Non-factoid question answering across tasks and domains
Inventors
Yu, Wenhao • Wu, Lingfei • Deng, Yu • Zeng, Qingkai • Mahindru, Ruchi • Guven Kaya, Sinem • Jiang, Meng
Assignees
International Business Machines Corp • University of Notre Dame
Publication Number
US-11797611-B2
Publication Date
2023-10-24
Expiration Date
2041-07-07
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
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.
Core Innovation
The invention presents a framework for non-factoid question answering across tasks and domains. It employs a multi-task joint learning model that is first trained in a general domain and then initialized and tuned in a specific target domain. This model handles both reading comprehension and document retrieval tasks in parallel. By encoding question/candidate document pairs and producing both a reading score and a matching score for each pair, the system enables both local identification of answer snippets and global semantic alignment between questions and candidate documents.
The problem addressed by the invention is the limitation of current non-factoid question answering models, particularly their reliance on a two-step process involving document retrieval and answer snippet identification. This can result in relevant documents being overlooked and lowers the accuracy for non-factoid answers that usually require actionable solutions or explanations. Additionally, challenges arise due to the small size of domain-specific datasets, making it difficult for existing frameworks to semantically associate data with documents.
The core solution involves dynamic optimization within the joint learning model, where the relative difficulty of the reading comprehension and document retrieval tasks is assessed and task weights are adjusted accordingly. The system's decision process, which includes task weighting, ranking, and updating based on the calculated joint learning loss, allows it to focus on the more challenging task to improve performance. This approach is implemented in a computer-implemented method, a computer system, and a computer program product as described in the various embodiments.
Claims Coverage
The claims cover multiple inventive features focused on the training and operation of a multi-task joint learning model for non-factoid question answering across tasks and domains.
Multi-task joint learning model for non-factoid question answering
A framework that initializes, tunes, and operates a multi-task joint learning model in a specific target domain for answering non-factoid questions. The model is first trained in a general domain and later adapted to a target domain by processing question/candidate document pairs, generating reading and matching scores, and outputting answer snippets.
Parallel computation of reading and matching scores
Reading scores (identifying answer snippets within documents) and matching scores (assessing semantic alignment between a question and document) are computed for each question/candidate document pair. These scores are generated in parallel and are used to evaluate and rank candidate answer snippets.
Dynamic task weighting based on task difficulty
The multi-task learning process includes determining which task—reading comprehension or document retrieval—is more difficult for the joint model to learn. Task weights are dynamically adjusted by the processor based on this determination during training, allowing the system to concentrate on the more challenging task.
Tokenization and vector encoding for each question/candidate document pair
Each question/candidate document pair is tokenized so that every word receives a token, and each token is encoded into a vector representation. This encoding facilitates the parallel processing of reading and matching tasks.
Joint prediction and ranking of answer snippets
A joint prediction model combines the reading and matching scores to rank identified answer snippets from multiple candidate documents. The highest-ranked answer snippet is selected as the output for a given question.
The claims define a comprehensive system that uses a multi-task joint learning approach, parallel task execution, dynamic task weighting, thorough data encoding, and a joint prediction model to advance non-factoid question answering across tasks and domains.
Stated Advantages
Provides improved non-factoid answer questioning abilities by performing both document retrieval and reading comprehension in parallel, reducing the risk of missing answers.
Enables global and local analysis for each question and candidate document pair, allowing both semantic document alignment and detailed answer snippet identification.
Supports fast and efficient training of non-factoid question answering frameworks in specific domains, addressing the challenge of limited domain-specific data.
Optimizes model performance by dynamically focusing training on the more difficult of the two tasks (reading comprehension or document retrieval), improving accuracy and adaptability.
Documented Applications
Non-factoid question answering for technical domains, such as information technology, automotive repair, and computer hardware troubleshooting.
Question answering over domain-specific document repositories containing technical documents, reports, research articles, blog posts, and similar sources.
Interested in licensing this patent?