Neural Semantic Parsing with Extremely Rich Symbolic Meaning Representations

TL;DR

It is shown through neural model probing that training on a taxonomic representation enhances the model’s ability to learn the taxonomical hierarchy, encouraging for research in computational semantics that aims to combine data-driven distributional meanings with knowledge-based symbolic representations.

Abstract

Current open-domain neural semantics parsers show impressive performance. However, closer inspection of the symbolic meaning representations they produce reveals significant weaknesses: sometimes they tend to merely copy character sequences from the source text to form symbolic concepts, defaulting to the most frequent word sense based in the training distribution. By leveraging the hierarchical structure of a lexical ontology, we introduce a novel compositional symbolic representation for concepts based on their position in the taxonomical hierarchy. This representation provides richer semantic information and enhances interpretability. We introduce a neural "taxonomical" semantic parser to utilize this new representation system of predicates, and compare it with a standard neural semantic parser trained on the traditional meaning representation format, employing a novel challenge set and evaluation metric for evaluation. Our experimental findings demonstrate that the taxonomical model, trained on much richer and complex meaning representations, is slightly subordinate in performance to the traditional model using the standard metrics for evaluation, but outperforms it when dealing with out-of-vocabulary concepts. This finding is encouraging for research in computational semantics that aims to combine data-driven distributional meanings with knowledge-based symbolic representations.

Figures (10)

• Figure 1: Graphical display of a meaning representation for the sentence John, a keen birdwatcher, was delighted to see a hobby in the style of the Parallel Meaning Bank. Oval nodes represent concepts, boxed nodes introduce contexts, and labeled edges denote thematic roles or semantic relations.
• Figure 2: Simple illustration of a taxonomical encoding. Each box denotes a concept within a typical ontological ISA-hierarchy. The longer the encoding, the more specific its concepts is.
• Figure 3: Discourse Representation Structure for a sentence shown in box format (left) and sequence notation (right). The corresponding graph for this DRS is shown in Figure \ref{['fig:standard']}.
• Figure 4: Distribution of word senses in the different data splits of the Parallel Meaning Bank. Note that except sense "01", sense "02" is prominent because every person name incorporates either the female.n.02 or male.n.02 and "08" also stands out because every meaning for a tensed clause includes the time.n.08 concept.
• Figure 5: The pipelines for semantic parsing in comparison. Route (1) shows a neural semantic parsing system based on traditional concepts representations. Route (2) illustrates the taxonomy-based parsing system where a mapper interprets the produced symbols.