Document Summarization Model Training with Named Entity Loss Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document summarization models lack accuracy in preserving factual consistency, particularly with regards to proper nouns and numerical information, leading to potentially misleading summaries.
Innovation Solution
The proposed solution involves a training apparatus for document summarization models that utilizes named entity recognition to differentiate between named and non-named entities, calculates weighted losses for each type of entity, and employs a masking process to concentrate information on important tokens, thereby improving factual consistency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If document summarization models are trained to rapidly summarize input documents with natural sentence structures, then productivity is improved, but manufacturing precision deteriorates due to altered proper nouns and numerical information
Solution Approach 1:
The patent segments the summarization task by introducing a named entity recognition unit that identifies and separates named entities from other tokens. This segmentation allows the model to handle named entities differently during summarization, preserving their factual accuracy while maintaining overall summarization efficiency. The loss function is also segmented into separate components for named entities and non-named entities, enabling targeted optimization.
Solution Approach 2:
The patent applies local quality by assigning different treatment and weighting to different parts of the input text. Specifically, named entities receive special attention through dedicated loss calculation and weighted integration in the total loss function. This ensures that critical factual elements (proper nouns, numerical information) are preserved with higher precision while other parts of the text can be summarized more freely for productivity.
2Ease of operation
If the model prioritizes natural sentence structures in summarization, then ease of operation is improved, but reliability deteriorates due to misleading information
Solution Approach 1:
The patent changes the parameter weighting in the loss function to reflect the different importance of named entities versus other text elements. By introducing weighted loss calculation where named entity loss components receive higher weights, the model optimizes for both natural sentence structures and factual accuracy. This parameter adjustment ensures that named entities are preserved accurately while maintaining overall summarization quality.
3Device complexity
If uniform loss calculation is applied to all tokens, then device complexity is reduced, but measurement precision deteriorates regarding named entity accuracy
Solution Approach 1:
The patent segments the loss calculation into distinct components: one for named entities and another for non-named entities. This segmentation is implemented through separate loss function calculations that are then combined with appropriate weighting. While this increases computational steps slightly, it dramatically improves named entity preservation accuracy by treating these critical tokens differently from ordinary text tokens.
Solution Approach 2:
The patent applies local quality by creating specialized loss calculation pathways for named entities. Instead of a single uniform loss function, the system implements localized loss calculations that specifically target named entity tokens, ensuring they receive appropriate attention during training. This localized approach maintains reasonable system complexity while significantly improving measurement precision for named entities.
Data Source
AI summary
An apparatus for training a document summarization model includes a token generation unit, a named entity recognition unit, and a model training unit. The token generation unit generates document tokens and summarization tokens. The named entity recognition unit assigns named entity token status to a summarization token, recognized as a named entity through NER, and assigns non-named entity token status to the other tokens. The model training unit obtains feature vectors by inputting the plurality of document tokens into an encoder inside a document summarization model, obtains a first loss related to the named entity token, a second loss related to the other tokens, and a total loss using a weighted value by inputting the feature vectors, the summarization tokens, the named entity token, and the non-named entity token into a decoder inside the document summarization model, and trains the document summarization model on the basis of the total loss.