This month, we released a breakthrough update to our natural language processing technology: an enhanced semantic classification of adjectives designed to advance machine understanding. We have engineered this new linguistic framework to help AI systems comprehend the complex and nuanced meanings of adjectives in context.
Like other parts of speech, adjectives are words rich in meanings, allowing us to express various concepts like qualities, states or emotions. However, adjectives can also be ambiguous and open to interpretation. This updated classification provides a structured taxonomy of adjectives divided into seven meaningful branches: characteristics, situations, quantities, feelings, taste, actions and relationships. Our new categorization helps AI models recognize and represent the connections between adjectives and the concepts they modify.
As we discussed in our recent article, The Importance of Disambiguation in Natural Language Processing, accurately understanding meaning and intent is crucial for NLP projects. Our enhanced semantic classification builds upon Lettria's existing disambiguation capabilities to provide AI models with an even stronger foundation in linguistics.
With this improved foundation in linguistics, Lettria continues to push the boundaries of natural language processing for business. Our new semantic classification translates directly into better performance in key NLP techniques like sentiment analysis, product catalog enrichment and conversational AI. This guide details how the updated taxonomy will enhance our machine learning models and empower organizations with optimized artificial intelligence.
Discover our Semantic Hierarchy
The characteristics branch includes adjectives describing living things, objects, or concepts, whether concrete or abstract, permanent or not. This information is typically found in semantic structuring or ontologies as class or individual attributes. In addition to very general categories concerning measurement, quality or importance, there are categories describing physical properties like smell, taste, sound, texture, shape, color, and other visual characteristics. Human (and sometimes animal) characteristics like intelligence or kindness are also included.
The situation branch contains adjectives indicating something or someone's state or situation: ways of being, spatial characteristics, being related to something without expressly mentioning that thing (those adjectives are in another category), relative situation in space, time, or hierarchy, anteriority, posteriority, or simultaneity, being ancient or new, regularity, and other temporal characteristics.
The quantity branch allows expressing numerical properties: number, order, fraction, multiplication, small or large quantity, and being total or partial.
The feeling branch is divided into two sub-branches. The first contains adjectives indicating the referent experiences a feeling or emotion. The second indicates the referent arouses a feeling or emotion in someone else. This distinction between adjectives qualifying a patient and those qualifying an agent (in the linguistic meanings) is critical for properly structuring information and avoiding misinterpretation.
The taste branch also divides into two sub-branches. The first contains adjectives indicating being attracted, repelled, or indifferent to something or someone. The second expresses attracting, repelling, or leaving someone indifferent.
The action branch divides into two categories grouping adjectives related to actions. The "is_doing" adjectives mention the referent is performing an action. The "likely_to" indicate the possibility of performing or undergoing an action.
Finally, the relational category is a branch of its own for relational adjectives indicating a relationship with something. This is a clearly identified adjective category in contemporary grammar with quite different syntactic properties than other adjectives.
Want to learn how to build a private ChatGPT using open-source technology?
Due to polysemy, some adjectives can be in multiple categories depending on context. For example, sweet could be in the flavor category or in the good_human_characteristic category. The context determines which meaning is intended.
The names of some categories could be debated. For example, "making_need" in the feeling branch is not particularly transparent. Alternative names might be "covet-inducing" or "desirous." However, the names are just labels, while the important thing is the underlying concepts and meanings.
As languages are not mathematical, there are borderline cases that could arguably be in one category or another. The categorization is not always clear-cut or definitive. However, the approach of assigning the most representative or central category according to usage and meaning is pragmatic.
The taxonomy covers not only qualifying adjectives but also relational adjectives, which provide a crucial mechanism for linking entities or concepts. The relational branch is key for semantic networks and ontologies.
How will this affect projects built with Lettria?
Taking sentiment analysis projects as a key example, the expanded "feeling" branch provides more nuanced categorization of emotion-conveying adjectives. By distinguishing between adjectives describing a subject's own feelings and those describing the feelings the subject arouses in others, our models can gain a richer understanding of the sentiment being expressed. Recognizing these nuances will result in more accurate classification of positive, negative or neutral sentiment.
For product catalog enrichment, the characteristics and attributes expressed by adjectives are essential to capturing a product's properties and qualities. The categories under "characteristics" and "quantity" map directly to the types of attributes needed to describe products in categories like apparel, food and beverages, mechanical parts, and more. Our models can now identify more types of attributes from product descriptions, allowing us to suggest additional structured attributes to include in product catalogs. The "relationships" branch also provides a way to identify connections between products and components or accessories.
The whole process of disambiguation and structuring within the Lettria platform has seen a major update with these latest adjective enhancements. By enriching our modeling of adjective meaning, the Lettria platform continues to push the boundaries of machine understanding of language. This improved foundation in linguistics translates to better performance in key NLP applications for business. Our mission is to build AI with true language intelligence, and advancing semantic classification is fundamental to achieving that goal.
Our updated adjective taxonomy is a practical framework for representing and understanding adjective meaning. The categorization could continue to be improved and expanded; however, as a broad-coverage foundation, it achieves the goal of facilitating natural language processing, semantic interoperability and ontology development. The relational branch, in particular, provides a structure for linking entities via adjectives that denote relationships. On the whole, the taxonomy is an informative model for adjective semantics.