Computation and Language (cs.CL)

Abstract: Mathematics is a highly specialized domain with its own unique set of challenges that has seen limited study in natural language processing. However, mathematics is used in a wide variety of fields and multidisciplinary research in many different domains often relies on an understanding of mathematical concepts. To aid researchers coming from other fields, we develop a prototype system for searching for and defining mathematical concepts in context, focusing on the field of category theory. This system, Parmesan, depends on natural language processing components including concept extraction, relation extraction, definition extraction, and entity linking. In developing this system, we show that existing techniques cannot be applied directly to the category theory domain, and suggest hybrid techniques that do perform well, though we expect the system to evolve over time. We also provide two cleaned mathematical corpora that power the prototype system, which are based on journal articles and wiki pages, respectively. The corpora have been annotated with dependency trees, lemmas, and part-of-speech tags.

Abstract: Answer selection in open-domain dialogues aims to select an accurate answer from candidates. Recent success of answer selection models hinges on training with large amounts of labeled data. However, collecting large-scale labeled data is labor-intensive and time-consuming. In this paper, we introduce the predicted intent labels to calibrate answer labels in a self-training paradigm. Specifically, we propose the intent-calibrated self-training (ICAST) to improve the quality of pseudo answer labels through the intent-calibrated answer selection paradigm, in which we employ pseudo intent labels to help improve pseudo answer labels. We carry out extensive experiments on two benchmark datasets with open-domain dialogues. The experimental results show that ICAST outperforms baselines consistently with 1%, 5% and 10% labeled data. Specifically, it improves 2.06% and 1.00% of F1 score on the two datasets, compared with the strongest baseline with only 5% labeled data.

Abstract: Although the NLP community has adopted central differential privacy as a go-to framework for privacy-preserving model training or data sharing, the choice and interpretation of the key parameter, privacy budget $\varepsilon$ that governs the strength of privacy protection, remains largely arbitrary. We argue that determining the $\varepsilon$ value should not be solely in the hands of researchers or system developers, but must also take into account the actual people who share their potentially sensitive data. In other words: Would you share your instant messages for $\varepsilon$ of 10? We address this research gap by designing, implementing, and conducting a behavioral experiment (311 lay participants) to study the behavior of people in uncertain decision-making situations with respect to privacy-threatening situations. Framing the risk perception in terms of two realistic NLP scenarios and using a vignette behavioral study help us determine what $\varepsilon$ thresholds would lead lay people to be willing to share sensitive textual data - to our knowledge, the first study of its kind.

Abstract: A wide variety of natural language tasks are currently being addressed with large-scale language models (LLMs). These models are usually trained with a very large amount of unsupervised text data and adapted to perform a downstream natural language task using methods like fine-tuning, calibration or in-context learning. In this work, we propose an approach to adapt the prior class distribution to perform text classification tasks without the need for labelled samples and only few in-domain sample queries. The proposed approach treats the LLM as a black box, adding a stage where the model posteriors are calibrated to the task. Results show that these methods outperform the un-adapted model for different number of training shots in the prompt and a previous approach were calibration is performed without using any adaptation data.

Abstract: Dialog policies, which determine a system's action based on the current state at each dialog turn, are crucial to the success of the dialog. In recent years, reinforcement learning (RL) has emerged as a promising option for dialog policy learning (DPL). In RL-based DPL, dialog policies are updated according to rewards. The manual construction of fine-grained rewards, such as state-action-based ones, to effectively guide the dialog policy is challenging in multi-domain task-oriented dialog scenarios with numerous state-action pair combinations. One way to estimate rewards from collected data is to train the reward estimator and dialog policy simultaneously using adversarial learning (AL). Although this method has demonstrated superior performance experimentally, it is fraught with the inherent problems of AL, such as mode collapse. This paper first identifies the role of AL in DPL through detailed analyses of the objective functions of dialog policy and reward estimator. Next, based on these analyses, we propose a method that eliminates AL from reward estimation and DPL while retaining its advantages. We evaluate our method using MultiWOZ, a multi-domain task-oriented dialog corpus.

