Long-form Answer Extraction via Sentence Index Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing question answering systems are inadequate for long-form answer extraction, particularly in longer contexts, as they fail to effectively address the sparsity of answer sentences and perform poorly compared to factoid answers, necessitating a more efficient approach that leverages multiple techniques.
Innovation Solution
A system and method for long-form answer extraction using a combination of sentence index generation techniques, involving pre-processing, training generative sequence-to-sequence models, and post-processing to generate and expand sentence indices and spans, improving answer extraction in longer contexts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If span extraction techniques are used for long-form answer extraction, then the approach can handle general QA tasks, but the extraction results are poorer for long-form answers compared to factoid answers
Solution Approach 1:
The patent segments the answer extraction task into two distinct models: one for extracting sentence indices (identifying which sentences contain answer information) and another for extracting token spans (identifying the exact answer boundaries within those sentences). This segmentation allows each model to specialize in its specific function, improving overall performance for long-form answers while maintaining capability for factoid answers.
2Measurement precision
If sentence extraction techniques are used, then the discriminative approach can identify answer sentences, but it does not address the sparsity of answer sentences in the context well, especially in longer contexts
Solution Approach 1:
The patent introduces sentence indices as an intermediary representation between the input context and the final answer extraction. The first model generates sentence indices that point to relevant answer sentences, and the second model uses these indices to extract precise token spans. This intermediary approach helps manage the sparsity of answer sentences in long contexts by providing a structured intermediate representation that guides the extraction process.
3Device complexity
If a single extraction technique is used, then the system is simpler to implement, but it cannot leverage the advantages of multiple techniques for more efficient long-form extractive QA
Solution Approach 1:
The patent merges two different extraction techniques into a unified system: sentence-level extraction (identifying which sentences contain answers) and token-level extraction (identifying exact answer boundaries). By combining these techniques in a multi-model architecture, the system leverages the strengths of both approaches to improve long-form answer extraction efficiency while maintaining a structured and manageable system design.
Data Source
AI summary
This disclosure relates generally to long-form answer extraction and, more particularly, to long-form answer extraction based on combination of sentence index generation techniques. Existing answer extractions techniques have achieved significant progress for extractive short answers; however, less progress has been made for long form questions that require explanations. Further the state-of-art long-answer extractions techniques result in poorer long-form answers or not address sparsity which becomes an issue longer contexts. Additionally, pre-trained generative sequence-to-sequence models are gaining popularity for factoid answer extraction tasks. Hence the disclosure proposes a long-form answer extraction based on several steps including training a set of generative sequence-to-sequence models comprising a sentence indices generation model and a sentence index spans generation. The trained set of generative sequence-to-sequence models is further utilized for model long-form answer extraction based on a union of several sentence index generation techniques comprising a sentence indices and a sentence index spans.