KNOWLEDGE DISCOVERY FROM CITATION NETWORKS

• US 20150186789A1
• Filed: 12/29/2014
• Published: 07/02/2015
• Est. Priority Date: 12/06/2010
• Status: Active Grant

• Current Assignee: The Research Foundation for The State University of New York (State University of New York)
• Original Assignee: The Research Foundation for The State University of New York (State University of New York)
• Inventors: Guo, Zhen, Zhang, Mark
• Granted Patent: US 9,269,051 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/24573 using data annotations, e.g. user-defined metadata
• CPC Class Codes: G06F 16/24578 using ranking
• CPC Class Codes: G06F 16/93 Document management systems
• CPC Class Codes: G06F 16/9558 Details of hyperlinks; Management of linked annotations
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G06Q 10/10 Office automation; Time management

Abstract

In a corpus of scientific articles such as a digital library, documents are connected by citations and one document plays two different roles in the corpus: document itself and a citation of other documents. A Bernoulli Process Topic (BPT) model is provided which models the corpus at two levels: document level and citation level. In the BPT model, each document has two different representations in the latent topic space associated with its roles. Moreover, the multi-level hierarchical structure of the citation network is captured by a generative process involving a Bernoulli process. The distribution parameters of the BPT model are estimated by a variational approximation approach.

Claims

1. A method of modeling a set of documents, each document within the set of documents comprising content, and at least a portion of the documents being linked to other documents through citations, comprising:

defining a Bernoulli Process Topic model for the set of documents representing each document as a represented as a mixture over latent topics, dependent on (1) an intrinsic content of each respective document, and (2) a content of other documents related to the respective document through a multi-level citation structure of direct and indirect linkages to the other documents;

for each respective document, determining a first set of latent topic characteristics based on an intrinsic content of the respective document;

for each document, determining a second set of latent topic characteristics based on a respective content of the other documents which are directly and indirectly linked, the indirectly linked documents contributing transitively to the latent topic characteristics of the respective document;

representing a third set of latent topics for the respective document based on a joint probability distribution of at least the first and second sets of latent topic characteristics, Bernoulli Process Topic model; and

outputting at least one document in response to an input, based on at least the third set of latent topics.

2. The method according to claim 1, further comprising ranking at last a portion of the set of documents, in dependence on at least the input and the third set of latent topics.

3. The method according to claim 1, further comprising determining a Bayesian network structure based on at least one linkage, wherein each linkage implies a content relationship of the linked documents.

4. The method according to claim 1, further comprising iteratively determining for each location in each document, a topic from a topic distribution of the document having a Dirichlet probability distribution, and a word which follows a multinomial distribution conditioned on the determined topic for the respective location.

5. The method according to claim 1, further comprising, for each document at a document level, and for each location in each document, iteratively until convergence within a convergence criterion is achieved:

choosing a document linked to the respective document based on the location in the document and a multinomial distribution conditioned on the respective document,

choosing a topic from a topic distribution of the linked document based on the respective location in the respective document, and

choosing a word which follows the multinomial distribution conditioned on the chosen topic.

6. A method for characterizing a corpus of documents, each having one or more semantically significant citation linkages between respective documents, comprising:

identifying a multilevel hierarchy of directly cited documents and indirectly cited documents;

for each respective document, determining latent topic characteristics based on an intrinsic semantic content of the respective document, semantic content associated with directly cited documents, and semantic content associated with indirectly cited documents, wherein a semantic content significance of a citation has a transitive property;

representing latent topics for documents within the corpus based on a joint probability distribution of the latent topic characteristics; and

storing, in a memory, the represented set of latent topics.

7. The method according to claim 6, wherein relationships among the documents of the corpus of documents are modeled by a Bernoulli random process, such that a topic distribution of each respective document is a mixture of distributions associated with the directly linked documents and the indirectly linked documents.

8. The method according to claim 7, wherein a joint probability distribution is estimated by an iterative process at a citation level, comprising,

for each document d_j, and for the i-th location in document d_j,

choosing a topic z_ji from a topic distribution of document d_j, p(z|d_j,θ

_dj) where the distribution parameter θ

_dj is drawn from a Dirichlet distribution Dir(α

),

choosing a word w_ji which follows a multinomial distribution p(w|z_ji,Λ

) conditioned on the topic z_ji, and

respectively incrementing the locations i and documents d_j to process each location in each document

9. The method according to claim 8, wherein a joint probability distribution is estimated by an iterative process at a document level comprising,

for each document d_s, and for the i-th location in each document d_s,

choosing a cited document c_si from p(c|d_s,Ξ

), a multinomial distribution conditioned on the document d_s,

choosing a topic t_si from a topic distribution of the document c_si at a citation level, and

choosing a word w_si which follows the multinomial distribution p(w|t_si,Λ

) conditioned on the topic t_si,

wherein Ξ

is a mixing coefficient matrix which represents how much of a content of a respective document is from direct citations and how much of the content is from indirect citations, and

a composition of Ξ and

θ

models the topic distribution at the document level.