Abstract: Larger language models, such as GPT-3, have shown to be excellent in many tasks. However, we demonstrate that out-of-ordinary questions can throw the model off guard. This work focuses on finding answers to negated complementary questions in commonsense scenarios. We illustrate how such questions adversely affect the model responses. We propose a model-agnostic methodology to improve the performance in negated complementary scenarios. Our method outperforms few-shot generation from GPT-3 (by more than 11 points) and, more importantly, highlights the significance of studying the response of large language models in negated complementary questions. The code, data, and experiments are available under: https://github.com/navidre/negated_complementary_commonsense.

Abstract: The generations of large language models are commonly controlled through prompting techniques, where a user's query to the model is prefixed with a prompt that aims to guide the model's behaviour on the query. The prompts used by companies to guide their models are often treated as secrets, to be hidden from the user making the query. They have even been treated as commodities to be bought and sold. However, there has been anecdotal evidence showing that the prompts can be extracted by a user even when they are kept secret. In this paper, we present a framework for systematically measuring the success of prompt extraction attacks. In experiments with multiple sources of prompts and multiple underlying language models, we find that simple text-based attacks can in fact reveal prompts with high probability.

Abstract: Existing evaluation metrics for natural language generation (NLG) tasks face the challenges on generalization ability and interpretability. Specifically, most of the well-performed metrics are required to train on evaluation datasets of specific NLG tasks and evaluation dimensions, which may cause over-fitting to task-specific datasets. Furthermore, existing metrics only provide an evaluation score for each dimension without revealing the evidence to interpret how this score is obtained. To deal with these challenges, we propose a simple yet effective metric called DecompEval. This metric formulates NLG evaluation as an instruction-style question answering task and utilizes instruction-tuned pre-trained language models (PLMs) without training on evaluation datasets, aiming to enhance the generalization ability. To make the evaluation process more interpretable, we decompose our devised instruction-style question about the quality of generated texts into the subquestions that measure the quality of each sentence. The subquestions with their answers generated by PLMs are then recomposed as evidence to obtain the evaluation result. Experimental results show that DecompEval achieves state-of-the-art performance in untrained metrics for evaluating text summarization and dialogue generation, which also exhibits strong dimension-level / task-level generalization ability and interpretability.

Abstract: Before deploying a language model (LM) within a given domain, it is important to measure its tendency to generate factually incorrect information in that domain. Existing factual generation evaluation methods focus on facts sampled from the LM itself, and thus do not control the set of evaluated facts and might under-represent rare and unlikely facts. We propose FACTOR: Factual Assessment via Corpus TransfORmation, a scalable approach for evaluating LM factuality. FACTOR automatically transforms a factual corpus of interest into a benchmark evaluating an LM's propensity to generate true facts from the corpus vs. similar but incorrect statements. We use our framework to create two benchmarks: Wiki-FACTOR and News-FACTOR. We show that: (i) our benchmark scores increase with model size and improve when the LM is augmented with retrieval; (ii) benchmark score correlates with perplexity, but the two metrics do not always agree on model ranking; and (iii) when perplexity and benchmark score disagree, the latter better reflects factuality in open-ended generation, as measured by human annotators. We make our data and code publicly available in https://github.com/AI21Labs/factor.

Abstract: We propose the In-context Autoencoder (ICAE) for context compression in a large language model (LLM). The ICAE has two modules: a learnable encoder adapted with LoRA from an LLM for compressing a long context into a limited number of memory slots, and a fixed decoder which is the target LLM that can condition on the memory slots for various purposes. We first pretrain the ICAE using both autoencoding and language modeling objectives on massive text data, enabling it to generate memory slots that accurately and comprehensively represent the original context. Then, we fine-tune the pretrained ICAE on a small amount of instruct data to enhance its interaction with various prompts for producing desirable responses. Our experimental results demonstrate that the ICAE learned with our proposed pretraining and fine-tuning paradigm can effectively produce memory slots with $4\times$ context compression, which can be well conditioned on by the target LLM to respond to various prompts. The promising results demonstrate significant implications of the ICAE for its novel approach to the long context problem and its potential to reduce computation and memory overheads for LLM inference in practice, suggesting further research effort in context management for an LLM. Our code and data will be released shortly.