wherein the second coverage score is indicative of coverage of documents in which the second set of selected features occur among the documents in the given document collection, and wherein the second set of selected features are used for summarizing the document collection.
2. The method of claim 1 , wherein obtaining a corpus comprises:
retrieving each of documents in the corpus, the corpus having more than 1,000 documents with a median word length of more than 500 words;
generating a feature vector for each of the retrieved documents, wherein the feature vector includes a plurality of n-grams in text of the document, a topic keyword describing content of the document, a document classification code, or a combination thereof; and
clustering the retrieved documents into the plurality of document collections.
3. The method of claim 1 , wherein computing a pertinence score for each of the plurality of features comprises:
computing a first frequency in which the feature occurs in the given document collection;
computing a second frequency in which the feature occurs in the other document collections in the corpus; and
computing a product of the first frequency times the logarithm of the inverse of the second frequency to generate the pertinence score.
4. The method of claim 1 , wherein re-ranking the first set of selected features comprises:
swapping rankings of first and second selected features in response to determining that a third coverage score for a third set of selected features including the first selected feature is smaller than a fourth coverage score for a fourth set of selected features including the second selected feature, and wherein the first selected feature is ranked higher than the second selected feature.
5. The method of claim 4 , wherein
the third set of selected features includes the first selected feature and one or more selected features with consecutive highest pertinence scores higher than the pertinence score of the first selected feature; and
the fourth set of selected features includes the second selected feature and the one or more selected features with consecutive highest pertinence scores higher than the pertinence score of the first selected feature.
6. The method of claim 5 , wherein the swapping operation is repeated for each pair of selected features ranked consecutively.
7. The method of claim 1 , wherein obtaining a corpus comprising a plurality of document collections comprises:
performing steps for clustering documents; and
performing steps for calculating feature vectors.
8. A computer-implemented method of summarizing a collection of documents in a corpus, the method comprising:
obtaining a corpus comprising a plurality of document collections, wherein each of the document collections is associated with a plurality of features obtained from documents in the respective document collection, each of the plurality of document collection having more than 100 documents, the plurality of features comprising n-grams, topic keywords, document classification codes, or a combination thereof;
for a given document collection among the plurality of document collections, computing, by one or more processors, a pertinence score for each of the plurality of features of the given document collection, wherein the pertinence score is indicative of pertinence of the respective feature to the given document collection relative to other document collections in the corpus;
ranking, by one or more processors, the plurality of features of the given document collection based on the features' pertinence scores;
selecting, by one or more processors, a plurality of highest ranked features from the ranked plurality of features of the given document collection to form a first set of selected features, wherein the first set of selected features are selected, at least in part, by determining a first coverage score for the first set of selected features is above a threshold, and wherein the first coverage score is indicative of coverage of documents in which the first set of selected features occur among the documents in the given document collection;
re-ranking, by one or more processors, the first set of selected features based on the features' relevance to the given document collection; and
selecting, by one or more processors, a plurality of highest re-ranked features from the first set of selected features to form a second set of selected features, wherein the second set of selected features are selected, at least in part, by determining a second coverage score for the second set of selected features is above the threshold, wherein the second coverage score is indicative of coverage of documents in which the second set of selected features occur among the documents in the given document collection, and wherein the second set of selected features are used for summarizing the document collection, wherein selecting a plurality of highest re-ranked features comprises:
(i) for a given set of selected features with n consecutive highest pertinence scores, where n is a natural number, computing an amount of the number of documents in which the given set of selected features occur relative to the number of the documents in the given document collection; and
(ii) in response to determining the amount is above the threshold, selecting the given set of selected features; and in response to determining the amount is not above the threshold, updating the given set of selected features by adding another selected feature with the (n+1)th highest pertinence score and repeating the computing and determining operations (i) and (ii) at least once.
