Automated enrichment of entity descriptions in unstructured text
Automatically enriching the descriptions of an entity mentioned in a sentence corpus includes generating multiple enriched descriptions corresponding to a label of the entity. Each of the multiple enriched descriptions is ranked. The ranking is generated by a machine learning model that is configured to determine a likelihood that an enriched description correctly describes the entity. The sentence corpus are annotated by coupling each mention of the entity with one or more of the enriched descriptions. The one or more enriched descriptions are selected based on the ranking. As annotated, the sentence corpus can be output.
1 . A method, comprising:
generating, by a processor of a computer, a plurality of enriched descriptions corresponding to an entity, wherein the entity is mentioned in a sentence corpus input to the computer;
computing, by the processor, using a machine learning model, and based on determining a probability that each mention of the entity in the sentence corpus belongs to one of a plurality of classes, an entropy associated with the plurality of enriched descriptions,
wherein the machine learning model is configured to learn without additional training data, and
wherein the machine learning model identifies and extracts the entity from the sentence corpus based on inferences;
ranking, by the processor, using a semantic similarity model, and based on computing the entropy using the machine learning model, each of the plurality of enriched descriptions, wherein a ranking of an enriched description, of the plurality of enriched descriptions, corresponds to likelihood that a machine learning classifier correctly classifies the entity mentioned in the sentence corpus using information from the enriched description;
selecting, based on the ranking, a subset of the plurality of enriched descriptions that reduces the entropy; and
outputting the subset of the plurality of enriched descriptions selected.
2 . The method of claim 1 ,
wherein the likelihood is determined based on a semantic similarity of embedding space encodings of each of the enriched descriptions and the entity.
3 . The method of claim 1 ,
wherein the method further comprises:
revising the ranking in response to user-provided feedback,
wherein the user-provided feedback includes at least one of a user-generated description or a re-ranking of the plurality of enriched descriptions.
4 . The method of claim 1 ,
wherein the generating includes enriching one or more initial descriptions input to the computer by a user.
5 . The method of claim 4 ,
wherein the enriching one or more initial descriptions includes extending an initial description using a language model that predicts a next word based on one or more previous words.
6 . The method of claim 4 ,
wherein the enriching one or more initial descriptions includes generating a new description using a language model that paraphrases an initial description, wherein language model is trained as a sequence-to-sequence generator.
7 . The method of claim 1 ,
wherein the generating includes automatically generating the enriched descriptions in response to an input of the entity without at least one initial description.
8 . The method of claim 7 ,
wherein the automatically generating is performed using a template and language model configured to generate an enriched description based on the template.
9 . The method of claim 7 ,
wherein the automatically generating is performed using a template and language model configured to use sequence-to-sequence text generation conditioned on an entity label.
10 . A system, comprising:
one or more processors configured to initiate operations including:
generating a plurality of enriched descriptions corresponding to an entity mentioned in a sentence corpus input to a computer;
computing, using a zero-shot machine learning model, and based on determining a probability that each mention of the entity in the sentence corpus belongs to one of a plurality of classes, an entropy associated with the plurality of enriched descriptions;
ranking, using a semantic similarity model and based on computing the entropy using the zero-shot machine learning model, each of the plurality of enriched descriptions, wherein a ranking of an enriched description, of the plurality of enriched descriptions, corresponds to likelihood that a machine learning classifier correctly classifies the entity mentioned in the sentence corpus using information from the enriched description;
selecting, based on the ranking, a subset of the plurality of enriched descriptions that reduces the entropy; and
outputting the subset of the plurality of enriched descriptions selected.
11 . The system of claim 10 ,
wherein the likelihood is determined based on a semantic similarity of embedding space encodings of each of the enriched descriptions and the entity.
12 . The system of claim 10 ,
wherein the one or more processors are configured to initiate operations further including:
revising the ranking in response to user-provided feedback,
wherein the user-provided feedback includes at least one of a user-generated description or a re-ranking of the plurality of enriched descriptions.
13 . The system of claim 10 ,
wherein the generating includes enriching one or more initial descriptions input to the computer by a user.
14 . The system of claim 13 ,
wherein the enriching one or more initial descriptions includes extending an initial description using a language model that predicts a next word based on one or more previous words.
15 . The system of claim 13 ,
wherein the enriching one or more initial descriptions includes generating a new description using a language model that paraphrases an initial description, wherein language model is trained as a sequence-to-sequence generator.
16 . The system of claim 10 ,
wherein the generating includes automatically generating the enriched descriptions in response to an input of the entity without at least one initial description.
17 . The system of claim 16 ,
wherein the automatically generating is performed using a template and language model configured to generate an enriched description based on the template.
18 . A computer program product, the computer program product comprising:
one or more computer-readable storage media and program instructions collectively stored on the one or more computer-readable storage media, the program instructions executable by a processor to cause the processor to initiate operations including:
generating a plurality of enriched descriptions corresponding to an entity mentioned in a sentence corpus input to a computer;
computing, using a machine learning model, and based on determining a probability that each mention of the entity in the sentence corpus belongs to one of a plurality of classes, an entropy associated with the plurality of enriched descriptions;
ranking, using a semantic similarity model based on computing the entropy using the machine learning model, each enriched description of the plurality of enriched descriptions, wherein a ranking of an enriched description, of the plurality of enriched descriptions, corresponds to likelihood that a machine learning classifier correctly classifies the entity mentioned in the sentence corpus using information from the enriched description;
selecting, based on the ranking, a subset of the plurality of the enriched descriptions that reduces the entropy; and
outputting the subset of the plurality of enriched descriptions selected.
19 . The computer program product of claim 18 ,
wherein the machine learning model identifies and extracts the entity from the sentence corpus based on inferences.
20 . The computer program product of claim 18 ,
wherein the likelihood is determined based on a semantic similarity of embedding space encodings of each of the enriched descriptions and the entity.