Large language models (LLMs) can solve problems step-by-step.While this chain-of-thought (CoT) reasoning boosts LLMs’ performance, it is unclear if LLMs know when to use CoT and whether those CoT are always necessary to answer the question. This paper shows that LLMs tend to generate redundant calculations and reasoning on a manually constructed math QA dataset, GSM8K-Zero.GSM8K-Zero is constructed such that the questions can be answered without any calculations, but LLMs, including Llama-2 models and Claude-2, tend to generate lengthy and unnecessary calculations to answer the questions.We also conduct experiments to explain why LLMs generate redundant calculations and reasonings.

In this paper, we explore the following question: Are synonym substitution attacks really synonym substitution attacks (SSAs)?We approach this question by examining how SSAs replace words in the original sentence and show that there are still unresolved obstacles that make current SSAs generate invalid adversarial samples. We reveal that four widely used word substitution methods generate a large fraction of invalid substitution words that are ungrammatical or do not preserve the original sentence’s semantics. Next, we show that the semantic and grammatical constraints used in SSAs for detecting invalid word replacements are highly insufficient in detecting invalid adversarial samples.

Using large language models (LLMs) to evaluate text quality has recently gained popularity. Some existing prior works explore the idea of using LLMs for evaluation, while they differ in some details of the evaluation process. In this paper, we analyze *LLM evaluation* and *G-Eval*, and we discuss how those details in the evaluation process change how well the ratings given by LLMs correlate with human ratings. We find that the auto Chain-of-Thought (CoT) used in G-Eval does not always make G-Eval more aligned with human ratings. We also show that forcing the LLM to output only a numeric rating, as in G-Eval, is suboptimal. Last, we reveal that asking the LLM to explain its own ratings consistently improves the correlation between the ChatGPT and human ratings and pushes state-of-the-art (SoTA) correlations on two meta-evaluation datasets.

With the power of large pretrained language models, various research works have integrated knowledge into dialogue systems. The traditional techniques treat knowledge as part of the input sequence for the dialogue system, prepending a set of knowledge statements in front of dialogue history. However, such a mechanism forces knowledge sets to be concatenated in an ordered manner, making models implicitly pay imbalanced attention to the sets during training. In this paper, we first investigate how the order of the knowledge set can influence autoregressive dialogue systems’ responses. We conduct experiments on two commonly used dialogue datasets with two types of transformer-based models and find that models view the input knowledge unequally. To this end, we propose a simple and novel technique to alleviate the order effect by modifying the position embeddings of knowledge input in these models. With the proposed position embedding method, the experimental results show that each knowledge statement is uniformly considered to generate responses.

Spoken language understanding (SLU) tasks have been studied for many decades in the speech research community, but have not received as much attention as lower-level tasks like speech and speaker recognition. In this work, we introduce several new annotated SLU benchmark tasks based on freely available speech data, which complement existing benchmarks and address gaps in the SLU evaluation landscape. We contribute four tasks: question answering and summarization involve inference over longer speech sequences; named entity localization addresses the speech-specific task of locating the targeted content in the signal; dialog act classification identifies the function of a given speech utterance. In order to facilitate the development of SLU models that leverage the success of pre-trained speech representations, we will release a new benchmark suite, including for each task (i) curated annotations for a relatively small fine-tuning set, (ii) reproducible pipeline (speech recognizer + text model) and end-to-end baseline models and evaluation metrics, (iii) baseline model performance in various types of systems for easy comparisons. We present the details of data collection and annotation and the performance of the baseline models. We also analyze the sensitivity of pipeline models’ performance to the speech recognition accuracy, using more than 20 publicly availablespeech recognition models.

Recent studies find existing self-supervised speech encoders contain primarily acoustic rather than semantic information. As a result, pipelined supervised automatic speech recognition (ASR) to large language model (LLM) systems achieve state-of-the-art results on semantic spoken language tasks by utilizing rich semantic representations from the LLM. These systems come at the cost of labeled audio transcriptions, which is expensive and time-consuming to obtain. We propose a task-agnostic unsupervised way of incorporating semantic information from LLMs into self-supervised speech encoders without labeled audio transcriptions. By introducing semantics, we improve existing speech encoder spoken language understanding (SLU) performance by over 5% on intent classification (IC), with modest gains in named entity resolution (NER) and slot filling (SF), and spoken question answering (SQA) FF1 score by over 2%. Our approach, which uses no ASR data, achieves similar performance as methods trained on over 100 hours of labeled audio transcripts, demonstrating the feasibility of unsupervised semantic augmentations to existing speech encoders.

Human evaluation is indispensable and inevitable for assessing the quality of texts generated by machine learning models or written by humans. However, human evaluation is very difficult to reproduce and its quality is notoriously unstable, hindering fair comparisons among different natural language processing (NLP) models and algorithms. Recently, large language models (LLMs) have demonstrated exceptional performance on unseen tasks when only the task instructions are provided. In this paper, we explore if such an ability of the LLMs can be used as an alternative to human evaluation. We present the LLMs with the exact same instructions, samples to be evaluated, and questions used to conduct human evaluation, and then ask the LLMs to generate responses to those questions; we dub this LLM evaluation. We use human evaluation and LLM evaluation to evaluate the texts in two NLP tasks: open-ended story generation and adversarial attacks. We show that the result of LLM evaluation is consistent with the results obtained by expert human evaluation: the texts rated higher by human experts are also rated higher by the LLMs.We also find that the results of LLM evaluation are stable over different formatting of the task instructions and the sampling algorithm used to generate the answer. We are the first to show the potential of using LLMs to assess the quality of texts and discuss the limitations and ethical considerations of LLM evaluation.

