How to Bootstrap Ontology Generation and Mapping for eCommerce chatbots

First, an ontology generation tool automatically generates a prototype ontology based on values extracted from existing sources: catalogues, search engine facets… We specify that the values of the “product_type” node or facet, for example,  should be mapped to objects in the ontology, and all other attributes for those objects. The NLU Framework includes support for special data types (currency, quantities/measurements…) and the ontology generation tool will take care of detecting these whenever possible, but manual configuration may be necessary.

Second, a separate tool scans the ontology, and generates possible synonyms or variants for object names and attribute values by querying various sources, including word embeddings trained on domain-specific texts. The results are reviewed manually and incorporated into the ontology as needed.

Finally, the ontology is reviewed manually to ensure that data types have been correctly assigned, and to specify default attributes for certain data types. For example, currency amounts will be mapped to the “price” attribute by default; similarly, any standalone number will be interpreted as prices by default.

When processing a parse tree corresponding to a user query, the NLU Framework will map objects and modifiers to elements in the ontology. In the example above, “LG” can be found in the ontology as a value of the “brand” attribute, whereas “french door” is a “refrigerator_type”.

Special handling is provided for numerical data types when they appear as modifiers. This allows the system to understand expressions such as “more than 200 lbs” or “for less than $100”.

As we have seen, an ontology involves a large number of tasks but automation is starting to enter this space as well.