semantic search

What is semantic search?

Semantic search is a data searching technique that uses natural language processing (NLP) and machine learning algorithms to improve the accuracy of search results by considering the searcher's intent and the contextual meaning of the terms used in their query. Semantic search is widely used in web search engines such as Google, but it also has applications in areas such as content management systems, internal corporate chatbots and e-commerce platforms.

Traditional keyword-based search methodologies, known as lexical search, focus only on finding exact matches for terms used in the searcher's query. While this technique is useful for surfacing direct matches, it cannot account for linguistic nuances such as homonyms, synonyms and context-dependent meanings.

Semantic search, in contrast, aims to identify the searcher's underlying objective and find contextually relevant results, even if they do not contain the exact words used in the original query. In other words, semantic search algorithms aim to understand what users actually mean, not just what they say.

To produce these results, semantic search algorithms draw on external sources such as knowledge graph databases, specialized lists of terms called ontologies and subject-specific collections of texts. In some cases, they also incorporate contextual information about the user, such as their location and search history.

How does semantic search work?

Semantic search algorithms have complex architectures that integrate several branches of machine learning, including NLP, question answering and knowledge graphs. While the entire process of returning search results might take only a fraction of a second using a search engine like Google, it involves multiple steps behind the scenes.

When a semantic search system receives a user's search query, it starts by using NLP to tokenize the query, breaking it down into smaller units such as words and phrases. The algorithm then marks each token as a particular part of speech, such as an adjective or verb, known as part-of-speech tagging, and analyzes their grammatical relationships, known as dependency parsing.

During this stage, the algorithm might also transform tokens into word embeddings, or numerical vector representations in which words with similar meanings are mapped close together in space. This step helps the algorithm comprehend semantic relationships among words, thereby further improving its understanding of context.

For example, imagine that a user Googles the phrase "tallest mountain in the United States." After decomposing that query into tokens tagged as certain parts of speech, the algorithm finds their interrelationships -- for example, the adjective tallest modifies the noun mountain. The algorithm also performs named entity recognition to categorize known named entities, such as people's names, locations, quantities and so on. In this case, the algorithm would recognize the term United States and categorize it as a known geographical entity.

After this initial query processing, the algorithm begins the semantic analysis stage. This involves steps such as deciding which definition fits best for words with multiple meanings, known as word-sense disambiguation; identifying ideas and themes, known as concept extraction; and broadening the search to include synonyms and related terms, known as query expansion.

To continue with the above example, the search algorithm might recognize the term mountain not just as a term, but as a concept associated with natural landscapes. Likewise, it might broaden the search to include North America as a related term to United States.

The system then uses semantic indexing to access presorted information about these terms. Web search engines like Google rely on indexed documents and data entries, ranking content based on relevance and authority. These engines prioritize content that is not only semantically relevant, such as a list of the heights of various mountains, but also the most authoritative -- for example, websites associated with government agencies, reputable universities and established news outlets.

In many cases, semantic search algorithms are also trained on examples of user queries and can continuously adapt based on new user data. For example, an algorithm might use information about which links a user clicked on and how much time they spent on the results page when returning results for that user's future queries.

In addition, changes that users make to their search terms after an initial query can serve as feedback about the results. If users frequently alter their language and try again after making a certain initial search, for instance, this could indicate dissatisfaction with the first page of results.

Knowledge graphs also play an important role in ensuring that algorithms can quickly return relevant information for search queries. For example, Google's proprietary Knowledge Graph, launched in 2012, contains billions of data records about people, locations and other known entities. For a query like "tallest mountain in the United States," Google's search algorithm can make use of the Knowledge Graph's structured data on mountains and their important attributes, including height.

Thus, to arrive at its answer, the search algorithm parses the user's query, understanding mountain as a type of geographical feature and tallest as a request to compare heights within the region United States. The algorithm then consults the Knowledge Graph and identifies Denali as the relevant entity, subsequently telling the user that Denali is the tallest mountain in the United States. The results might also include additional information the algorithm identifies as possibly relevant, such as Denali's former name, Mount McKinley, and the fact that it is also the highest peak in North America, not just the United States.

Pros and cons of semantic search

As noted above, a semantic search approach offers several benefits over its simpler keyword-based predecessors, but it also comes with several limitations and challenges.

Advantages of semantic search include the following:

• Better relevance and accuracy. The most important benefit of semantic search algorithms is their ability to improve the quality of search results. The ability to infer a searcher's intended meaning and context is especially useful for queries containing ambiguous language or that have different implications based on location or time. For example, with a semantic search algorithm, the query "local restaurants" would yield results in the user's current town.
• Flexibility and adaptability. Semantic search algorithms are dynamic, adjusting in response to new data and user interactions over time. This flexibility enables the algorithm to better reflect emerging trends and changes in language usage, as well as users' preferences. For example, a semantic search algorithm can learn to recognize a new slang term and associate it with older synonyms.
• Improved user experience. Because semantic search algorithms can understand the underlying meaning of a user's question, rather than relying only on the exact words typed, they facilitate simpler, more natural interactions with search engines. For example, if a user types in the natural language query "what time is the NFL game tonight," the search algorithm could provide an answer that takes into account the current date, the football season schedule and the user's time zone.
• Efficient information retrieval. Using semantic analysis with databases such as knowledge graphs is significantly faster than traditional keyword search methods, particularly when combined with predictive analytics and pattern-matching machine learning algorithms. This benefit is particularly important for search engines like Google, which need to sort through an unimaginably vast amount of internet content to provide results.

Disadvantages of semantic search include the following:

• Complexity. While semantic search algorithms' complex architectures give them an advantage over lexical search algorithms, these architectures are also more difficult to plan, build and maintain. They require ongoing updates and algorithm tuning to remain effective, which demands a level of machine learning skills and tooling that might be beyond the reach of many smaller organizations and individual researchers.
• Computational load. Semantic search algorithms' size and complexity also means they require a great deal of computational resources to function, including processing power and extensive memory. Moreover, these compute and memory requirements scale with the amount of data being analyzed. Acquiring, operating and monitoring this compute infrastructure can be highly costly -- not to mention energy-intensive, which raises environmental sustainability concerns.
• Data privacy. Part of what makes semantic search algorithms useful is their ability to understand the specific context in which a user is making their search. But this involves tracking and analyzing user data, such as location, internet browsing behavior and search history. In addition to raising obvious individual information privacy concerns, these practices could lead to regulatory compliance issues in regions with strong data protection laws, such as the European Union's General Data Protection Regulation.
• Algorithmic bias. Like any machine learning model, semantic search algorithms reflect the biases found in their training data. For example, if a semantic search algorithm's training data mainly reflects the experiences of a majority group, it might not accurately represent the diverse realities of minority populations. This can lead to misunderstandings of cultural context and skewed outcomes. Strategies to mitigate algorithmic bias include regular algorithmic audits and constructing diverse training data sets.