Saturday 05 April 2025
Researchers have developed a new way to search for documents that contain information about the causes of an event, rather than just searching for related topics. This approach could be useful in fields such as journalism and law enforcement, where understanding the underlying causes of events is crucial.
The traditional method of searching for documents involves matching keywords from the query with words in the document. However, this can lead to irrelevant results if the keywords are not specific enough or if the document contains related but not causally relevant information.
To address this issue, researchers have developed a new approach that uses a combination of lexical and semantic indexing techniques. Lexical indexing involves creating an index of words from the query and matching them with corresponding words in the documents. Semantic indexing, on the other hand, involves creating a vector representation of each document based on its meaning, rather than just its individual words.
The researchers used a dataset of news articles to train their model. The dataset consisted of 303,291 articles from the Telegraph India newspaper, along with labels indicating which articles were relevant to specific query events. The team then used this data to train a sentence embedding model that could be used to represent each document as a vector.
The researchers then developed a semantic search pipeline that combined results from multiple query strategies across both lexical and semantic indexes. The pipeline first embedded the query text using the sentence embedding model, then retrieved the most relevant documents based on cosine similarity scores between the query embedding and document embeddings. It also used a lexical index optimized for Okapi BM25 scoring to retrieve additional relevant documents.
The team evaluated their approach using two metrics: Mean Average Precision (MAP) and Precision at 5 (P@5). MAP measures the average precision of the top-ranked results, while P@5 measures the proportion of correct results in the top 5 ranked documents. The researchers found that their semantic search pipeline outperformed all the baseline methods and was the top method on the CAIR 2021 leader board.
This approach could have important implications for fields such as journalism and law enforcement, where understanding the underlying causes of events is crucial. For example, a journalist searching for documents related to a scandal might use this approach to identify relevant documents that contain information about the causes of the scandal, rather than just documents that mention related topics.
The researchers are now exploring ways to further improve their approach, including developing better models of causality and retrieving results using the query title only. This could enable even more accurate searches for causally relevant documents.
Cite this article: “Revolutionizing Causal Search: A Semantic Pipeline for Identifying Likely Causes of Query Events”, The Science Archive, 2025.
Causal Search, Document Retrieval, Semantic Indexing, Lexical Indexing, Sentence Embedding Model, Cosine Similarity, Okapi Bm25 Scoring, Mean Average Precision, Precision At 5, Information Extraction, Event Analysis
