Investigating the Impacts of Misspellings in Patent Search by Combining Natural Language Tools and Rule-Based Approaches

Among all sources of technical information, patent information is one of the richest and most comprehensive. Knowing how to search in this mass of documents is becoming increasingly crucial. However, many users have limited knowledge of patents and search strategies, so they must use intuitive, often approximate approaches that can lead to highly inaccurate searches and be time-consuming. To address this problem, there are tools that help expand queries to increase recall so as not to miss good documents, however, it remains an open problem dealing with misspellings-based strategies. Typically, the problem of the presence of misspellings in patent text is underestimated even by experts in the field, and there is no specific functionality to handle it in the tools available, both free and paid. The goal of the article is to raise awareness about the difficulties in making a proper patent strategy that also takes into account the possible presence of misspellings. It is important to know where we expect to find them and how much these may affect the final result. In particular, it is chosen to divide misspellings into categories, distinguishing between misspellings associated with a generic keyword or multiword from misspellings in acronyms, chemical formulas, names of applicants, inventors, or names of specific formulas or theorems. At least one example case is given for each category, showing when and how it may affect the result. Finally, an integrated approach combining word and contextual embedding models based on deep learning with a rule-based algorithm based on wild cards and truncation operators is suggested for correcting the query, automatically suggesting the most consistent misspellings, thus achieving a more accurate and reliable result