Question Answering System Sentence Graph Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing topic-oriented question answering systems often return complete topics as answers, which can include irrelevant sentences, obfuscating the actual response to the question due to the breakdown of documents into multiple sub-documents based on formatting structures.
Innovation Solution
A method that breaks down ingested corpus into topics and further into constituent sentences, creating a graph structure with edge weights based on sentence similarity, allowing for the identification of strongly connected sentences to form concise answers by traversing the graph and updating edge weights based on the input question.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If complete topics are returned as answers from document breakdown, then comprehensive information coverage is achieved, but answer clarity and relevance deteriorate due to inclusion of irrelevant sentences
Solution Approach 1:
The patent segments the complete topic into individual sentences and evaluates each sentence's relevance to the question separately. By breaking down the topic at the sentence level rather than returning the entire topic, the system can selectively include only relevant sentences in the answer, thus maintaining information coverage while improving answer precision and clarity.
Solution Approach 2:
The patent applies different quality standards to different parts of the topic by evaluating each sentence individually for its relevance to the question. Instead of treating the entire topic uniformly, the system assigns local relevance scores to each sentence and selectively includes high-quality (relevant) sentences while excluding low-quality (irrelevant) ones, thereby improving overall answer relevance.
2Measurement precision
If sentence similarity calculation is performed for all sentences in a topic, then accurate answer filtering is achieved, but computational complexity increases
Solution Approach 1:
The patent performs similarity calculation only on sentences that are potentially relevant, rather than unnecessarily processing all sentences in the topic. By using threshold-based filtering and selective computation, the system achieves accurate relevance measurement while avoiding the excessive computational burden of evaluating every sentence in the corpus.
Data Source
AI summary
According to one exemplary embodiment, a method for generating an answer in a question answering system is provided. The method may include receiving a question. The method may also include identifying a candidate answer from a corpus. The method may then include determining a plurality of sentences based on the identified candidate answer. The method may further include calculating a similarity value for each sentence within the plurality of sentences based on comparing the plurality of sentences to the candidate answer and the received question. The method may also include identifying at least one sentence within the plurality of sentences with a calculated similarity value that exceeds a threshold value. The method may then include presenting the answer, whereby the answer comprises the plurality of sentences, the candidate answer, and metadata.