Do language models have beliefs about the world? Dennett (1995) famously argues that even thermostats have beliefs, on the view that a belief is simply an informational state decoupled from any motivational state. In this paper, we discuss approaches to detecting when models have beliefs about the world, and we improve on methods for updating model beliefs to be more truthful, with a focus on methods based on learned optimizers or hypernetworks. Our main contributions include: (1) new metrics for evaluating belief-updating methods that focus on the logical consistency of beliefs, (2) a training objective for Sequential, Local, and Generalizing model updates (SLAG) that improves the performance of learned optimizers, and (3) the introduction of the belief graph, which is a new form of interface with language models that shows the interdependencies between model beliefs. Our experiments suggest that models possess belief-like qualities to only a limited extent, but update methods can both fix incorrect model beliefs and greatly improve their consistency. Although off-the-shelf optimizers are surprisingly strong belief-updating baselines, our learned optimizers can outperform them in more difficult settings than have been considered in past work. Code is available at https://github.com/peterbhase/SLAG-Belief-Updating
Current language models can generate high-quality text. Are they simply copying text they have seen before, or have they learned generalizable linguistic abstractions? To tease apart these possibilities, we introduce RAVEN, a suite of analyses for assessing the novelty of generated text, focusing on sequential structure (n-grams) and syntactic structure. We apply these analyses to four neural language models (an LSTM, a Transformer, Transformer-XL, and GPT-2). For local structure - e.g., individual dependencies - model-generated text is substantially less novel than our baseline of human-generated text from each model's test set. For larger-scale structure - e.g., overall sentence structure - model-generated text is as novel or even more novel than the human-generated baseline, but models still sometimes copy substantially, in some cases duplicating passages over 1,000 words long from the training set. We also perform extensive manual analysis showing that GPT-2's novel text is usually well-formed morphologically and syntactically but has reasonably frequent semantic issues (e.g., being self-contradictory).
Dialogue summarization helps readers capture salient information from long conversations in meetings, interviews, and TV series. However, real-world dialogues pose a great challenge to current summarization models, as the dialogue length typically exceeds the input limits imposed by recent transformer-based pre-trained models, and the interactive nature of dialogues makes relevant information more context-dependent and sparsely distributed than news articles. In this work, we perform a comprehensive study on long dialogue summarization by investigating three strategies to deal with the lengthy input problem and locate relevant information: (1) extended transformer models such as Longformer, (2) retrieve-then-summarize pipeline models with several dialogue utterance retrieval methods, and (3) hierarchical dialogue encoding models such as HMNet. Our experimental results on three long dialogue datasets (QMSum, MediaSum, SummScreen) show that the retrieve-then-summarize pipeline models yield the best performance. We also demonstrate that the summary quality can be further improved with a stronger retrieval model and pretraining on proper external summarization datasets.
Recent years have brought about an interest in the challenging task of summarizing conversation threads (meetings, online discussions, etc.). Such summaries help analysis of the long text to quickly catch up with the decisions made and thus improve our work or communication efficiency. To spur research in thread summarization, we have developed an abstractive Email Thread Summarization (EmailSum) dataset, which contains human-annotated short (<30 words) and long (<100 words) summaries of 2549 email threads (each containing 3 to 10 emails) over a wide variety of topics. We perform a comprehensive empirical study to explore different summarization techniques (including extractive and abstractive methods, single-document and hierarchical models, as well as transfer and semisupervised learning) and conduct human evaluations on both short and long summary generation tasks. Our results reveal the key challenges of current abstractive summarization models in this task, such as understanding the sender's intent and identifying the roles of sender and receiver. Furthermore, we find that widely used automatic evaluation metrics (ROUGE, BERTScore) are weakly correlated with human judgments on this email thread summarization task. Hence, we emphasize the importance of human evaluation and the development of better metrics by the community. Our code and summary data have been made available at: https://github.com/ZhangShiyue/EmailSum
Abstractive summarization, the task of generating a concise summary of input documents, requires: (1) reasoning over the source document to determine the salient pieces of information scattered across the long document, and (2) composing a cohesive text by reconstructing these salient facts into a shorter summary that faithfully reflects the complex relations connecting these facts. In this paper, we adapt TP-TRANSFORMER (Schlag et al., 2019), an architecture that enriches the original Transformer (Vaswani et al., 2017) with the explicitly compositional Tensor Product Representation (TPR), for the task of abstractive summarization. The key feature of our model is a structural bias that we introduce by encoding two separate representations for each token to represent the syntactic structure (with role vectors) and semantic content (with filler vectors) separately. The model then binds the role and filler vectors into the TPR as the layer output. We argue that the structured intermediate representations enable the model to take better control of the contents (salient facts) and structures (the syntax that connects the facts) when generating the summary. Empirically, we show that our TP-TRANSFORMER outperforms the Transformer and the original TP-TRANSFORMER significantly on several abstractive summarization datasets based on both automatic and human evaluations. On several syntactic and semantic probing tasks, we demonstrate the emergent structural information in the role vectors and improved syntactic interpretability in the TPR layer outputs. Code and models are available at https://github.com/jiangycTarheel/TPT-Summ.
