Learning as Abduction: Trainable Natural Logic Theorem Prover for Natural Language Inference

Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 20–31 Barcelona, Spain (Online), December 12–13, 2020 20 Learning as Abduction: Trainable Natural Logic Theorem Prover for Natural Language Inference Lasha Abzianidze UiL OTS, Utrecht University l.abzianidze@uu.nl Abstract Tackling Natural Language Inference with a logic-based method is becoming less and less common. While this might have been coun- terintuitive several decades ago, nowadays it seems pretty obvious. The main reasons for such a conception are that (a) logic-based methods are usually brittle when it comes to processing wide-coverage texts, and (b) instead of automatically learning from data, they require much of manual effort for devel- opment. We make a step towards to over- come such shortcomings by modeling learning from data as abduction: reversing a theorem- proving procedure to abduce semantic rela- tions that serve as the best explanation for the gold label of an inference problem. In other words, instead of proving sentence-level infer- ence relations with the help of lexical relations, the lexical relations are proved taking into account the sentence-level inference relations. We implement the learning method in a tableau theorem prover for natural language and show that it improves the performance of the theo- rem prover on the SICK dataset by 1.4% while still maintaining high precision (> 94%). The obtained results are competitive with the state of the art among logic-based systems. 1 Introduction Natural language inference (NLI) is a well- established task for measuring intelligent sys- tems’ capacity of natural language understanding (Cooper et al., 1996; Dagan et al., 2005). To im- prove and better evaluate the systems on the NLI task, many annotated NLI datasets are prepared and used for training and evaluating NLP models. Gen- erally speaking, an NLI dataset is a set of natural language sentence pairs, called premise-hypothesis This work is licensed under a Creative Commons Attri- bution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/. A clown is putting eyeshadow on The person is quickly removing make-up A clown is putting eyeshadow on The person is quickly removing make-up Proving Abduction eyeshadow ⊑ make-up clown ⊑ person Neutral Contradiction Figure 1: Theorem proving retrieves necessary seman- tic relations from KB, e.g., eyeshadow is make-up. To learn domain-specific relations, e.g., put on is incom- patible with remove (not found in KB), abduction re- verse proves semantic relations from training samples. pairs, that are annotated by crowd workers with one of three inference labels (entailment, contradiction, and neutral), representing a semantic relation from a premise to a hypothesis. Currently, the state-of-the-art systems in NLI are exclusively based on Deep Learning (DL). One of the reasons for this is that DL-based systems can ea- gerly learn from task-related data and also take an advantage form high-quality pre-trained word em- beddings. The training phase helps them to obtain competitive scores on the in-domain test part. On the other hand, logic-based systems are becoming less favored for NLI since it is hard to scale them up for reasoning with wide-coverage sentences. De- …

Content truncated to 500 words