US 10,372,743 B2Granted

Systems and methods for homogeneous entity grouping

AssigneeBaidu USA LLC

InventorsShulong Tan, Hongliang Fei, Yi Zhen, Yu Cao, Bocong Liu, Chaochun Liu, Richard Chun Ching Wang, Dawen Zhou, Wei Fan

Life of this patent

Estimating the term…

patentgrep

Loading patent…

Abstract

Systems and methods are disclosed to identify entities that have a similar meaning, and may, in embodiments, be grouped into entity groups for knowledge base construction. In embodiments, the entity relations of similarity or non-similarity for an entity pair are predicted as a binary relationship. In embodiments, the prediction may be based upon similarity score between the entities and the entity features, which features are constructed using an entity feature or representation model. In embodiments, the prediction may be an iterative process involving minimum human checking and existing knowledge update. In embodiments, one or more entity groups are formed using graph search from the predicted entity pairs. In embodiments, a group centroid entity may be selected to represent each group based on one or more factors, such as its generality or popularity.

Claims20

Dependent claims

A computer-implemented method to group synonymous entities from among a set of entities, based at least in part on their semantics, the method comprising:

for each entity pair from a set of entity pairs formed from the set of entities, determining a similarity score based upon character similarity between the entities that form the entity pair;
using a prior knowledge dataset to identify one or more sets of similar entities having same or similar meanings as indicated in the prior knowledge dataset, the one or more sets of similar entities comprising entities from an entity dictionary, the entity dictionary comprising the set of entities to be grouped relative to whether they have a same or similar meaning;
using large scale text data to construct embeddings of entities in the set of entities in which embedding training using the large scale text data is constrained using at least some of the sets of similar entities so that entities from a set of similar entities are represented by a same or similar embedding;
for each entity pair from the set of entity pairs, predicting entity relations between the entities that form the entity pair using as a feature set the entity embeddings of the entities that form the entity pair and the similarity score for that entity pair, the prediction being a binary prediction of positive relation of similarity or non-positive relation of similarity;
responsive to one or more criteria not being met:
sampling a set of predicted entity relations for verification;
using verification results to update the prior knowledge dataset;
and
returning to the step of using the prior knowledge dataset to identify one or more sets of similar entities having same or similar meanings to re-identify sets of similar entities, re-construct the embeddings of the entities using the re-identified sets of similar entities as constraints to the embedding training, and re-predict entity relations;
responsive to one or more criteria being met, outputting the entity pairs with positive relation predictions;
and
using the positive relation predictions to group the entities from the set of entities into one or more entity groups, each entity group comprising entities that have a positive entity relationship linking the entities in the entity group to have a same or similar meaning.

9 more elements

A computer-implemented method for identifying similar entities, the method comprising:

[a] constructing entity features, using an entity feature constructor, for each entity of a plurality of entities using large scale text data and based on prior knowledge, which identifies at least some of the entities from the plurality of entities as having a same or similar meaning, the entity feature constructor using at least some of the entities that have been identified as having a same or similar meaning to constrain entity feature construction so that entities that have been identified as having a same or similar meaning are represented by same or similar entity features;
[b] for each entity pair from a set of entity pairs formed from entities from the plurality of entities, predicting, using an entity relation predictor and at least entity features for two entities of the entity pair, an entity relation regarding whether the two entities that form the entity pair share a similar meaning;
[c] randomly sampling some entity pairs for entity relations verification and verifying if the entity relation for each randomly sampled entity pair is correct;
[d] using the verified entity relations to update the prior knowledge;
[e] re-constructing at least some of the entity features using the updated prior knowledge of entities that have been identified as having a same or similar meaning to constrain entity feature construction so that entities that have been identified as having a same or similar meaning are represented by same or similar entity features;
[f] re-predicting at least some of the entity relations using the re-constructed entity features;
[g] iterating steps [d], [e], and [f] until a difference in an entity relations change measure between an evaluation set of entity pairs and their corresponding entity relations for a current iteration and their corresponding entity relations from a prior iteration is less than a threshold;
and
[h] outputting each entity pair that has an entity relation that indicates that the entities in the entity pair have a same or similar meaning.

6 more elements

A non-transitory computer-readable medium or media comprising one or more sequences of instructions which, when executed by one or more processors, causes the steps to be performed comprising:

identifying, using an existing knowledge dataset, one or more sets of similar entities, in which a set of similar entities comprises entities from an entity dictionary that have a same or similar meaning as indicated by the existing knowledge dataset;
for each entity pair from a set of entity pairs formed from entities in the entity dictionary, determining a similarity score based upon character similarity between the entities that form the entity pair;
for each entity in the entity dictionary, constructing an embedding using at least some of the one or more sets of similar entities as construction constraints to have entities in a set of similar entities represented with a same or similar embedding;
for each entity pair from the set of entity pairs formed from entities in the entity dictionary, predicting the entity relations between the two entities of the entity pair based on at least the embeddings of the two entities and the similarity score of the entity pair, the prediction being a prediction of whether the two entities of the entity pair are a related entity pair or not related entity pair;
outputting related entity pairs after the prediction;
grouping the related entity pairs as one or more entity groups, each entity group comprising entities in the group that are all related;
and
selecting a group representative for each entity group.

5 more elements

Dependencies are inferred from the claim text.