Supervised NRE models have a lack of large-scale high-quality training data because manually labeling data is time-consuming and human intensive. To help with this, distant supervision (DS) assumption has been used to automatically label data by aligning existing knowledge graphs with plain text. For any entity pair in knowledge graphs, sentences mentioning both entities will be labeled with their corresponding relations in knowledge graphs. Large-scale training examples can be easily constructed using this training scheme. Although DS provides a feasible approach to use more data, it is inevitable for its automatic labeling mechanism to be accompanied with the wrong labeling problem, since not all sentences mentioning two given entities express their relations in knowledge graphs exactly. 

Utilizing more data for RE models with distant supervision

Real-world relation distributions are long-tail. This means only the common relations obtain sufficient training instances and most relations have very limited relational facts and corresponding sentences. Few-shot learning focuses on grasping tasks with only a few training examples and is a good fit for this need. The general idea of few-shot models is to train good representations of instances or learn ways of fast adaptation from existing large-scale data, and then transfer to new tasks.

Let T be some task, E be past experience related to this classes of task T, and P be performance measure of T. Few-Shot learning is a type of machine learning problem which is specified by E, T, and P, but only has limited E either quantitatively or qualitatively. E typically includes supervised information and prior knowledge.

Performing more efficient learning with few-shot learning for RE

Most RE systems work within pre-specific relation sets designed by human experts. Since the world undergoes open-ended growth of relations, it is not possible to handle all the emerging relation types only by humans. RE system need to be able not to rely on predefined relation schemas and have the ability to work in open scenarios. 


Orienting more open domains for RE


In Relation Extraction (RE), a single document often mentions multiple entities exhibiting complex cross-sentence relationships. Most existing RE methods focus on intra-sentence extraction and are inadequate for collectively identifying relational facts expressed across long paragraphs. Extracting these facts requires models capable of handling complicated contexts, such as full documents, rather than relying solely on isolated single sentences.

Handling Complicated Contexts in Relation Extraction

- Utilizing more data
- Performing more efficient learning
- Handling more complicated context
- Orienting more open domains

Columbia University

Most RE methods work in a simplified setting, and mainly focus on training models with large amounts of human annotations to classify two given entities within one sentence into pre-defined relations. To build an effective and robust RE system that can be deployed in the real world, more complex scenarios must be further investigated.  
- Collecting high-quality human annotation is expensive and time-consuming
- Many long-tail relations cannot provide large amounts of training examples 
- Most facts are expressed by long context consisting of multiple sentences
- Using a predefined set to cover those relations with open-ended growth is not easy

Learn Before

Related

Learn After