Meetings are a key component of human collaboration. As increasing numbers of meetings are recorded and transcribed, meeting summaries have become essential to remind those who may or may not have attended the meetings about the key decisions made and the tasks to be completed. However, it is hard to create a single short summary that covers all the content of a long meeting involving multiple people and topics. In order to satisfy the needs of different types of users, we define a new query-based multi-domain meeting summarization task, where models have to select and summarize relevant spans of meetings in response to a query, and we introduce QMSum, a new benchmark for this task. QMSum consists of 1,808 query-summary pairs over 232 meetings in multiple domains. Besides, we investigate a locate-then-summarize method and evaluate a set of strong summarization baselines on the task. Experimental results and manual analysis reveal that QMSum presents significant challenges in long meeting summarization for future research. Dataset is available at \url{https://github.com/Yale-LILY/QMSum}.
The progress in Query-focused Multi-Document Summarization (QMDS) has been limited by the lack of sufficient largescale high-quality training datasets. We present two QMDS training datasets, which we construct using two data augmentation methods: (1) transferring the commonly used single-document CNN/Daily Mail summarization dataset to create the QMDSCNN dataset, and (2) mining search-query logs to create the QMDSIR dataset. These two datasets have complementary properties, i.e., QMDSCNN has real summaries but queries are simulated, while QMDSIR has real queries but simulated summaries. To cover both these real summary and query aspects, we build abstractive end-to-end neural network models on the combined datasets that yield new state-of-the-art transfer results on DUC datasets. We also introduce new hierarchical encoders that enable a more efficient encoding of the query together with multiple documents. Empirical results demonstrate that our data augmentation and encoding methods outperform baseline models on automatic metrics, as well as on human evaluations along multiple attributes.
Neuro-symbolic representations have proved effective in learning structure information in vision and language. In this paper, we propose a new model architecture for learning multi-modal neuro-symbolic representations for video captioning. Our approach uses a dictionary learning-based method of learning relations between videos and their paired text descriptions. We refer to these relations as relative roles and leverage them to make each token role-aware using attention. This results in a more structured and interpretable architecture that incorporates modality-specific inductive biases for the captioning task. Intuitively, the model is able to learn spatial, temporal, and cross-modal relations in a given pair of video and text. The disentanglement achieved by our proposal gives the model more capacity to capture multi-modal structures which result in captions with higher quality for videos. Our experiments on two established video captioning datasets verifies the effectiveness of the proposed approach based on automatic metrics. We further conduct a human evaluation to measure the grounding and relevance of the generated captions and observe consistent improvement for the proposed model. The codes and trained models can be found at https://github.com/hassanhub/R3Transformer
Text generation models can generate factually inconsistent text containing distorted or fabricated facts about the source text. Recent work has focused on building evaluation models to verify the factual correctness of semantically constrained text generation tasks such as document summarization. While the field of factuality evaluation is growing fast, we don't have well-defined criteria for measuring the effectiveness, generalizability, reliability, or sensitivity of the factuality metrics. Focusing on these aspects, in this paper, we introduce a meta-evaluation framework for evaluating factual consistency metrics. We introduce five necessary, common-sense conditions for effective factuality metrics and experiment with nine recent factuality metrics using synthetic and human-labeled factuality data from short news, long news and dialogue summarization domains. Our framework enables assessing the efficiency of any new factual consistency metric on a variety of dimensions over multiple summarization domains and can be easily extended with new meta-evaluation criteria. We also present our conclusions towards standardizing the factuality evaluation metrics.
Existing language models excel at writing from scratch, but many real-world scenarios require rewriting an existing document to fit a set of constraints. Although sentence-level rewriting has been fairly well-studied, little work has addressed the challenge of rewriting an entire document coherently. In this work, we introduce the task of document-level targeted content transfer and address it in the recipe domain, with a recipe as the document and a dietary restriction (such as vegan or dairy-free) as the targeted constraint. We propose a novel model for this task based on the generative pre-trained language model (GPT-2) and train on a large number of roughly-aligned recipe pairs (https://github.com/microsoft/document-level-targeted-content-transfer). Both automatic and human evaluations show that our model out-performs existing methods by generating coherent and diverse rewrites that obey the constraint while remaining close to the original document. Finally, we analyze our model's rewrites to assess progress toward the goal of making language generation more attuned to constraints that are substantive rather than stylistic.