Semantic Search

Semantic vs. Lexical Search: What is the Difference?

Historically, search engines have operated by matching the keywords in a user’s query to occurrences of those words in large collections of documents.

Formally referred to as lexical search, the approach is reliant on directly matching words in a user’s query to their appearances in documents.

Lexical search allows for minor variations in spelling or word formation through techniques—however, these techniques are limited in their ability to identify the equivalence of different words or phrases that are semantically related but lexically distinct.

The limitation of the inverted index approach used in lexical search lies in the inability to capture semantically equivalent words or phrases.

Semantic equivalence refers to words or sentences that differ in form but convey the same (or contextually similar) meaning.

For instance, “Operating Income” is interchangeable with the term “EBIT”; however, an inverted index data structure might not recognize the equivalence (and thus, neglect the semantically related terms).

One potential solution to the limitation involves creating an index of synonyms to identify similar versions of words. However, such an approach requires extensive fine-tuning (and prompting), which is generally only well-defined for individual words or phrases and does not account for semantically equivalent sentences—not to mention, the surrounding context and position must also be considered (e.g. “Adjusted EBITDA” vs. “Management Adjusted EBITDA).

To address these challenges, embeddings can be used to represent words and sentences in a manner that captures their semantic meaning, and these representations, or vectors, contain a list of real-valued numbers, where words or sentences with similar meanings are located close to each other in the vector space.

Semantic search, comprised of knowledge graphs and natural language processing (NLP) techniques, enable search engines to handle unstructured textual data and match documents to queries based on semantics rather than lexical overlap.

Natural language processing (NLP) helps the search system understand content on a deeper level, while machine learning (ML) refines the results through data-driven patterns (and “trial-and-error”).

How Does Semantic Search Work?

Semantic search operates by leveraging vector search, which enables the delivery and ranking of content based on contextual and intent relevance.

Vector search encodes details of searchable information into vectors—fields of related terms or items—and then compares these vectors to determine which are most similar. The process utilizes text embeddings to transform words into vectors, which are essentially lists of numbers.

By measuring the similarity between these vectors, the system identifies the response that is most similar to the vector corresponding to the query, and the response associated with the most similar vector becomes the output.

When searching for information in a large corpus of text, two primary approaches are employed:

• Lexical Search ➝ Keyword-Based (Exact-Match)
• Vector Semantic Search ➝ Context-Oriented

Keyword or lexical search relies on matching exact words or phrases in the query with those in the documents.

The lexical search approach is relatively simple and fast but has limitations, considering lexical search does not consider the context or meaning of words. The lack of semantic understanding can hinder the retrieval of truly relevant information or data points.

In contrast, vector semantic similarity search uses natural language processing (NLP) techniques to analyze the meanings of words and their relationships, wherein words are represented as vectors in a high-dimensional space, where the distance between vectors indicates their semantic similarity.

Vector semantic similarity search can handle more subtle relationships between words, including nuances.

In effect, more accurate and relevant search results are produced under semantic search, for the most part.

Once a user inputs a search query, the step-by-step process by which semantic search functions is as follows:

• Step 1 ➝ Analyze Intent and Context: The large language model (LLM) performs an analysis of the user’s query to comprehend the user’s intent and the contextual meaning of the query.
• Step 2 ➝ Extract Intent and Relationships: The semantic search system processes the query to identify the relationships among the terms and to determine the semantic meaning.
• Step 3 ➝ Return Intent and Relationships: The extracted intent and relational information are conveyed back to the LLM.
• Step 4 ➝ Retrieve Relevant Data: Utilizing the comprehended intent, the LLM retrieves data that are pertinent to the query.
• Step 5 ➝ Rank Data Based on Relevance: A ranking algorithm assesses the retrieved data from a vector database, ordering them according to their relevance to the query.
• Step 6 ➝ Generate Output: The LLM responds with the generated content or search results to the user, thereby concluding the semantic search process.

What are the Mechanics of Semantic Search?

To comprehend the underlying mechanisms of semantic search, the concept of semantic indexing and embedding must be understood.

Collectively, the two techniques—semantic indexing and embedding—facilitate efficient and relevant information retrieval based on similarity, resulting in more accurate results.

• Semantic Indexing ➝ Semantic indexing is a technique used in information retrieval to organize and categorize documents based on their meaning rather than words by themselves. The indexing process involves analyzing the content of each document and assigning it to a set of keywords or concepts that describe its main ideas.
• Embedding ➝ Embedding is the process of creating numerical representations as vectors in a high-dimensional space of words or documents that capture their meanings.

Unlike traditional search engines that are reliant on keyword matching, semantic search focuses on context and intent. The underlying meaning of queries can be deciphered, even if exact keywords present in the results are not included.

The shift from keyword-centric to meaning-centric searching significantly enhances the precision and relevance of search results.

Semantic search assists users in obtaining better results by allowing natural language questions, rather than requiring specific keywords.

Matching and ranking of content extend beyond standard lexical matching through the use of artificial intelligence (AI).

The intent of the user query can be captured by extending the query with context, in which the meaning of words can be part of that context.

Semantic search engines utilize several key technologies to achieve a deeper understanding of search queries and content:

How Does Semantic Search Improve Accuracy?

Semantic search enhances the accuracy of search results by comprehending the meaning and context of terms in the user’s query and the documents being searched.

Unlike traditional keyword matching, semantic search goes beyond surface-level associations to deliver more relevant and contextually appropriate results.

Semantic search engines utilize structured knowledge bases such as ontologies and knowledge graphs, which define entities, concepts, and their interrelationships.

By mapping query terms and document content to these knowledge bases, the search engine can understand the semantic connections between them.