10. The method according to claim 9, wherein:

the mixing coefficient matrix Ξ

is an N×

N matrix, where Ξ

_js=p(c_si=d_j|d_s), which is treated as a fixed quantity computed from the citation information of the corpus of documents;

K is a number of latent topics;

matrix Θ

is a K×

N matrix representing a set of topic distributions at a citation level, where Θ

_lj=p(z_ji=l|d_j),

matrix Λ

is a M×

K matrix representing word probability, where Λ

_hl=p(w_si^h=1|t_si=l), and

each document d_s has a set of citations Q_ds,

the method further comprising constructing a matrix S to denote direct relationships among the documents d_s wherein

11. A method for determining a set of latent topics of a Bayesian network of multilevel hierarchically related documents having direct and indirect references associated with content relationships, comprising:

for each respective document, determining a set of latent topic characteristics based on an intrinsic content of the respective document and a set of latent topic characteristics based on a respective content of other documents which are directly referenced and indirectly referenced through at least one other document to the respective document; and

representing a set of latent topics for each respective document based on a joint probability distribution of at least the latent topic characteristics based on the intrinsic content and the respective content of other documents which are directly referenced and indirectly referenced through at least one other document to the respective document, dependent on the identified network and a random process.

12. The method according to claim 11, wherein relationships among the Bayesian network of multilevel hierarchically related documents are modeled by a Bernoulli process such that a latent topic distribution of each respective document is a mixture of distributions associated with the related document.

13. The method according to claim 11, wherein the Bayesian network of multilevel hierarchically related documents is modeled by a generative probabilistic model of a latent topic content of the set of documents along with the references among the Bayesian network of multilevel hierarchically related documents.

14. The method according to claim 11, wherein relationships among the Bayesian network of multilevel hierarchically related documents are modeled by a Bernoulli process such that a topic distribution of each respective document is a mixture of distributions associated with documents linked to the respective documents.

15. The method according to claim 11, wherein the latent topics are modeled at both a document level and a citation level, and distinctions in the multilevel hierarchical network are captured by a Bernoulli random process.

16. The method according to claim 11, wherein the Bayesian network of multilevel hierarchically related documents is modeled by a generative probabilistic model of a topic content of each document of the Bayesian network of multilevel hierarchically related documents along with the linkages among members of the Bayesian network of multilevel hierarchically related documents.

17. The method according to claim 11, further comprising:

performing an iterative process for each document referenced by a respective document, and for each location in each document, comprising choosing a topic from a Dirichlet topic distribution of a respective document, choosing a word which follows a multinomial distribution conditioned on the topic; and

performing an iterative process for each document, and for each location in each document, comprising choosing a referenced document from a multinomial distribution conditioned on the respective document, choosing a topic from the topic distribution of the document at the referenced document level, and choosing a word which follows a multinomial distribution conditioned on the topic.

18. The method according to claim 11, further comprising:

performing an iterative process for each document d_j referenced by a respective document, and for each location i in each document d_j, comprising;

choosing a topic z_ji from a topic distribution of document d_j, p(z|d_j,θ

_dj) where the distribution parameter θ

_dj is drawn from a Dirichlet distribution Dir(α

),

choosing a word w_ji which follows a multinomial distribution p(w|z_ji,Λ

) conditioned on the topic z_ji,

wherein a composition of θ

models the topic distribution at the document level;

performing an iterative process for each document d_s, and for each location i in each document d_s, comprising;

choosing a referenced document c_si from p(c|d_s,Ξ

), a multinomial distribution conditioned on the document d_s,

choosing a topic t_si from the topic distribution of the document c_si at the referenced document level, and

choosing a word w_si which follows a multinomial distribution p(w|t_si,Λ

) conditioned on the topic t_si,

wherein Ξ

is a mixing coefficient matrix which represents contributions to the content of the respective document from direct references and indirect references, and a composition of Ξ

models the topic distribution at the document level,

wherein a number of latent topics is K and the mixing coefficients are parameterized by an N×

N matrix Ξ

where Ξ

_js=p(c_si=d_j|d_s), which are treated as a fixed quantity computed from the reference information of the network of multilevel hierarchically related documents, and

wherein topic distributions at a level of the referenced documents are parameterized by a K×

N matrix Θ

where Θ

_lj=p(z_ji=l|d_j), and an M×

K word probability matrix Λ

, where Λ

_hl=p(w_si^h=1|t_si=l),

wherein each document d_s has a set of references Q_ds,

further comprising constructing a matrix S to denote direct relationships among the documents wherein

19. The method according to claim 11, further comprising using the set of latent topics to rank at least a portion of the Bayesian network of multilevel hierarchically related documents.

20. The method according to claim 11, further comprising using the set of latent topics as an index for the Bayesian network of multilevel hierarchically related documents.