Efficient Candidate Discovery for Chinese Domain Lexicon Construction via Recursive Lexical Combination
This paper proposes Recursive Lexical Combination (RLC), a corpus-driven method that efficiently constructs Chinese domain lexicons by recursively expanding lexical units to significantly reduce candidate search space while retaining high recall of emerging terminology, as demonstrated through validation on national-level data asset policy texts.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast landscape of modern research, scholars increasingly turn to text itself as a primary source of data. Whether analyzing corporate reports, government policies, or news articles, the goal is often to extract specific meanings hidden within unstructured sentences. To do this effectively, researchers rely on specialized dictionaries, known as domain lexicons, which act as a filter to identify the precise words and phrases that matter for a particular field. However, language is fluid, especially in rapidly evolving areas like data economics. New concepts emerge constantly, often appearing as combinations of words that standard dictionaries have not yet recorded. The challenge lies in finding these new terms without drowning in a sea of irrelevant word combinations. If a computer tries to list every possible way to join words together, it creates an overwhelming number of candidates, most of which are nonsense. If it relies too heavily on existing lists, it misses the new ideas entirely.
A team of researchers from China University of Geosciences (Beijing) and the University of International Business and Economics has proposed a new way to navigate this problem. They focused on Chinese policy texts regarding data assets, a field where new terminology is appearing daily. Instead of guessing which words might be important or trying to list every possible combination, they developed a method called Recursive Lexical Combination. This approach treats the text as a sequence of building blocks. It starts with short, common word pairings and asks a simple question: does this specific combination appear frequently enough in the text to be worth extending? If the answer is yes, the method adds the next word to the right and checks again. If the combination becomes rare or disappears, the process stops immediately for that path. This acts as a filter, pruning away unlikely candidates before they can grow into long, confusing strings.
The researchers tested this method on a collection of national-level policy documents. They compared their new approach against two traditional strategies: one that simply matches text against a fixed list of known terms, and another that blindly generates every possible contiguous word combination. The results showed that the new method successfully trimmed the list of potential candidates by more than half, reducing the number of types from over 35,000 to roughly 15,000. Despite this significant reduction, the method still captured about 80 percent of the true, important terms that existed in the test documents. Crucially, it found nearly 80 percent of the new terms that were missing from the original fixed list, proving that it could discover emerging concepts without needing to know them in advance.
The study highlights that the efficiency of this pruning depends on the length of the terms. For short, two-word combinations, the method kept every single possibility, ensuring nothing was missed. As the terms grew longer, the method became more selective, cutting away more candidates. While this meant that some very long, rare terms were not found, the trade-off was a much smaller, more manageable list for human experts to review. After the computer generated this refined list, human experts reviewed the remaining candidates to confirm their meaning and relevance. This final step boosted the accuracy of the results significantly, turning a rough list of possibilities into a reliable dictionary.
The findings suggest that for fields with rapidly changing language, the best approach is not to rely solely on existing knowledge or to try to process everything at once. Instead, a hybrid workflow works best: use the text itself to guide the discovery of new combinations, stop the search early when the evidence is weak, and then let human judgment finalize the list. This process allows researchers to build dictionaries that are both comprehensive enough to catch new ideas and focused enough to be practical. The study confirms that by separating the task of finding candidates from the task of validating their meaning, it is possible to construct domain lexicons that are both efficient and adaptable to the evolving nature of specialized language.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.