The capability allows the identification of relevant results even if they do not contain the exact keywords used in the query.

The process involves analyzing the user’s query to discern their true intent and information need.

The query is parsed to extract key entities and understand its grammatical structure — these entities are then linked to corresponding entries in the knowledge graph.

The original query can be expanded with synonyms, related concepts, and contextual information identified from the knowledge graph, which aids in retrieving relevant documents that do not contain the verbatim query terms.

Instead of merely matching keywords, semantic search engines match the expanded query against a semantic index that encodes the meanings of documents.

Search results are then ranked by relevance using semantic similarity measures that quantify how closely the meanings of the query and the documents align.

Some semantic ranking signals that can be implemented include the semantic relatedness of query entities to document entities, the centrality, or importance of matched entities within the knowledge graph, and personalized signals based on the user’s search history and preferences.

By understanding the semantics of queries and documents, expanding queries with related concepts, and ranking based on semantic similarity, semantic search can surface more relevant results than traditional keyword-based search.

The combination of knowledge graphs and advanced natural language processing techniques (NLP) enables semantic search engines to better meet the user’s particular information needs.

Semantic search is facilitated by natural language processing (NLP) and machine learning (ML), which enhance search results by comprehending the user’s intent rather than merely matching keywords to documents.

NLP enables the search engine to understand concepts at a deeper level, while ML employs data and iterative patterns to refine the user experience.

Semantic search engines also utilize structured knowledge bases like ontologies and knowledge graphs that define entities, concepts, and their relationships. By mapping query terms and document content to these knowledge bases, the search engine can comprehend the semantic connections between them, allowing it to identify relevant results even if they do not contain the exact query keywords.

The original query can be expanded with synonyms, related concepts, and contextual information identified from the knowledge graph, which aids in retrieving relevant documents that do not contain the exact query terms.

^{Semantic Search Pinecone Vector Database (Source: Pinecone)}

What are the Benefits of Semantic Search?

Why Does Semantic Search Matter in Finance?

The finance industry manages vast amounts of complex and unstructured data as part of their daily workflow.

• Public Filings (10-K and 10-Q)
• Earnings Call Transcripts
• Equity Research Reports
• Financial News Articles
• Industry Research Reports

Semantic search offers significant benefits by enabling users to efficiently locate the information they require within these extensive datasets. By understanding the meaning and context of terms in both the user’s query and the documents being searched, semantic search improves the accuracy of search results, surpassing the limitations of simple keyword matching to deliver more relevant outcomes.

The more comprehensive and diverse the data sets and documents an LLM is trained on, the more refined and accurate the semantic search performance will be.

Semantic search is a methodology wherein the semantic meaning of words is utilized to retrieve pertinent content from document collections or data sets, which differs from keyword-based search, which retrieves documents by matching exact keywords.

Semantic search effectively retrieves content that shares the same meaning as a user’s query, even if different words are used.

The iterative process addresses the shortcomings of the inverted index approach, which, while effective at retrieving exact or similar variants of a given word, cannot capture semantically equivalent words or sentences that differ in form but share similar meanings.

Dense Retrieval vs. Reranking: What is the Difference?

There are two main approaches to semantic search, both helping improve RAG outputs.

• Dense Retrieval ➝ The first type of semantic search is known as dense retrieval, which relies on vector similarity using dense vector embeddings in a high-dimensional vector space. It contrasts with the traditional sparse vector representations used in keyword-based retrieval. Dense retrieval allows a user to write a specific query relating to a portion of the knowledge source, and have the system return a concise response with correlating embedding references.
• Reranking ➝ The second type of semantic search is known as reranking. Reranking requires a systematic approach to assign relevance scores to matches within the knowledge base. These scores are then used to change the order in which results are displayed to the user, to optimize result relevance.

Methods like TF-IDF can also be used to distinguish relevant from non-relevant words.

However, due to the ambiguity of language, synonyms, and other roadblocks, keyword search will sometimes fail to find the right response.

A vector search-enabled semantic search produces results by working at both ends of the query pipeline simultaneously.

Once a query is received, the search engine transforms the query into embeddings, which are numerical representations of data and related contexts, and stored in vectors.

The kNN algorithm, or k-nearest neighbor algorithm, then matches vectors of existing documents (a semantic search concerns text) to the query vectors. The semantic search then generates results and ranks them based on conceptual relevance.

How to Extract Structured Data from Unstructured Documents

Financial reports come in a wide variety of unstructured formats like PDFs, HTML, and text documents.

Semantic search engines use natural language processing (NLP) techniques to extract structured data from these sources:

By converting unstructured financial data into a structured format, the LLM can better understand the data and the relationships between different pieces of information, and therefore, provide more accurate and relevant results.

Semantic Search (NLP): Applications in Finance

The finance industry routinely handles vast amounts of complex, unstructured data. Implementing semantic search technologies offers significant benefits by enabling users to efficiently locate pertinent information within these extensive datasets. Key applications include:

Investment analysts and portfolio managers can leverage semantic search to gather relevant information from financial news, company filings, market reports, and other data sources more effectively.

By adopting semantic search, financial institutions can substantially enhance the efficiency and quality of their financial statement and valuation analysis for use-cases such as the following:

• Extract financial metrics and ratios across a wide array of companies and industries.
• Identify risk factors and forward-looking statements essential for assessing investment potential.
• Extract insights and discern trends from unstructured data, such as earnings call transcripts.
• Minimize the manual effort required to comb through financial filings and reports.

In closing, semantic search engines extract structured data, map it to standard taxonomies, integrate semantic and keyword techniques to understand the contextual meaning, and further improve via interative refinement, thereby facilitating more accurate and reliable research with improved efficiency.