Generative language models have recently shown remarkable success in generating answers to questions in a given textual context. However, these answers may suffer from hallucination, wrongly cite evidence, and spread misleading information. In this work, we address this problem by employing ChatGPT, a state-of-the-art generative model, as a machine-reading system. We ask it to retrieve answers to lexically varied and open-ended questions from trustworthy instructive texts. We introduce WHERE (WikiHow Evidence REtrieval), a new high-quality evaluation benchmark of a set of WikiHow articles exhaustively annotated with evidence sentences to questions that comes with a special challenge: All questions are about the article’s topic, but not all can be answered using the provided context. We interestingly find that when using a regular question-answering prompt, ChatGPT neglects to detect the unanswerable cases. When provided with a few examples, it learns to better judge whether a text provides answer evidence or not. Alongside this important finding, our dataset defines a new benchmark for evidence retrieval in question answering, which we argue is one of the necessary next steps for making large language models more trustworthy.

Natural language inherently consists of implicit and underspecified phrases, which represent potential sources of misunderstanding. In this paper, we present a data set of such phrases in English from instructional texts together with multiple possible clarifications. Our data set, henceforth called CLAIRE, is based on a corpus of revision histories from wikiHow, from which we extract human clarifications that resolve an implicit or underspecified phrase. We show how language modeling can be used to generate alternate clarifications, which may or may not be compatible with the human clarification. Based on plausibility judgements for each clarification, we define the task of distinguishing between plausible and implausible clarifications. We provide several baseline models for this task and analyze to what extent different clarifications represent multiple readings as a first step to investigate misunderstandings caused by implicit/underspecified language in instructional texts.

Communicating efficiently in natural language requires that we often leave information implicit, especially in spontaneous speech. This frequently results in phenomena of incompleteness, such as omitted references, that pose challenges for language processing. In this survey paper, we review the state of the art in research regarding the automatic processing of such implicit references in dialog scenarios, discuss weaknesses with respect to inconsistencies in task definitions and terminologies, and outline directions for future work. Among others, these include a unification of existing tasks, addressing data scarcity, and taking into account model and annotator uncertainties.

We describe SemEval-2022 Task 7, a shared task on rating the plausibility of clarifications in instructional texts. The dataset for this task consists of manually clarified how-to guides for which we generated alternative clarifications and collected human plausibility judgements. The task of participating systems was to automatically determine the plausibility of a clarification in the respective context. In total, 21 participants took part in this task, with the best system achieving an accuracy of 68.9%. This report summarizes the results and findings from 8 teams and their system descriptions. Finally, we show in an additional evaluation that predictions by the top participating team make it possible to identify contexts with multiple plausible clarifications with an accuracy of 75.2%.

This paper describes the data, task setup, and results of the shared task at the First Workshop on Understanding Implicit and Underspecified Language (UnImplicit). The task requires computational models to predict whether a sentence contains aspects of meaning that are contextually unspecified and thus require clarification. Two teams participated and the best scoring system achieved an accuracy of 68%.

The usage of (co-)referring expressions in discourse contributes to the coherence of a text. However, text comprehension can be difficult when referring expressions are non-verbalized and have to be resolved in the discourse context. In this paper, we propose a novel dataset of such implicit references, which we automatically derive from insertions of references in collaboratively edited how-to guides. Our dataset consists of 6,014 instances, making it one of the largest datasets of implicit references and a useful starting point to investigate misunderstandings caused by underspecified language. We test different methods for resolving implicit references in our dataset based on the Generative Pre-trained Transformer model (GPT) and compare them to heuristic baselines. Our experiments indicate that GPT can accurately resolve the majority of implicit references in our data. Finally, we investigate remaining errors and examine human preferences regarding different resolutions of an implicit reference given the discourse context.

In community-edited resources such as wikiHow, sentences are subject to revisions on a daily basis. Recent work has shown that resulting improvements over time can be modelled computationally, assuming that each revision contributes to the improvement. We take a closer look at a subset of such revisions, for which we attempt to improve a computational model and validate in how far the assumption that ‘revised means better’ actually holds. The subset of revisions considered here are noun substitutions, which often involve interesting semantic relations, including synonymy, antonymy and hypernymy. Despite the high semantic relatedness, we find that a supervised classifier can distinguish the revised version of a sentence from an original version with an accuracy close to 70%, when taking context into account. In a human annotation study, we observe that annotators identify the revised sentence as the ‘better version’ with similar performance. Our analysis reveals a fair agreement among annotators when a revision improves fluency. In contrast, noun substitutions that involve other lexical-semantic relationships are often perceived as being equally good or tend to cause disagreements. While these findings are also reflected in classification scores, a comparison of results shows that our model fails in cases where humans can resort to factual knowledge or intuitions about the required level of specificity.

Instructional texts, such as articles in wikiHow, describe the actions necessary to accomplish a certain goal. In wikiHow and other resources, such instructions are subject to revision edits on a regular basis. Do these edits improve instructions only in terms of style and correctness, or do they provide clarifications necessary to follow the instructions and to accomplish the goal? We describe a resource and first studies towards answering this question. Specifically, we create wikiHowToImprove, a collection of revision histories for about 2.7 million sentences from about 246000 wikiHow articles. We describe human annotation studies on categorizing a subset of sentence-level edits and provide baseline models for the task of automatically distinguishing “older” from “newer” revisions of a sentence.

wikiHow is a resource of how-to guidesthat describe the steps necessary to accomplish a goal. Guides in this resource are regularly edited by a community of users, who try to improve instructions in terms of style, clarity and correctness. In this work, we test whether the need for such edits can be predicted automatically. For this task, we extend an existing resource of textual edits with a complementary set of approx. 4 million sentences that remain unedited over time and report on the outcome of two revision modeling experiments.

In this paper we describe our participation in the SemEval 2019 shared task on hyperpartisan news detection. We present the system that we submitted for final evaluation and the three approaches that we used: sentiment, bias-laden words and filtered n-gram features. Our submitted model is a Linear SVM that solely relies on the negative sentiment of a document. We achieved an accuracy of 0.621 and a f1 score of 0.694 in the competition, revealing the predictive power of negative sentiment for this task. There was no major improvement by adding or substituting the features of the other two approaches that we tried.