Low-resource Sentence Compression via Semi-Markov CRF and Linguistic Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional sentence compression systems require large amounts of data and are computationally expensive, making them inefficient for devices with limited resources and prone to poor grammatical accuracy due to their topic specificity and reliance on extensive training data.
Innovation Solution
A low-resource, structured, language-aware neural network-based sentence compression system that predicts the inclusion of words and phrases in a compressed sentence using a Semi-Markov Conditional Random Field algorithm, with a language model for grammatical accuracy scoring, allowing for efficient compression with limited training data and resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional sentence compression algorithms are used, then text can be compressed to fit within allotted space, but large amounts of training data are required which are infeasible to collect in certain instances
Solution Approach 1:
The patent changes the fundamental parameters of the compression system by switching from neural network-based approaches requiring extensive training data to rule-based and statistical methods that operate with minimal or no training data, while maintaining compression quality through linguistic rules and probability models
Solution Approach 2:
The patent employs lightweight, computationally inexpensive compression algorithms that do not require expensive large-scale training datasets, using instead simple statistical models and linguistic rules that can be applied directly without extensive pre-training
2Manufacturing precision
If conventional sentence compression systems are used, then text compression can be achieved, but the systems are computationally expensive and not available for use in devices with limited resources
Solution Approach 1:
The patent replaces computationally intensive neural network mechanisms with lighter statistical and rule-based mechanisms, substituting heavy deep learning computations with efficient probability calculations and linguistic rule applications that consume fewer computational resources
Solution Approach 2:
The patent uses lightweight algorithms that require minimal computational power and memory, making them suitable for deployment on resource-constrained devices like mobile phones and tablets without requiring expensive GPU or TPU hardware
3Area of stationary object
If simple truncation of long sentences is used, then space constraints are satisfied, but context is mischaracterized completely or gibberish is generated
Solution Approach 1:
The patent extracts and preserves the most important semantic elements and key information from long sentences using statistical importance scoring and linguistic rule-based selection, rather than simply cutting off text, ensuring that the compressed sentence retains contextual accuracy and meaningful content
Solution Approach 2:
The patent changes the compression approach from fixed-length truncation to intelligent selection of sentence components based on statistical importance and linguistic structure, dynamically determining which words and phrases to retain based on their contribution to sentence meaning
4Device complexity
If systems requiring less structure and training data are used, then resource requirements are reduced, but grammatical accuracy deteriorates
Solution Approach 1:
The patent introduces linguistic rules and statistical probability models as intermediary mechanisms that bridge the gap between simple low-resource algorithms and high grammatical accuracy, using pre-defined linguistic knowledge and statistical patterns to ensure grammatically correct compression without requiring extensive training data
Data Source
AI summary
Embodiments of the present invention provide systems, methods, and computer storage media for sentence compression in which a provided sentence is compressed to fit within an allotted space. Portions of the input sentence are copied to generate the compressed sentence. Upon receipt of a sentence, top candidate compressed sentences may be determined based on probabilities of segments of the input sentence to be included in a potential compressed sentence. The top candidate compressed sentences are re-ranked based on grammatical accuracy scores for each of the candidate compressed sentences using a language model trained using linguistic features of words and/or phrases. The highest scoring candidate compressed sentence may be presented to the user.