9. A system, comprising:
one or more processors; and
memory storing instructions that when executed by at least some of the processors effectuate operations comprising:
obtaining a corpus comprising a plurality of document collections, wherein each of the document collections is associated with a plurality of features obtained from documents in the respective document collection, the plurality of features comprising n-grams, topic keywords, document classification codes, or a combination thereof;
for a given document collection among the plurality of document collections, computing a pertinence score for each of the plurality of features of the given document collection, wherein the pertinence score is indicative of pertinence of the respective feature to the given document collection relative to other document collections in the corpus;
ranking the plurality of features of the given document collection based on the features' pertinence scores;
selecting a plurality of highest ranked features from the ranked plurality of features of the given document collection to form a first set of selected features, wherein the first set of selected features are selected, at least in part, by determining a first coverage score for the first set of selected features is above a threshold, and wherein the first coverage score is indicative of coverage of documents in which the first set of selected features occur among the documents in the given document collection, wherein selecting a plurality of highest ranked features comprises:
(i) for a given set of features with n consecutive highest pertinence scores, where n is a natural number, computing an amount of the number of documents in which the given set of features occur relative to the number of the documents in the given document collection; and
(ii) in response to determining the amount is above the threshold, selecting the given set of features; and in response to determining the amount is not above the threshold, updating the given set of features by adding another feature with the (n+1)th highest pertinence score and repeating the computing and determining operations (i) and (ii) at least once;
re-ranking the first set of selected features based on the features' relevance to the given document collection; and
selecting a plurality of highest re-ranked features from the first set of selected features to form a second set of selected features, wherein the second set of selected features are selected, at least in part, by determining a second coverage score for the second set of selected features is above the threshold, wherein the second coverage score is indicative of coverage of documents in which the second set of selected features occur among the documents in the given document collection, and wherein the second set of selected features are used for summarizing the document collection.
10. The system of claim 9 , wherein obtaining a corpus comprises:
retrieving each of documents in the corpus, the corpus having more than 1,000 documents with a median word length of more than 500 words;
generating a feature vector for each of the retrieved documents, wherein the feature vector includes a plurality of n-grams in text of the document, a topic keyword describing content of the document, a document classification code, or a combination thereof; and
clustering the retrieved documents into the plurality of document collections.
11. The system of claim 9 , wherein computing a pertinence score for each of the plurality of features comprises:
computing a first frequency in which the feature occurs in the given document collection;
computing a second frequency in which the feature occurs in the other document collections in the corpus; and
computing a product of the first frequency times the logarithm of the inverse of the second frequency to generate the pertinence score.
12. The system of claim 9 , wherein re-ranking the first set of selected features comprises:
swapping rankings of first and second selected features in response to determining that a third coverage score for a third set of selected features including the first selected feature is smaller than a fourth coverage score for a fourth set of selected features including the second selected feature, and wherein the first selected feature is ranked higher than the second selected feature.
13. The system of claim 12 , wherein
the third set of selected features includes the first selected feature and one or more selected features with consecutive highest pertinence scores higher than the pertinence score of the first selected feature; and
the fourth set of selected features includes the second selected feature and the one or more selected features with consecutive highest pertinence scores higher than the pertinence score of the first selected feature.
14. The system of claim 13 , wherein the swapping operation is repeated for each pair of selected features ranked consecutively.
15. The system of claim 9 , wherein obtaining a corpus comprising a plurality of document collections comprises:
performing steps for clustering documents; and
performing steps for calculating feature vectors.
16. A system, comprising:
one or more processors; and
memory storing instructions that when executed by at least some of the processors effectuate operations comprising:
obtaining a corpus comprising a plurality of document collections, wherein each of the document collections is associated with a plurality of features obtained from documents in the respective document collection, the plurality of features comprising n-grams, topic keywords, document classification codes, or a combination thereof;
for a given document collection among the plurality of document collections, computing a pertinence score for each of the plurality of features of the given document collection, wherein the pertinence score is indicative of pertinence of the respective feature to the given document collection relative to other document collections in the corpus;
ranking the plurality of features of the given document collection based on the features' pertinence scores;
selecting a plurality of highest ranked features from the ranked plurality of features of the given document collection to form a first set of selected features, wherein the first set of selected features are selected, at least in part, by determining a first coverage score for the first set of selected features is above a threshold, and wherein the first coverage score is indicative of coverage of documents in which the first set of selected features occur among the documents in the given document collection;
re-ranking the first set of selected features based on the features' relevance to the given document collection; and
selecting a plurality of highest re-ranked features from the first set of selected features to form a second set of selected features, wherein the second set of selected features are selected, at least in part, by determining a second coverage score for the second set of selected features is above the threshold, wherein the second coverage score is indicative of coverage of documents in which the second set of selected features occur among the documents in the given document collection, and wherein the second set of selected features are used for summarizing the document collection, wherein selecting a plurality of highest re-ranked features comprises:
(i) for a given set of selected features with n consecutive highest pertinence scores, where n is a natural number, computing an amount of the number of documents in which the given set of selected features occur relative to the number of the documents in the given document collection; and
(ii) in response to determining the amount is above the threshold, selecting the given set of selected features; and in response to determining the amount is not above the threshold, updating the given set of selected features by adding another selected feature with the (n+1)th highest pertinence score and repeating the computing and determining operations (i) and (ii) at least once.