Analyzing the Effects of Reasoning Types on Cross-Lingual Transfer Performance
- Karthikeyan K ,
- Aalok Sathe ,
- Somak Aditya ,
- Monojit Choudhury
Multilingual language models achieve impressive zero-shot accuracies in many languages in complex tasks such as Natural Language Inference (NLI). Examples in NLI (and equivalent complex tasks) often pertain to various types of sub-tasks, requiring different kinds of reasoning. Certain types of reasoning have proven to be more difficult to learn in a monolingual context, and in the crosslingual context, similar observations may shed light on zero-shot transfer efficiency and few-shot sample selection. Hence, to investigate the effects of types of reasoning on transfer performance, we propose a category-annotated multilingual NLI dataset and discuss the challenges to scale monolingual annotations to multiple languages. We statistically observe interesting effects that the confluence of reasoning types and language similarities have on transfer performance.
Publication Downloads
TaxiXNLI
October 14, 2021
This repository contains necessary data associated with the Analyzing the Effects of Reasoning Types on Cross-Lingual Transfer Performance, published in EMNLP 2021 Multilingual Representation Learning workshop. This data is a multi-lingual extension of ID 3860 released dataset. Similar to the ID 3397 and 3860, the broad goal for this project is to research and develop Neuro-Symbolic systems for Natural Language Inferencing (NLI) to leverage the "correctness" guarantees and interpretability of symbolic systems in neural network inference. In ID 3860, we proposed TaxiNLI data where we annotated an already public NLI dataset (MultiNLI) with the types of reasoning required to solve each example. Multi-NLI, each example has a premise, hypothesis sentence and a label. In TaxiNLI, the annotations will only add 0s/1s against reasoning categories such as lexical, syntactic, spatial etc. In this data, we translate the TaxiNLI dataset automatically using Bing to Spanish, French, Russian, Hindi, Arabic, Vietnamese, Chinese, Swahili, and Urdu. And, we also use another public dataset (XNLI), sample 1.4k XNLI examples, and annotate with a few selected interesting categories (Negation, Boolean, Spatial, Causal, Temporal, Knowledge).