A hybrid approach to Natural Language Inference for the SICK dataset

Natural Language Inference (NLI) can be described as the task of answering if a short text called Hypothesis (H) can be inferred from another text called Premise (P) (Poliak, 2020; Dagan et al., 2013). Affirmative answers are considered as semantic entailments and negative ones are either contradictions or semantically “neutral” statements. In the last three decades, many Natural Language Processing (NLP) methods have been put to use for solving this task. As it so happened to almost every other NLP task, Deep Learning (DL) techniques in general (and Transformer neural networks in particular) have been achieving the best results in this task in recent years, progressively increasing their outcomes when compared to classical, symbolic Knowledge Representation models in solving NLI.

Nevertheless, however successful DL models are in measurable results like accuracy and F-score, their outcomes are far from being explicable, and this is an undesirable feature specially in a task such as NLI, which is meant to deal with language understanding together with rational reasoning inherent to entailment and to contradiction judgements. It is therefore tempting to evaluate how more explainable models would perform in NLI and to compare their performance with DL models later on.

This paper puts forth a pipeline that we called IsoLex. It provides explainable, transparent NLP models for NLI. It has been tested on a partial version of the SICK corpus (Marelli, 2014) called SICK-CE, containing only the contradiction and the entailment pairs (4245 in total), thus leaving aside the neutral pairs, as an attempt to concentrate on unambiguous semantic relationships, which arguably favor the intelligibility of the results.

The pipeline consists of three serialized commonly used NLP models: first, an Isolation Forest module is used to filter off highly dissimilar Premise-Hypothesis pairs; second, a WordNet-based Lexical Relations module is employed to check whether the Premise and the Hypothesis textual contents are related to each other in terms of synonymy, hyperonymy, or holonymy; finally, similarities between Premise and Hypothesis texts are evaluated by a simple cosine similarity function based on Word2Vec embeddings.

IsoLex has achieved 92% accuracy and 94% F-1 on SICK-CE. This is close to SOTA models for this kind of task, such as RoBERTa with a 98% accuracy and 99% F-1 on the same dataset.

The small performance gap between IsoLex and SOTA DL models is largely compensated by intelligibility on every step of the proposed pipeline. At anytime it is possible to evaluate the role of similarity, lexical relatedness and so forth in the overall process of inference.