Semantic parsing is the process of mapping natural language into component parts, with the intention of translating the text into a logical form that a machine can understand. One example of this is translating a search engine query into a SQL query, which can then be used to search a database. While there are similarities between semantic parsing and machine translation, the two processes differ in one important way – while in machine translation the target output is human readable, in semantic parsing the target output is machine readable.
In semantic parsing development and research, there are four major models, with most instances using one or a combination of the models. These models are the Compositional Semantic Model, the Translation Model, Rule Extraction, and the Probabilistic Model. In the Compositional Semantic Model, algorithms generate compositional meaning representations such as logical forms and functional representations. These representations form the backbone of machine language. This process essentially transforms grammar structures into logical ones. The Translation Model works to map text inputs to representations structures like context-free grammar, which are recursive rules used to generate patterns of strings. Because context-free grammar applies defined rules to the text and translates it into a logical form, it is then easier to convert into a final, machine-readable product.
In Rule Extraction, algorithms find the candidate translation rules. One method of doing this is through the construction of SILT trees, or trees for the Semantic Interpretation by Learning Transformations. SILT uses pattern-based transformation rules to map phrases in natural language to productions in machine-languages. In the Probabilistic Model, a variety of techniques are applied to learn and find the best translations. One such technique is known as semantic-based Probabilistic Context Free Grammar (PCFG). PCFG takes context-free grammar techniques one step further by associating rules with probabilities that are derived through annotated training data. The probabilistic data is then used in the case of multiple translation possibilities – the final selection can be made based on the overall probability of a parse tree for a given statement.