▍ Can ChatGPT Be Used for Patent Search Work?

“If users do not understand the principles of patent searching, and simply hand over the technical text to ChatGPT for the construction of search formulas without analyzing the technical contents, the search results will be problematic.”

Recently, ChatGPT (Generative Pre-trained Transformer), an artificial intelligence (AI) chatbot program developed by OpenAI, has become a popular topic, attracting much attention and discussion. Its applications in the fields of natural language processing and text analysis have been well documented and have aroused great interest. It can be used to generate various language models, such as natural language texts, dialogues, and question-answering. It is currently one of the most advanced and efficient technologies in the language field.

ChatGPT has a wide range of applications. In fields like medical, financial, legal, and media, ChatGPT can also be used to generate and analyze text data, thereby improving work efficiency and accuracy. Recently, the technology has even been used in the realm of intellectual property, with some having used it to draft patent applications.

Traditional patent search work requires a patent attorney to manually search and screen, while ChatGPT can automatically process text data, extract and analyze information. Imagine a scenario in which users only need to input technical content or patent publication numbers into ChatGPT to automatically complete patent document retrieval, without the need to specifically delegate patent agents for retrieval. This has also led to some patent search professionals feeling worried that their work may be replaced by ChatGPT.

But could ChatGPT really ever be used in patent search? What is its search capability? Let’s explore by conducting some tests.

Below is the object to be searched:

As shown in Figure 1 above, a cup is disclosed, comprising a cup body, a cup lid, and a spiral straw. The top of the straw serves as a water suction port, and the bottom of the straw serves as a water inlet. The feature is that the spiral straw is wound around the outer wall of the cup, which can provide a shockproof effect to the cup body.

In response to the above, we directly used ChatGPT for the search work. In the ChatGPT dialogue box, we entered the following statement and obtained the answer from ChatGPT:

As seen in Figure 2, ChatGPT almost immediately completed the search work and provided the publication number, patent name, and reason for similarity. Based on the content output by ChatGPT, it seems that it has found patent literature that is very similar to this application. Does this prove that ChatGPT can easily complete patent search work?

After locating the patents provided by ChatGPT in response to our question, we found that they were completely irrelevant to what we were searching for. So why did ChatGPT give such unreliable answers?

In fact, ChatGPT is a semantic response model, which means it does not have the ability to search the patent database in real time. In terms of patent search work, the answers given may have no authenticity.

Does this mean that ChatGPT can’t be used for patent search? Not exactly. Theoretically speaking, ChatGPT is good at semantic recognition and extraction, so we can extract key information from the text and automatically construct a search formula. Users can then directly search the database using this formula, which could replace the work of patent search professionals, at least to some extent. Based on this idea, we will further operate and let ChatGPT extract features from technical texts and automatically construct a search formula.

As we can see, ChatGPT extracted features such as cup, spiral straw, winding, and shock resistance from the technical text based on the technical effect to construct the search formula, which seems to be relatively accurate, but this is based on the clear description of the beneficial effect. The technical effect we input is inferred from the technical features, which means that we have done some processing work on the technical effect. In most cases, the technical effect copied from the invention disclosure form is relatively general, which can cause problems in ChatGPT’s recognition. For example:

As seen above, when the technical effect is not described in combination with the features, ChatGPT cannot accurately extract the search elements. So, if we don’t even provide the beneficial effects, can ChatGPT extract accurate search elements?

As we can see, if the search formula is not limited by the beneficial effects, all the keywords from the technical text provided will be used to construct the search formula, and the search scope is too small, resulting in the failure to search for patents. Comparing the search formulas provided by ChatGPT in Figures 4 to Figure 6, we believe that the search formula in Figure 4 is more in line with our expectation. We used the “title + abstract + claims” field to verify the search formula given in Figure 4, and obtained the following results:

From Figure 7, it can be seen that there are only nine search results, and most of them are irrelevant to what we are actually looking for. From the above, we could tell that ChatGPT is currently not very competent in extracting keywords and constructing search formulas. In other words, if users do not understand the principles of patent searching, and simply hand over the technical text to ChatGPT for the construction of search formulas without analyzing the technical contents, the search results will be problematic, which will lead to incorrect conclusions.

Well, it’s clear that ChatGPT can’t finish the work independently. How about working together with a patent attorney? We conducted another test, and this time, we took full advantages of ChatGPT’s strength, and had it extract keywords from the technical text first:

As can be seen from Figure 8, ChatGPT’s keyword extraction from the technical text is very accurate, and the word segmentation is also reasonable. In this way, a patent attorney can directly select the search elements they need from the keyword results and let ChatGPT construct search formulas (that is, the patent attorney participates in the selection of basic search elements). For example:

From the results, it can be seen that ChatGPT’s ability to expand keywords is relatively excellent, and the generated search formulas can be directly copied to the search database for searching, without manually entering parentheses and logical operators. We validated the search formulas and obtained the following results:

ery surprisingly, the search results this time were relatively satisfactory, and the search results were all target patents. In other words, by analyzing the technical solution, determining the exact search elements, and then having ChatGPT perform keyword expansion and search formula construction, the patent attorney can improve the efficiency of the patent search work to a certain extent, especially during the preliminary examination. Upon finishing reading the technical solution, the patent attorney can get the preliminary examination results of the target patent faster with the assistance of ChatGPT, and can easily adjust the search formulas (such as adding or deleting keywords, replacing a search element with a classification number, adjusting search fields, etc.).

We embarked upon one more test with ChatGPT to see if it could find classification numbers automatically.

As can be seen, ChatGPT can also quickly provide classification numbers, but it should be noted that whenever ChatGPT outputs a numerical or code answer, it must be verified. As expected, the classification numbers are inaccurate.

From Figure 12, it can be seen that the classification number provided by ChatGPT is for the drinking vessel on the dining table, not for the traditional water cup. Therefore, the accuracy of using ChatGPT for classification number extraction is low. It is better and more efficient to classify the search results in Figure 10 to obtain the classification number of the water cup.

In summary, ChatGPT can’t automatically complete patent search tasks. Although ChatGPT can extract keywords well, it can’t determine which keywords are the basic retrieval elements. The grasp of basic search elements is the key to a patent attorney’s ability to complete search tasks well and is the core value of the patent attorney. This is an area where ChatGPT can’t replace patent attorneys. Therefore, the current ChatGPT will not cause unemployment among patent search professionals.

On the other hand, can ChatGPT be used as a search tool for patent attorneys to improve search efficiency? I would say that it is limited.

On the one hand, the battlefield for the patent search work of patent attorneys is in the patent database. Most of the time and effort spent on search is in adjusting search formulas and browsing the patent literature searched, and ChatGPT only provides a quick way to enter the “battlefield”. Compared to humans, ChatGPT has better keyword expansion capabilities, but such feature is also available with the general patent databases, which also perform well. In this sense, the benefit of using ChatGPT for patent search work is limited to keyword expansions, which can’t substantively improve quality and efficiency.

Of course, with the iterative update of ChatGPT, accurate identification of basic search elements may be achieved in the future, or there may be better performance in determining the similarity of patent documents. We will also pay close attention to the technical progress and update readers with related information and usage tips.

Note: the figures included are machine translations of Chinese language questions and answers.