Facts as Experts: Adaptable and Interpretable Neural Memory over Symbolic Knowledge

By Pat Verga, Haitian Sun, Livio Baldini Soares, William W. Cohen. 2020

54 citations recorded by the source corpus.

Massive language models are the core of modern NLP modeling and have been shown to encode impressive amounts of commonsense and factual information. However, that knowledge exists only within the latent parameters of the model, inaccessible to inspection and interpretation, and even worse, factual information memorized from the training corpora is likely to become stale as the world changes. Knowledge stored as parameters will also inevitably exhibit all of the biases inherent in the source materials. To address these problems, we develop a neural language model that includes an explicit interface between symbolically interpretable factual information and subsymbolic neural knowledge. We show that this model dramatically improves performance on two knowledge-intensive question-answering tasks. More interestingly, the model can be updated without re-training by manipulating its symbolic representations. In particular this model allows us to add new facts and overwrite existing ones in ways that are not possible for earlier models.

Subfields: Artificial Intelligence, Safety, Robustness & Interpretability, Knowledge Graphs & Representation

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Safety, Robustness & Interpretability

In Knowledge Graphs & Representation

Citations in the corpus

Modifying Memories in Transformer Models (2020, methodology): “FaE (Verga et al., 2020) .”
Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets (2020, background): “Closest to our work, Verga et al. (2020) observe similar answer overlap in knowledge-base QA, and explore results on non-overlapping sub-sets.”
Pre-Trained Models: Past, Present and Future (2021, background): “Besides tensorizing the textual corpora (Verga et al., 2020; F ´ evry et al., 2020), propose to tensorize entities and triples in existing knowledge bases.”
Knowledge Graph Based Synthetic Corpus Generation for Knowledge-Enhanced Language Model Pre-training (2021, background): “Verga et al. (2020) extend Févry et al. (2020) by adding a triple memory with (subject, relation) encoding as the key and the object encoding as the value.”
Fine-tuning Image Transformers using Learnable Memory (2022, background): “More recently, in the context of transformers for language modeling, several papers [10, 12, 14, 25, 39] propose to augment a model with access to fixed knowledge base.”
A Review on Language Models as Knowledge Bases (2022, methodology): “Building off of this, FAE (Verga et al., 2020) adds an additional memory that encodes triples from a symbolic KB that can be accessed with key-value memory to extract facts and improve predictions.”
Large Scale Knowledge Graph Based Synthetic Corpus Generation for Knowledge-Enhanced Language Model Pre-training (2020, background): “Verga et al. (2020) extend (Févry et al., 2020) by adding a triple memory in addition to an entity memory.”
Relation-Guided Pre-Training for Open-Domain Question Answering (2021, unknown): “Verga et al. (2020) incorporate the knowledge graph triplets into language model, so the model could utilize the triplets to predict correct entity.”
Relational World Knowledge Representation in Contextual Language Models: A Review (2021, background): “These ﬁndings suggest that storing some entity knowledge externally (e.g., in a dense memory, Févry et al., 2020) may be more robust, for example by allowing for ef-ﬁcient updates to the LM’s knowledge (Verga et al., 2020).”
HittER: Hierarchical Transformers for Knowledge Graph Embeddings (2020, background): “, 2020) and reasoning (Riedel et al., 2013; Xiong et al., 2017; Bauer et al., 2018; Verga et al., 2020), learning representations of knowledge graphs has been studied in a large body of literature.”
Reasoning Over Virtual Knowledge Bases With Open Predicate Relations (2021, methodology): “FaE (Verga et al., 2020), learns to retrieve relevant entries from an external memory and mix the retrieved information into the language model to make its final predictions (see Figure 3).”