Answering open-domain questions through unsupervised methods poses challenges for both machine-reading (MR) and language model (LM) -based approaches. The MR-based approach suffers from sparsity issues in extracted knowledge graphs (KGs), while the performance of the LM-based approach significantly depends on the quality of the retrieved context for questions. In this paper, we compare these approaches and propose a novel methodology that leverages directional predicate entailment (inference) to address these limitations. We use entailment graphs (EGs), with natural language predicates as nodes and entailment as edges, to enhance parsed KGs by inferring unseen assertions, effectively mitigating the sparsity problem in the MR-based approach. We also show EGs improve context retrieval for the LM-based approach. Additionally, we present a Boolean QA task, demonstrating that EGs exhibit comparable directional inference capabilities to large language models (LLMs). Our results highlight the importance of inference in open-domain QA and the improvements brought by leveraging EGs.

Large Language Models (LLMs) are claimed to be capable of Natural Language Inference (NLI), necessary for applied tasks like question answering and summarization. We present a series of behavioral studies on several LLM families (LLaMA, GPT-3.5, and PaLM) which probe their behavior using controlled experiments. We establish two biases originating from pretraining which predict much of their behavior, and show that these are major sources of hallucination in generative LLMs. First, memorization at the level of sentences: we show that, regardless of the premise, models falsely label NLI test samples as entailing when the hypothesis is attested in training data, and that entities are used as “indices’ to access the memorized data. Second, statistical patterns of usage learned at the level of corpora: we further show a similar effect when the premise predicate is less frequent than that of the hypothesis in the training data, a bias following from previous studies. We demonstrate that LLMs perform significantly worse on NLI test samples which do not conform to these biases than those which do, and we offer these as valuable controls for future LLM evaluation.