Linking tabular columns to unseen ontologies
An embodiment for improved linking of tabular columns to column types in an ontology unseen during training. The embodiment may for a target table, encode a target tabular query column, table headers, and target types independently to generate permutation invariant representations of tabular data associated with the target table. The embodiment may, for each of the target types, extract and further encode auxiliary information. The embodiment may process the encoded tabular data to obtain a first vector and a second vector. The embodiment may concatenate the first vector and the second vector to generate a final query vector. The embodiment may process the encoded target types through a third transformer to obtain a third vector. The embodiment may calculate a score to model interactions between the target tabular query column of the target table and the target types.
1. A computer-based method of linking tabular columns to column types in an ontology unseen during training, the method comprising:
for a target table, encoding a target tabular query column, table headers, and target types independently to generate permutation invariant representations of type data associated with a target ontology and tabular data associated with the target table, wherein encoding the target types further includes encoding associated auxiliary information, wherein the auxiliary information includes a partial taxonomy structure comprising linearized two-hop is-a ancestor labels;
processing the encoded tabular query column using a first transformer to obtain a first vector and the encoded table headers using a second transformer to obtain a second vector;
concatenating the first vector and the second vector to obtain a combined vector;
processing the combined vector through a linear layer and a Gaussian Error Linear Unit layer to obtain a final query vector;
processing the encoded target types through a third transformer to obtain a third vector; and
calculating a score for the target tabular query column as a dot product between the final query vector and the third vector to model interactions between the target tabular query column of the target table and the target types and provide a column-type annotation.
2. The computer-based method of claim 1 , wherein encoding the target tabular query column, the table headers, and the target types independently to generate the permutation invariant representations of the type data and the tabular data associated with the target table further comprises:
linearizing the tabular data;
automatically assigning an increasing sequence of position identifiers to each non-delimiting tokenized cell in the linearized tabular data until a separator token is reached; and
in response to reaching the separator token, automatically assigning a monotonically increasing sequence of position identifiers for each non-delimiting tokenized cell positioned after the separator token, restarting from an integer corresponding to 1 greater than the position identifier assigned to the separator token.
3. The computer-based method of claim 1 , wherein the auxiliary information further includes one a glossary comprising a short description of the target label type.
4. The computer-based method of claim 1 , wherein encoding the target tabular query column, the table headers, and the target types independently to generate the permutation invariant representations of the type data and the tabular data associated with the target table further comprises:
assigning the target types, associated glossaries, associated parent labels, and associated grand-parent labels different token identifications.
5. The computer-based method of claim 4 , wherein the different token identifications are initialized to two pre-trained token type embeddings of a pre-trained transformer encoder model comprising two token type identifiers.
6. The computer-based method of claim 1 , further comprising:
leveraging a library for large-scale vector-similarity searches to increase a speed of an inferencing step.
7. The computer-based method of claim 1 , wherein the calculated score is determined using a MaxSim function.
8. A computer system, the computer system comprising:
one or more processors, one or more computer-readable memories, one or more computer-readable tangible storage medium, and program instructions stored on at least one of the one or more computer-readable tangible storage medium for execution by at least one of the one or more processors via at least one of the one or more computer-readable memories, wherein the computer system is capable of performing a method comprising:
for a target table, encoding a target tabular query column, table headers, and target types independently to generate permutation invariant representations of type data associated with a target ontology and tabular data associated with the target table, wherein encoding the target types further includes encoding associated auxiliary information, wherein the auxiliary information includes a partial taxonomy structure comprising linearized two-hop is-a ancestor labels;
processing the encoded tabular query column using a first transformer to obtain a first vector and the encoded table headers using a second transformer to obtain a second vector;
concatenating the first vector and the second vector to obtain a combined vector;
processing the combined vector through a linear layer and a Gaussian Error Linear Unit layer to obtain a final query vector;
processing the encoded target types through a third transformer to obtain a third vector; and
calculating a score for the target tabular query column as a dot product between the final query vector and the third vector to model interactions between the target tabular query column of the target table and the target types and provide a column-type annotation.
9. The computer system of claim 8 , wherein encoding the target tabular query column, the table headers, and the target types independently to generate the permutation invariant representations of the type data and the tabular data associated with the target table further comprises:
linearizing the tabular data;
automatically assigning an increasing sequence of position identifiers to each non-delimiting tokenized cell in the linearized tabular data until a separator token is reached; and
in response to reaching the separator token, automatically assigning a monotonically increasing sequence of position identifiers for each non-delimiting tokenized cell positioned after the separator token, restarting from an integer corresponding to 1 greater than the position identifier assigned to the separator token.
10. The computer system of claim 8 , wherein the auxiliary information further includes a glossary comprising a short description of the target label type.
11. The computer system of claim 8 , wherein encoding the target tabular query column, the table headers, and the target types independently to generate the permutation invariant representations of the type data and the tabular data associated with the target table further comprises:
assigning the target types, associated glossaries, associated parent labels, and associated grand-parent labels different token identifications.
12. The computer system of claim 11 , wherein the different token identifications are initialized to two pre-trained token type embeddings of a pre-trained transformer encoder model comprising two token type identifiers.
13. The computer system of claim 8 , further comprising:
leveraging a library for large-scale vector-similarity searches to increase a speed of an inferencing step.
14. The computer system of claim 8 , wherein the calculated score is determined using a MaxSim function.
15. A computer program product, the computer program product comprising:
one or more computer-readable storage medium and program instructions stored on at least one of the one or more computer-readable storage medium, the program instructions executable by a processor capable of performing a method, the method comprising:
for a target table, encoding a target tabular query column, table headers, and target types independently to generate permutation invariant representations of type data associated with a target ontology and tabular data associated with the target table, wherein encoding the target types further includes encoding associated auxiliary information, wherein the auxiliary information includes a partial taxonomy structure comprising linearized two-hop is-a ancestor labels;
processing the encoded tabular query column using a first transformer to obtain a first vector and the encoded table headers using a second transformer to obtain a second vector;
concatenating the first vector and the second vector to obtain a combined vector;
processing the combined vector through a linear layer and a Gaussian Error Linear Unit layer to obtain a final query vector;
processing the encoded target types through a third transformer to obtain a third vector; and
calculating a score for the target tabular query column as a dot product between the final query vector and the third vector to model interactions between the target tabular query column of the target table and the target types and provide a column-type annotation.
16. The computer program product of claim 15 , wherein encoding the target tabular query column, the table headers, and the target types independently to generate the permutation invariant representations of the type data and the tabular data associated with the target table further comprises:
linearizing the tabular data;
automatically assigning an increasing sequence of position identifiers to each non-delimiting tokenized cell in the linearized tabular data until a separator token is reached; and
in response to reaching the separator token, automatically assigning a monotonically increasing sequence of position identifiers for each non-delimiting tokenized cell positioned after the separator token, restarting from an integer corresponding to 1 greater than the position identifier assigned to the separator token.
17. The computer program product of claim 15 , wherein the auxiliary information further includes a glossary comprising a short description of the target label type.
18. The computer program product of claim 15 , wherein encoding the target tabular query column, the table headers, and the target types independently to generate the permutation invariant representations of the type data and the tabular data associated with the target table further comprises:
assigning the target types, associated glossaries, associated parent labels, and associated grand-parent labels different token identifications.
19. The computer program product of claim 18 , wherein the different token identifications are initialized to two pre-trained token type embeddings of a pre-trained transformer encoder model comprising two token type identifiers.
20. The computer program product of claim 15 , further comprising:
leveraging a library for large-scale vector-similarity searches to increase a speed of an inferencing step.