A Novel Method in Detecting Real-Word Errors Using Textual Entailment

Abstract

Spellchecking has become ubiquitous in modern life, running on everything from giant cloud computing clusters to pocket-sized phones. Modern spell checkers can detect and address simple typos, certain types of real-word errors, and even some grammatical mistakes. Despite the progress made on morphological and syntactical correctness, ensuring semantic cohesiveness has proven more elusive.

In this thesis, a novel method in detecting real-word errors by analyzing the entailment relationship between sentences is presented. It scores the degree of entailment between two sentences and tries to find alternative sentences that maximize the entailment value. This method should theoretically be able to remediate a class of real-word errors and semantic contradictions that existing spell checkers are not currently able to handle.

The proposed algorithm is compared against the current open-source spellchecking standards, Aspell and Hunspell, and a similar algorithm that tries to minimize perplexity. Multiple corpora, in English and Spanish, with artificially induced real-word errors from a generated homophone list were used. Real-world corpora were also examined with their naturally occurring errors.

When examining an English corpus with synthetic errors, the algorithm suggested the correct word at the correct location 83% percent of the time, showing that in ideal conditions, entailment can be a powerful tool to correct contextual spelling errors. The accuracy rate dropped to 53% in the Spanish corpus, still showing cross-language utility. However, on real-world data, the method was only 5% to 16% accurate. This suggests some combination of natural errors only rarely causing issues and deficiencies with the algorithm's implementation. More work integrating this algorithm with other error correcting methods, testing and training on larger datasets, and improving the implementation is needed.