Non-factoid question answering across tasks and domains
An approach for a non-factoid question answering framework across tasks and domains may be provided. The approach may include training a multi-task joint learning model in a general domain. The approach may also include initializing the multi-task joint learning model in a specific target domain. The approach may include tuning the joint learning model in the target domain. The approach may include determining which task of the multiple tasks is more difficult for the multi-task joint learning model to learn. The approach may also include dynamically adjusting the weights of the multi-task joint learning model, allowing the model to concentrate on learning the more difficult learning task.
1. A computer-implemented method for training a machine learning framework for answering non-factoid questions across tasks and domains, the computer-implemented method comprising:
initializing, by the processor, a multi-task joint learning model in a specific target domain;
receiving, by the processor, an input sequence, wherein the input sequence is a non-factoid question in the specific target domain;
generating, by the processor, a plurality of question/candidate document pairs, wherein the question/candidate document pair is associated with a single document;
assigning, by the processor, one or more snippet scores for each of the question/candidate document pairs, wherein a snippet score corresponds to an answer snippet;
generating, by the processor, a reading score for each question/candidate document pair, based on the one or more assigned snippet scores;
generating, by the processor, a matching score for the document associated with each of the question/candidate document pairs, wherein the matching score is assigned to the candidate document of the question/candidate document pair; and
outputting, by the processor, an output sequence, based at least in part on the reading score and the matching score, wherein an output sequence is an answer snippet;
tuning, by the processor, a multi-task joint learning model in the target domain, based on the output sequence, the reading score and the matching score.
2. The computer-implemented method of claim 1 , wherein the multi-task joint learning model comprises the tasks of reading comprehension and document matching.
3. The computer-implemented method of claim 1 , wherein training further comprises:
determining, by the processor, the more difficult task between the multiple tasks of the multi-task joint learning model; and
adjusting, by the processor, the weights of the tasks in the multi-task joint learning model, dynamically, based on the determining.
4. The computer-implemented method of claim 1 , wherein the multi-task joint learning model is comprised of a language model.
5. The computer-implemented method of claim 1 , wherein the multi-task joint learning model is comprised of multiple multilayer perceptron classifiers.
6. The computer-implemented method of claim 1 , wherein the multi-task joint learning model is comprised of a joint prediction model, wherein during inference, the multi-task joint learning model ranks a plurality of identified answer snippets from a plurality of candidate documents.
7. The computer system of claim 1 , wherein the multi-task joint learning model is comprised of a joint prediction model, wherein during inference, the multi-task joint learning model ranks a plurality of identified answer snippets from a plurality of candidate documents.
8. The computer-implemented method of claim 1 , wherein generating the reading score comprises:
tokenizing, by the processor, each of the question/candidate document pairs, a token is generated for every word in each of the question/candidate document pairs; and
encoding, by the processor, a vector representation for each token in the question/candidate document pair.
9. A computer system for training a machine learning framework for answering non-factoid questions across tasks and domains, the system comprising:
a processor;
a readable storage media; and
computer program instructions to:
initialize a multi-task joint learning model in a specific target domain;
receive an input sequence, wherein the input sequence is a non-factoid question in the specific target domain;
generate a plurality of question/candidate document pairs, wherein the question/candidate document pair is associated with a single document;
assign one or more snippet scores for each of the question/candidate document pairs, wherein a snippet score corresponds to an answer snippet;
generate a reading score for each question/candidate document pair, based on the one or more assigned snippet scores;
generate a matching score for the document associated with each of the question/candidate document pairs, wherein the matching score is assigned to the candidate document of the question/candidate document pair; and
output an output sequence, based at least in part on the reading score and the matching score, wherein an output sequence is an answer snippet;
tune a multi-task joint learning model in the target domain, based on the output sequence, reading score and the matching score.
10. The computer system of claim 9 , wherein the multi-task joint learning model comprises the tasks of reading comprehension and document matching.
11. The computer system of claim 9 , wherein training further comprises:
determining, by the processor, the more difficult task between the multiple tasks of the multi-task joint learning model; and
adjusting, by the processor, the weights of the tasks in the multi-task joint learning model, dynamically, based on the determining.
12. The computer system of claim 9 , wherein the multi-task joint learning model is comprised of a language model.
13. The computer system of claim 9 , wherein the multi-task joint learning model is comprised of multiple multilayer perceptron classifiers.
14. The computer system of claim 9 , wherein generating the reading score comprises instructions to
tokenize each of the question/candidate document pairs, a token is generated for every word in each of the question/candidate document pairs; and
encode a vector representation for each token in the question/candidate document pair.
15. A computer program product for training a machine learning framework for answering non-factoid question across tasks and domains, the computer program product including one or more computer readable storage media and program instructions stored on the one or more computer readable storage media, the program instructions including instructions to:
initialize a multi-task joint learning model in a specific target domain;
receive an input sequence, wherein the input sequence is a non-factoid question in the specific target domain;
generate a plurality of question/candidate document pairs, wherein the question/candidate document pair is associated with a single document;
assign one or more snippet scores for each of the question/candidate document pairs, wherein a snippet score corresponds to an answer snippet;
generate a reading score for each question/candidate document pair, based on the one or more assigned snippet scores;
generate a matching score for the document associated with each of the question/candidate document pairs, wherein the matching score is assigned to the candidate document of the question/candidate document pair; and
output an output sequence, based at least in part on the reading score and the matching score, wherein an output sequence is an answer snippet;
tune a multi-task joint learning model in the target domain, based on the output sequence, reading score and the matching score.
16. The computer program product of claim 15 , wherein the general purpose machine learning framework is comprised of a language encoding model, a reading comprehension model, and a document retrieval model.
17. The computer program product of claim 15 , wherein the multi-task joint learning model comprises the tasks of reading comprehension and document matching.
18. The computer program product of claim 15 , wherein the multi-task joint learning model is comprised of a language model.
19. The computer program product of claim 15 , wherein the multi-task joint learning model is comprised of a joint prediction model, wherein during inference, the multi-task joint learning model ranks a plurality of identified answer snippets from a plurality of candidate documents.
20. The computer program product of claim 15 , wherein generating the reading score comprises instructions to:
tokenize each of the question/candidate document pairs, a token is generated for every word in each of the question/candidate document pairs; and
encode a vector representation for each token in the question/candidate document pair.