Answer Generation Apparatus for Polarity Questions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine reading comprehension technologies struggle to answer questions with polarity, such as yes or no, as they are limited to extracting answers from text without generating new responses.
Innovation Solution
An answer generation apparatus that includes an encoding unit to transform text and questions into vector representations, an evidence extraction unit to identify relevant text spans, an evidence search unit to determine the evidence range, and a classification unit to classify the answer type, enabling accurate generation of polarized answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If extractive-type task is used to answer questions, then the answer can be extracted directly from text, but the answer cannot be output in a format not written in the text (such as polarity answers yes/no)
Solution Approach 1:
The answer generation process is segmented into multiple independent modules: encoding unit for text representation, evidence extraction unit for identifying relevant spans, evidence search unit for locating evidence ranges, and classification unit for determining answer types. This segmentation allows each module to specialize in one function while collectively enabling both direct extraction and formatted output capabilities.
Solution Approach 2:
The system is designed with multi-functionality to handle both extractive tasks (extracting answers directly from text) and generative tasks (outputting formatted answers like polarity). The classification unit specifically adds the capability to generate polarity answers by classifying evidence spans as supporting or contradicting the question, thus achieving universality across different answer types.
2Productivity
If the machine only pays attention to relevant text parts, then processing efficiency is improved, but the machine cannot classify the answer to output in required formats
Solution Approach 1:
The evidence extraction unit and evidence search unit act as intermediaries between the encoding unit and classification unit. These intermediary components process the encoded text to identify and locate relevant evidence spans, then pass this structured evidence information to the classification unit. This intermediary processing maintains efficiency by focusing only on relevant text while enabling the final classification step to generate properly formatted answers.
3Device complexity
If a simple extraction model is used, then the model complexity is reduced, but the accuracy of identifying evidence spans and ranges decreases
Solution Approach 1:
The evidence identification process is segmented into two distinct stages: evidence extraction (identifying candidate spans) and evidence search (locating precise ranges). This segmentation allows each stage to use appropriately complex models for its specific task, balancing overall system complexity with identification accuracy. The extraction stage filters candidates efficiently, while the search stage refines precision on identified evidence ranges.
Data Source
AI summary
An encoding unit transforms a piece of text divided into a plurality of spans that are subdivided units of the piece of text and a question which have been input into a vector representation sequence representing a meaning of a span and the question based on the piece of text and the question which have been input using a pre-trained encoding model for transforming input text into a vector representation sequence representing a meaning of the input text. For each of the spans, an evidence extraction unit estimates an evidence score indicating the degree to which the span is suitable as the evidence for extracting the answer using a pre-trained extraction model for calculating the evidence score based on the vector representation sequence.