Short Summaries of Top Research Papers on AI Patent Search

Laborious, expensive, and exhaustive are a few adjectives that come to mind when describing the traditional ways of processing patent data, whether it is classifying patents, clustering documents, or highlighting paragraphs. However, bright minds like yourself are actively coming up with research to automate many such activities using artificial intelligence.

So, as a researcher, you always want to stay abreast of the latest developments in AI patent search. Just as the greatest ever ice hockey player, Wayne Gretzky, once said, “I skate to where the puck is going, not where it has been.”

In this piece, we will share concise summaries of recent research papers on AI patent search to assist you in your research endeavors. Let’s learn more about them.

PatentNet: Multi-label Classification of patent documents using deep learning-based language understanding

Roudsari et al. | 2022 | Springer

The researchers attempted to automate complicated, laborious, and expensive patent classification tasks such as multi-label classification using finely tuned pre-trained language models such as BERT, XLNet, RoBERTa, and ELECTRA. Next, the models underwent comparison against the baseline deep-learning approaches CNN, LSTM, BiLSTM, and CNN-BiLSTM. The research findings indicate that the pre-trained language models showed better multi-label patent classification performance, demonstrating their effectiveness in capturing patent document features better than the baseline models.

Patent Sentiment Analysis to Highlight Patent Paragraphs

Chikkamath et al. | 2021 | arXiv

Patent paragraph highlighting – a form of patent sentiment analysis and information retrieval – is a crucial stage in assessing patent documents. The authors of this paper have proposed a novel dataset to train Machine Learning algorithms to automate the patent paragraph highlighting process – cutting short a laborious and time-consuming task. The dataset to train the Machine Learning algorithms can automatically highlight patent paragraphs based on individual subject matter types.

Modular Development in Patent AI Space: A Case Study

Maan et al. | 2021 | CEUR-WS.org

This paper proposes researchers consider fitting their work into a common modular system of software components. This approach will enable faster prototyping, smoother upgrades, easier collaboration among research groups, and scale research and development in the patent-AI space. Artificial Intelligence can perform various patent-related tasks such as prior-art searching, technology landscaping, patent classification, etc. Such a system of modular components was created during the development of PQAI, an AI-based prior-art search platform.

AI-Assisted Patent Prior Art Searching - Feasibility Study

Rossi Setchi et al. | 2020 | Web | Wayback Machine

The UK patent office asked Cardiff University to evaluate the viability of current AI for patent prior art search. The authors checked several ML approaches to see if those could be used for tasks such as feature extraction, query expansion, document classification, document clustering, and topic modeling on patent data. The study showed that even though current approaches are not at a stage where they can fully automate every aspect of the prior-art search process, they can significantly assist examiners. Some parts of the process, for example, patent classification into CPCs, can be fully automated.

Optimizing Neural Networks for Patent Classification

Abdelgawad et al. | 2019 | PDF

In this paper, the authors compare recent approaches for text classification for patent categorization. It also investigates how much the AI models can be improved with different word embeddings and hyperparameter optimization. They found that CNN-based classification, custom FastText embeddings, and state-of-the-art optimizations outperform other approaches.

DeepPatent: Patent Classification with Convolutional Neural Networks and Word Embedding

Shaobao Li et al. | 2018 | Springer

A deep learning model based on CNN and word vectors is presented in this paper. The model does patent classification. It was trained on the title and abstract of the USPTO corpus. It showed high classification accuracy; unlike traditional approaches, it can practically be trained on vast amounts of data. An interesting finding of this paper is the relationship between the effect of input text size on the model and its accuracy. It was found that performance saturates at ~100 words.

Automating the Search for a Patent’s Prior Art with a Full Text Similarity Search

Helmers et al. | 2019 | arXiv

Most deep learning AI models that find patent similarity rely on a relatively small amount of text, the patent abstract. This paper’s authors went the opposite way and used full text of the patent for this task. They used tf-idf BOW, word2vec, and doc2vec to get patent vectors over full text, abstract, and claims of ~2,500 target patents and their citations from the A61 patent class. As a result, they found that nearly half of the citations suggested by the AI model come out to be accurate.

TechNet: Technology Semantic Network Based on Patent Data

Sarica et al. | 2019 | arXiv |  Github

In this paper, the authors created a technology semantic network, TechNet, which is trained on USPTO patent data. Technical terms were extracted and vectorized from the patent titles and abstract and their relationships were established in a vector space to form the technology semantic network. It represents an ‘engineering knowledge base’. TechNet was found to perform better than existing public semantic networks for knowledge retrieval and inference tasks in engineering.

Engineering Knowledge Graph for Keyword Discovery in Patent Search

Sarica et al. | 2019 | Wayback Machine

In this paper proposes a method to automate the keyword discovery process during patent search. Human searchers can read patent texts from prior searches to discover additional keywords that are not initially obvious and expand their search queries iteratively. In this approach, similar keywords from patents are returned for some keywords searched, and these returned keywords can be used again to search more keywords. This is done repetitively until no new keywords are found. It was found that this AI-based approach is faster and more accurate than a manual process.

Improving Chemical Named Entity Recognition in Patents with Contextualized Word Embeddings

Zenani Zhai et. al | 2019 | arXiv

This paper shows how chemical names can be found with high accuracy within patents. The authors used a concatenation of pre-trained word embeddings, CNN-based character-level word embeddings, and ELMO-based contextualized word embeddings as input to a neural network to boost performance of finding chemical compounds mentioned in patent text. This method greatly improves the identification of long chemical compound names and domain-specific and novel terms that were mostly missed by previous methods.

Summing it Up

Together we embark on a transformative journey of utilizing AI-assisted patent activities – classifications, sentiment analysis, extracting knowledge, inferring events, and more.

We sincerely hope this repository of research papers will tremendously assist you in your research endeavors to develop new ways of AI-based patent searches.

For further digging, you might want to check out this more exhaustive repository - Top AI-Based Research Papers on Prior Art Search.