Large Language Models (LLMs) tend to be unreliable on fact-based answers.To address this problem, NLP researchers have proposed a range of techniques to estimate LLM’s confidence over facts. However, due to the lack of a systematic comparison, it is not clear how the different methods compare to one other.To fill this gap, we present a rigorous survey and empirical comparison of estimators of factual confidence.We define an experimental framework allowing for fair comparison, covering both fact-verification and QA. Our experiments across a series of LLMs indicate that trained hidden-state probes provide the most reliable confidence estimates; albeit at the expense of requiring access to weights and supervision data. We also conduct a deeper assessment of the methods, in which we measure the consistency of model behavior under meaning-preserving variations in the input. We find that the factual confidence of LLMs is often unstable across semantically equivalent inputs, suggesting there is much room for improvement for the stability of models’ parametric knowledge.

Pre-trained language models (LMs) obtain state-of-the-art performance when adapted to text classification tasks. However, when using such models in real world applications, efficiency considerations are paramount. In this paper, we study how different training procedures that adapt LMs to text classification perform, as we vary model and train set size. More specifically, we compare standard fine-tuning, prompting, and knowledge distillation (KD) when the teacher was trained with either fine-tuning or prompting. Our findings suggest that even though fine-tuning and prompting work well to train large LMs on large train sets, there are more efficient alternatives that can reduce compute or data cost. Interestingly, we find that prompting combined with KD can reduce compute and data cost at the same time.

To successfully account for language, computational models need to take into account both the linguistic context (the content of the utterances) and the extra-linguistic context (for instance, the participants in a dialogue). We focus on a referential task that asks models to link entity mentions in a TV show to the corresponding characters, and design an architecture that attempts to account for both kinds of context. In particular, our architecture combines a previously proposed specialized module (an “entity library”) for character representation with transfer learning from a pre-trained language model. We find that, although the model does improve linguistic contextualization, it fails to successfully integrate extra-linguistic information about the participants in the dialogue. Our work shows that it is very challenging to incorporate extra-linguistic information into pre-trained language models.

Temporary syntactic ambiguities arise when the beginning of a sentence is compatible with multiple syntactic analyses. We inspect to which extent neural language models (LMs) exhibit uncertainty over such analyses when processing temporarily ambiguous inputs, and how that uncertainty is modulated by disambiguating cues. We probe the LM’s expectations by generating from it: we use stochastic decoding to derive a set of sentence completions, and estimate the probability that the LM assigns to each interpretation based on the distribution of parses across completions. Unlike scoring-based methods for targeted syntactic evaluation, this technique makes it possible to explore completions that are not hypothesized in advance by the researcher. We apply this method to study the behavior of two LMs (GPT2 and an LSTM) on three types of temporary ambiguity, using materials from human sentence processing experiments. We find that LMs can track multiple analyses simultaneously; the degree of uncertainty varies across constructions and contexts. As a response to disambiguating cues, the LMs often select the correct interpretation, but occasional errors point to potential areas of improvement

It is often posited that more predictable parts of a speaker’s meaning tend to be made less explicit, for instance using shorter, less informative words. Studying these dynamics in the domain of referring expressions has proven difficult, with existing studies, both psycholinguistic and corpus-based, providing contradictory results. We test the hypothesis that speakers produce less informative referring expressions (e.g., pronouns vs. full noun phrases) when the context is more informative about the referent, using novel computational estimates of referent predictability. We obtain these estimates training an existing coreference resolution system for English on a new task, masked coreference resolution, giving us a probability distribution over referents that is conditioned on the context but not the referring expression. The resulting system retains standard coreference resolution performance while yielding a better estimate of human-derived referent predictability than previous attempts. A statistical analysis of the relationship between model output and mention form supports the hypothesis that predictability affects the form of a mention, both its morphosyntactic type and its length.

In neural network models of language, words are commonly represented using context-invariant representations (word embeddings) which are then put in context in the hidden layers. Since words are often ambiguous, representing the contextually relevant information is not trivial. We investigate how an LSTM language model deals with lexical ambiguity in English, designing a method to probe its hidden representations for lexical and contextual information about words. We find that both types of information are represented to a large extent, but also that there is room for improvement for contextual information.

Humans use language to refer to entities in the external world. Motivated by this, in recent years several models that incorporate a bias towards learning entity representations have been proposed. Such entity-centric models have shown empirical success, but we still know little about why. In this paper we analyze the behavior of two recently proposed entity-centric models in a referential task, Entity Linking in Multi-party Dialogue (SemEval 2018 Task 4). We show that these models outperform the state of the art on this task, and that they do better on lower frequency entities than a counterpart model that is not entity-centric, with the same model size. We argue that making models entity-centric naturally fosters good architectural decisions. However, we also show that these models do not really build entity representations and that they make poor use of linguistic context. These negative results underscore the need for model analysis, to test whether the motivations for particular architectures are borne out in how models behave when deployed.

This paper describes our winning contribution to SemEval 2018 Task 4: Character Identification on Multiparty Dialogues. It is a simple, standard model with one key innovation, an entity library. Our results show that this innovation greatly facilitates the identification of infrequent characters. Because of the generic nature of our model, this finding is potentially relevant to any task that requires the effective learning from sparse or imbalanced data.

Recent state-of-the-art neural language models share the representations of words given by the input and output mappings. We propose a simple modification to these architectures that decouples the hidden state from the word embedding prediction. Our architecture leads to comparable or better results compared to previous tied models and models without tying, with a much smaller number of parameters. We also extend our proposal to word2vec models, showing that tying is appropriate for general word prediction tasks.