Techniques for classifying data using large language models
A system and method for classification. A method includes identifying candidate entities among text data by applying at least one entity identification rule to the text data. Inputs are constructed based on the identified candidate entities, where each input includes a first portion of text indicating a candidate entity and at least one second portion of text and where the at least one second portion of text of each input is adjacent to the first portion of text of the input. Multiple language models are applied to the inputs, where each language model is trained to identify a respective set of entities and where outputs of the language models include at least one portion of entity-indicating text for each input. Based on the outputs of the language models, at least one named entity in the text data is determined.
1 . A method for classification, comprising:
identifying a plurality of candidate entities among text data by applying at least one entity identification rule to the text data;
constructing a plurality of inputs based on the identified plurality of candidate entities, wherein each input includes a first portion of text indicating a candidate entity from among the plurality of candidate entities and at least one second portion of text, wherein the at least one second portion of text of each input is adjacent to the first portion of text of the input;
applying a plurality of language models to the plurality of inputs, wherein each language model is trained to identify a respective set of at least one entity, wherein outputs of the plurality of language models include at least one portion of entity-indicating text for each input; and
determining, based on the outputs of the plurality of language models, at least one named entity in the text data.
2 . The method of claim 1 , wherein applying the at least one entity identification rule further comprises:
identifying at least one entity-indicating term within the text data by applying a truth table to the text data, wherein the truth table includes a plurality of columns representing a set of factors and a column including a plurality of score outputs, wherein the plurality of candidate entities are identified based on the identified at least one entity-indicating term.
3 . The method of claim 2 , further comprising:
marking the first portion of text indicating the candidate entity for each input based on results of applying the truth table to the text data, wherein the plurality of inputs are constructed based further on the marked first portion of text for each input.
4 . The method of claim 1 , further comprising:
matching, for each input, between the candidate entity of the input and at least one classification output by the plurality of language models based on the input in order to determine a distance between the candidate entity and each of the at least one classification for each input, wherein the at least one named entity is determined based further on the distance determined for each input.
5 . The method of claim 4 , wherein the matching includes comparing each distance to a threshold.
6 . The method of claim 3 , further comprising:
training the plurality of language models using a plurality of sets of training data, wherein the set of training data used to train each language model includes a plurality of training inputs and a plurality of corresponding training task outputs of the at least one entity of each language model, wherein each training input includes a first portion of training text indicating a training candidate entity and at least one second portion of training text which is adjacent to the first portion of training text.
7 . The method of claim 6 , further comprising:
adding a plurality of task prefixes to the plurality of training inputs, wherein the plurality of language models is trained using the plurality of training inputs with the added plurality of task prefixes.
8 . The method of claim 6 , wherein the plurality of language models is a plurality of first language models, further comprising:
labeling the plurality of training inputs with the plurality of corresponding training task outputs, wherein labeling the plurality of training inputs further includes applying at least one second language model to the plurality of training inputs.
9 . The method of claim 8 , wherein the at least one second language model is a plurality of second language models, further comprising:
applying the plurality of second language models in a voting process, wherein outputs receiving a majority of votes from the plurality of second language models are used as the plurality of corresponding training task outputs for labeling the plurality of training inputs.
10 . The method of claim 1 , wherein each language model is trained using a respective set of training text inputs, wherein each language model is configured to only generate text predictions from among the respective set of training text inputs used to train the language model.
11 . The method of claim 1 , wherein each language model is fine-tuned only for entity extraction.
12 . A non-transitory computer readable medium having stored thereon instructions for causing a processing circuitry to execute a process, the process comprising:
identifying a plurality of candidate entities among text data by applying at least one entity identification rule to the text data;
constructing a plurality of inputs based on the identified plurality of candidate entities, wherein each input includes a first portion of text indicating a candidate entity from among the plurality of candidate entities and at least one second portion of text, wherein the at least one second portion of text of each input is adjacent to the first portion of text of the input;
applying a plurality of language models to the plurality of inputs, wherein each language model is trained to identify a respective set of at least one entity, wherein outputs of the plurality of language models include at least one portion of entity-indicating text for each input; and
determining, based on the outputs of the plurality of language models, at least one named entity in the text data.
13 . A system for classification, comprising:
a processing circuitry; and
a memory, the memory containing instructions that, when executed by the processing circuitry, configure the system to:
identify a plurality of candidate entities among text data by applying at least one entity identification rule to the text data;
construct a plurality of inputs based on the identified plurality of candidate entities, wherein each input includes a first portion of text indicating a candidate entity from among the plurality of candidate entities and at least one second portion of text, wherein the at least one second portion of text of each input is adjacent to the first portion of text of the input;
apply a plurality of language models to the plurality of inputs, wherein each language model is trained to identify a respective set of at least one entity, wherein outputs of the plurality of language models include at least one portion of entity-indicating text for each input; and
determining, based on the outputs of the plurality of language models, at least one named entity in the text data.
14 . The system of claim 13 , wherein the system is further configured to:
identify at least one entity-indicating term within the text data by applying a truth table to the text data, wherein the truth table includes a plurality of columns representing a set of factors and a column including a plurality of score outputs, wherein the plurality of candidate entities are identified based on the identified at least one entity-indicating term.
15 . The system of claim 14 , wherein the system is further configured to:
mark the first portion of text indicating the candidate entity for each input based on results of applying the truth table to the text data, wherein the plurality of inputs are constructed based further on the marked first portion of text for each input.
16 . The system of claim 13 , wherein the system is further configured to:
match, for each input, between the candidate entity of the input and at least one classification output by the plurality of language models based on the input in order to determine a distance between the candidate entity and each of the at least one classification for each input, wherein the at least one named entity is determined based further on the distance determined for each input.
17 . The system of claim 16 , wherein the matching includes comparing each distance to a threshold.
18 . The system of claim 15 , wherein the system is further configured to:
train the plurality of language models using a plurality of sets of training data, wherein the set of training data used to train each language model includes a plurality of training inputs and a plurality of corresponding training task outputs of the at least one entity of each language model, wherein each training input includes a first portion of training text indicating a training candidate entity and at least one second portion of training text which is adjacent to the first portion of training text.
19 . The system of claim 18 , wherein the system is further configured to:
add a plurality of task prefixes to the plurality of training inputs, wherein the plurality of language models is trained using the plurality of training inputs with the added plurality of task prefixes.
20 . The system of claim 18 , wherein the plurality of language models is a plurality of first language models, wherein the system is further configured to:
label the plurality of training inputs with the plurality of corresponding training task outputs, wherein labeling the plurality of training inputs further includes applying at least one second language model to the plurality of training inputs.
21 . The system of claim 20 , wherein the at least one second language model is a plurality of second language models, wherein the system is further configured to:
apply the plurality of second language models in a voting process, wherein outputs receiving a majority of votes from the plurality of second language models are used as the plurality of corresponding training task outputs for labeling the plurality of training inputs.
22 . The system of claim 13 , wherein each language model is trained using a respective set of training text inputs, wherein each language model is configured to only generate text predictions from among the respective set of training text inputs used to train the language model.
23 . The system of claim 13 , wherein each language model is fine-tuned only for entity extraction.