Large Language Models (LLMs) have issues with document question answering (QA) in situations where the document is unable to fit in the small context length of an LLM. To overcome this issue, most existing works focus on retrieving the relevant context from the document, representing them as plain text. However, documents such as PDFs, web pages, and presentations are naturally structured with different pages, tables, sections, and so on. Representing such structured documents as plain text is incongruous with the user’s mental model of these documents with rich structure. When a system has to query the document for context, this incongruity is brought to the fore, and seemingly trivial questions can trip up the QA system. To bridge this fundamental gap in handling structured documents, we propose an approach called PDFTriage that enables models to retrieve the context based on either structure or content. Our experiments demonstrate the effectiveness of the proposed PDFTriage-augmented models across several classes of questions where existing retrieval-augmented LLMs fail. To facilitate further research on this fundamental problem, we release our benchmark dataset consisting of 900+ human-generated questions over 80 structured documents from 10 different categories of question types for document QA. Our code and datasets will be released soon on Github.

Modern instruction-tuned models have become highly capable in text generation tasks such as summarization, and are expected to be released at a steady pace. In practice one may now wish to choose confidently, but with minimal effort, the best performing summarization model when applied to a new domain or purpose. In this work, we empirically investigate the test sample size necessary to select a preferred model in the context of news summarization. Empirical results reveal that comparative evaluation converges quickly for both automatic and human evaluation, with clear preferences for a system emerging from under 100 examples. The human preference data allows us to quantify how well automatic scores can reproduce preference rankings across a variety of downstream summarization tasks. We find that, while automatic metrics are stable at smaller sample sizes, only some automatic metrics are able to moderately predict model win rates according to human preference.

Rule-based reasoning, a fundamental type of legal reasoning, enables us to draw conclusions by accurately applying a rule to a set of facts. We explore causal language models as rule-based reasoners, specifically with respect to compositional rules - rules consisting of multiple elements which form a complex logical expression. Reasoning about compositional rules is challenging because it requires multiple reasoning steps, and attending to the logical relationships between elements. We introduce a new prompting method, Chain of Logic, which elicits rule-based reasoning through decomposition (solving elements as independent threads of logic), and recomposition (recombining these sub-answers to resolve the underlying logical expression). This method was inspired by the IRAC (Issue, Rule, Application, Conclusion) framework, a sequential reasoning approach used by lawyers. We evaluate chain of logic across eight rule-based reasoning tasks involving three distinct compositional rules from the LegalBench benchmark and demonstrate it consistently outperforms other prompting methods, including chain of thought and self-ask, using open-source and commercial language models.

Rapid advancements of large language models (LLMs) have enabled the processing, understanding, and generation of human-like text, with increasing integration into systems that touch our social sphere. Despite this success, these models can learn, perpetuate, and amplify harmful social biases. In this article, we present a comprehensive survey of bias evaluation and mitigation techniques for LLMs. We first consolidate, formalize, and expand notions of social bias and fairness in natural language processing, defining distinct facets of harm and introducing several desiderata to operationalize fairness for LLMs. We then unify the literature by proposing three intuitive taxonomies, two for bias evaluation, namely, metrics and datasets, and one for mitigation. Our first taxonomy of metrics for bias evaluation disambiguates the relationship between metrics and evaluation datasets, and organizes metrics by the different levels at which they operate in a model: embeddings, probabilities, and generated text. Our second taxonomy of datasets for bias evaluation categorizes datasets by their structure as counterfactual inputs or prompts, and identifies the targeted harms and social groups; we also release a consolidation of publicly available datasets for improved access. Our third taxonomy of techniques for bias mitigation classifies methods by their intervention during pre-processing, in-training, intra-processing, and post-processing, with granular subcategories that elucidate research trends. Finally, we identify open problems and challenges for future work. Synthesizing a wide range of recent research, we aim to provide a clear guide of the existing literature that empowers researchers and practitioners to better understand and prevent the propagation of bias in LLMs.

Approaches to computational argumentation tasks such as stance detection and aspect detection have largely focused on the text of independent claims, losing out on potentially valuable context provided by the rest of the collection. We introduce a general approach to these tasks motivated by syntopical reading, a reading process that emphasizes comparing and contrasting viewpoints in order to improve topic understanding. To capture collection-level context, we introduce the syntopical graph, a data structure for linking claims within a collection. A syntopical graph is a typed multi-graph where nodes represent claims and edges represent different possible pairwise relationships, such as entailment, paraphrase, or support. Experiments applying syntopical graphs to the problems of detecting stance and aspects demonstrate state-of-the-art performance in each domain, significantly outperforming approaches that do not utilize collection-level information.

Leaderboards are widely used in NLP and push the field forward. While leaderboards are a straightforward ranking of NLP models, this simplicity can mask nuances in evaluation items (examples) and subjects (NLP models). Rather than replace leaderboards, we advocate a re-imagining so that they better highlight if and where progress is made. Building on educational testing, we create a Bayesian leaderboard model where latent subject skill and latent item difficulty predict correct responses. Using this model, we analyze the ranking reliability of leaderboards. Afterwards, we show the model can guide what to annotate, identify annotation errors, detect overfitting, and identify informative examples. We conclude with recommendations for future benchmark tasks.

Text segmentation aims to uncover latent structure by dividing text from a document into coherent sections. Where previous work on text segmentation considers the tasks of document segmentation and segment labeling separately, we show that the tasks contain complementary information and are best addressed jointly. We introduce Segment Pooling LSTM (S-LSTM), which is capable of jointly segmenting a document and labeling segments. In support of joint training, we develop a method for teaching the model to recover from errors by aligning the predicted and ground truth segments. We show that S-LSTM reduces segmentation error by 30% on average, while also improving segment labeling.

Trust is implicit in many online text conversations—striking up new friendships, or asking for tech support. But trust can be betrayed through deception. We study the language and dynamics of deception in the negotiation-based game Diplomacy, where seven players compete for world domination by forging and breaking alliances with each other. Our study with players from the Diplomacy community gathers 17,289 messages annotated by the sender for their intended truthfulness and by the receiver for their perceived truthfulness. Unlike existing datasets, this captures deception in long-lasting relationships, where the interlocutors strategically combine truth with lies to advance objectives. A model that uses power dynamics and conversational contexts can predict when a lie occurs nearly as well as human players.

At about the midpoint of the IARPA MATERIAL program in October 2019, an evaluation was conducted on systems’ abilities to find Lithuanian documents based on English queries. Subsequently, both the Lithuanian test collection and results from all three teams were made available for detailed analysis. This paper capitalizes on that opportunity to begin to look at what’s working well at this stage of the program, and to identify some promising directions for future work.

We describe a modified shared-LSTM network for the Semantic Textual Similarity (STS) task at SemEval-2017. The network builds on previously explored Siamese network architectures. We treat max sentence length as an additional hyperparameter to be tuned (beyond learning rate, regularization, and dropout). Our results demonstrate that hand-tuning max sentence training length significantly improves final accuracy. After optimizing hyperparameters, we train the network on the multilingual semantic similarity task using pre-translated sentences. We achieved a correlation of 0.4792 for all the subtasks. We achieved the fourth highest team correlation for Task 4b, which was our best relative placement.