Zero-Shot Information Extraction as a Unified Text-to-Triple Translation

Abstract

We cast a suite of information extraction tasks into a text-to-triple translation framework. Instead of solving each task relying on taskspecific datasets and models, we formalize the task as a translation between task-specific input text and output triples. By taking the taskspecific input, we enable a task-agnostic translation by leveraging the latent knowledge that a pre-trained language model has about the task. We further demonstrate that a simple pretraining task of predicting which relational information corresponds to which input text is an effective way to produce task-specific outputs. This enables the zero-shot transfer of our framework to downstream tasks. We study the zero-shot performance of this framework on open information extraction (OIE2016, NYT, WEB, PENN), relation classification (FewRel and TACRED), and factual probe (Google-RE and T-REx). The model transfers non-trivially to most tasks and is often competitive with a fully supervised method without the need for any task-specific training. For instance, we significantly outperform the F1 score of the supervised open information extraction without needing to use its training set. 1

ZS4IE: A toolkit for Zero-Shot Information Extraction with simple Verbalizations

“…Furthermore, we compare our system with zeroshot task specific approaches from other authors when available. For RE, Wang et al (2021a) propose a text-to-triple translation method that given a text and a set of entities returns the existing relations. For EE, Lyu et al (2021) to us, the use of an entailment model, but in their case the input sentence is split in clauses according to the output of a Semantic Role Labelling system.…”

In-Context Learning for Data Generation in Aspect-Based Opinion Extraction

“…A few works [13,22,23] have explored LMs as-is or performing prompt engineering, which consists of finding the most appropriate prompt to solve some given task. Liu et al [13] surveys these methods for Natural Language Processing tasks.…”

Knowledge Mining: A Cross-disciplinary Survey

“…Accounting for early approaches in the literature, Yates et al [66] proposed the first Open IE system by using a self-supervised learning approach; Fader et al [67] leveraged POS tag patterns; Del Corro and Gemulla [68] decomposed a sentence into clauses, and Stanovsky et al [69] created the first annotated corpus by an automatic translation from the Question-Answer Meaning Representation dataset and developed an Open IE system using a Bi-LSTM with a BIO tagging scheme. More recently, Ro et al [70] included two classifiers for predicate and arguments; they use hidden states of a BERT model to extract predicates, and then the concatenation of predicate average, BERT hidden sequence, and position embedding are used as inputs for multi-head attention blocks for argument extraction. Wang et al [71] proposed a text-to-triple translation framework that includes generating and ranking steps; it uses Beam search over BERT attention score to generate relevant triples and then rank the generated results using a contrastive pre-trained model.…”

“…More recently, Ro et al [70] included two classifiers for predicate and arguments; they use hidden states of a BERT model to extract predicates, and then the concatenation of predicate average, BERT hidden sequence, and position embedding are used as inputs for multi-head attention blocks for argument extraction. Wang et al [71] proposed a text-to-triple translation framework that includes generating and ranking steps; it uses Beam search over BERT attention score to generate relevant triples and then rank the generated results using a contrastive pre-trained model. On the other hand, relation discovery aims at discovering unseen relation types from unsupervised data; e.g., [72] is a recent work in the literature that casts the task of relation discovery as a clustering task.…”