The majority of Neural Semantic Parsing (NSP) models are developed with the assumption that there are no concepts outside the ones such models can represent with their target symbols (closed-world assumption). This assumption leads to generate hallucinated outputs rather than admitting their lack of knowledge. Hallucinations can lead to wrong or potentially offensive responses to users. Hence, a mechanism to prevent this behavior is crucial to build trusted NSP-based Question Answering agents. To that end, we propose the Hallucination Simulation Framework (HSF), a general setting for stimulating and analyzing NSP model hallucinations. The framework can be applied to any NSP task with a closed-ontology. Using the proposed framework and KQA Pro as the benchmark dataset, we assess state-of-the-art techniques for hallucination detection. We then present a novel hallucination detection strategy that exploits the computational graph of the NSP model to detect the NSP hallucinations in the presence of ontology gaps, out-of-domain utterances, and to recognize NSP errors, improving the F1-Score respectively by similar to 21%, similar to 24% and similar to 1%. This is the first work in closed-ontology NSP that addresses the problem of recognizing ontology gaps. We release our code and checkpoints at https://github.com/amazon-science/ handling- ontology- gaps-in-semantic-parsing.
We explore the relationship between factuality and Natural Language Inference (NLI) by introducing FactRel -- a novel annotation scheme that models \textit{factual} rather than \textit{textual} entailment, and use it to annotate a dataset of naturally occurring sentences from news articles. Our analysis shows that 84\% of factually supporting pairs and 63\% of factually undermining pairs do not amount to NLI entailment or contradiction, respectively, suggesting that factual relationships are more apt for analyzing media discourse. We experiment with models for pairwise classification on the new dataset, and find that in some cases, generating synthetic data with GPT-4 on the basis of the annotated dataset can improve performance. Surprisingly, few-shot learning with GPT-4 yields strong results on par with medium LMs (DeBERTa) trained on the labelled dataset. We hypothesize that these results indicate the fundamental dependence of this task on both world knowledge and advanced reasoning abilities.
Attributing answer text to its source document for information-seeking questions is crucial for building trustworthy, reliable, and accountable systems. We formulate a new task of post-hoc answer attribution for long document comprehension (LDC). Owing to the lack of long-form abstractive and information-seeking LDC datasets, we refactor existing datasets to assess the strengths and weaknesses of existing retrieval-based and proposed answer decomposition and textual entailment-based optimal selection attribution systems for this task. We throw light on the limitations of existing datasets and the need for datasets to assess the actual performance of systems on this task.
With the aim of improving the state-of-the-art (SOTA) on a target task, a standard strategy in Natural Language Processing (NLP) research is to design a new model, or modify the existing SOTA, and then benchmark its performance on the target task. We argue in favor of enriching this chain of actions by a preliminary error-guided analysis: First, explore weaknesses by analyzing the hard cases where the existing model fails, and then target the improvement based on those. Interpretable evaluation has received little attention for structured prediction tasks. Therefore we propose the first in-depth analysis suite for Relation Classification (RC), and show its effectiveness through a case study. We propose a set of potentially influential attributes to focus on (e.g., entity distance, sentence length). Then, we bucket our datasets based on these attributes, and weight the importance of them through correlations. This allows us to identify highly challenging scenarios for the RC model. By exploiting the findings of our analysis, with a carefully targeted adjustment to our architecture, we effectively improve the performance over the baseline by >3 Micro-F1.
Recently, large pre-trained language models (LLMs) have demonstrated superior language understanding abilities, including zero-shot causal reasoning. However, it is unclear to what extent their capabilities are similar to human ones. We here study the processing of an event $B$ in a script-based story, which causally depends on a previous event $A$. In our manipulation, event $A$ is stated, negated, or omitted in an earlier section of the text. We first conducted a self-paced reading experiment, which showed that humans exhibit significantly longer reading times when causal conflicts exist ($\neg A \rightarrow B$) than under logical conditions ($A \rightarrow B$). However, reading times remain similar when cause A is not explicitly mentioned, indicating that humans can easily infer event B from their script knowledge. We then tested a variety of LLMs on the same data to check to what extent the models replicate human behavior. Our experiments show that 1) only recent LLMs, like GPT-3 or Vicuna, correlate with human behavior in the $\neg A \rightarrow B$ condition. 2) Despite this correlation, all models still fail to predict that $nil \rightarrow B$ is less surprising than $\neg A \rightarrow B$, indicating that LLMs still have difficulties integrating script knowledge. Our code and collected data set are available at https://github.com/tony-hong/causal-script.