Pre-trained language models (PLMs) are language models that are pre-trained on large-scaled corpora in a self-supervised fashion. These PLMs have fundamentally changed the natural language processing community in the past few years. In this tutorial, we aim to provide a broad and comprehensive introduction from two perspectives: why those PLMs work, and how to use them in NLP tasks. The first part of the tutorial shows some insightful analysis on PLMs that partially explain their exceptional downstream performance. The second part first focuses on emerging pre-training methods that enable PLMs to perform diverse downstream tasks and then illustrates how one can apply those PLMs to downstream tasks under different circumstances. These circumstances include fine-tuning PLMs when under data scarcity, and using PLMs with parameter efficiency. We believe that attendees of different backgrounds would find this tutorial informative and useful.

Transfer learning has proven to be crucial in advancing the state of speech and natural language processing research in recent years. In speech, a model pre-trained by self-supervised learning transfers remarkably well on multiple tasks. However, the lack of a consistent evaluation methodology is limiting towards a holistic understanding of the efficacy of such models. SUPERB was a step towards introducing a common benchmark to evaluate pre-trained models across various speech tasks. In this paper, we introduce SUPERB-SG, a new benchmark focusing on evaluating the semantic and generative capabilities of pre-trained models by increasing task diversity and difficulty over SUPERB. We use a lightweight methodology to test the robustness of representations learned by pre-trained models under shifts in data domain and quality across different types of tasks. It entails freezing pre-trained model parameters, only using simple task-specific trainable heads. The goal is to be inclusive of all researchers, and encourage efficient use of computational resources. We also show that the task diversity of SUPERB-SG coupled with limited task supervision is an effective recipe for evaluating the generalizability of model representation.

Transformer-based models are widely used in natural language understanding (NLU) tasks, and multimodal transformers have been effective in visual-language tasks. This study explores distilling visual information from pretrained multimodal transformers to pretrained language encoders. Our framework is inspired by cross-modal encoders’ success in visual-language tasks while we alter the learning objective to cater to the language-heavy characteristics of NLU. After training with a small number of extra adapting steps and finetuned, the proposed XDBERT (cross-modal distilled BERT) outperforms pretrained-BERT in general language understanding evaluation (GLUE), situations with adversarial generations (SWAG) benchmarks, and readability benchmarks. We analyze the performance of XDBERT on GLUE to show that the improvement is likely visually grounded.

Transformer-based pre-trained models with millions of parameters require large storage. Recent approaches tackle this shortcoming by training adapters, but these approaches still require a relatively large number of parameters. In this study, AdapterBias, a surprisingly simple yet effective adapter architecture, is proposed. AdapterBias adds a token-dependent shift to the hidden output of transformer layers to adapt to downstream tasks with only a vector and a linear layer. Extensive experiments are conducted to demonstrate the effectiveness of AdapterBias. The experiments show that our proposed method can dramatically reduce the trainable parameters compared to the previous works with a minimal decrease in task performances compared with fine-tuned pre-trained models. We further find that AdapterBias automatically learns to assign more significant representation shifts to the tokens related to the task in consideration.

Deep learning has been the mainstream technique in the natural language processing (NLP) area. However, deep learning requires many labeled data and is less generalizable across domains. Meta-learning is an arising field in machine learning. It studies approaches to learning better learning algorithms and aims to improve algorithms in various aspects, including data efficiency and generalizability. The efficacy of meta-learning has been shown in many NLP tasks, but there is no systematic survey of these approaches in NLP, which hinders more researchers from joining the field. Our goal with this survey paper is to offer researchers pointers to relevant meta-learning works in NLP and attract more attention from the NLP community to drive future innovation. This paper first introduces the general concepts of meta-learning and the common approaches. Then we summarize task construction settings, applications of meta-learning for various NLP problems and review the development of meta-learning in the NLP community.

There is a trend in the machine learning community to adopt self-supervised approaches to pre-train deep networks. Self-supervised representation learning (SSL) utilizes proxy supervised learning tasks, for example, distinguishing parts of the input signal from distractors, or generating masked input segments conditioned on the unmasked ones, to obtain training data from unlabeled corpora. BERT and GPT in NLP and SimCLR and BYOL in CV are famous examples in this direction. These approaches make it possible to use a tremendous amount of unlabeled data available on the web to train large networks and solve complicated tasks. Thus, SSL has the potential to scale up current machine learning technologies, especially for low-resourced, under-represented use cases, and democratize the technologies. Recently self-supervised approaches for speech processing are also gaining popularity. There are several workshops in relevant topics hosted at ICML 2020 (https://icml-sas.gitlab.io/), NeurIPS 2020 (https://neurips-sas-2020.github.io/), and AAAI 2022 (https://aaai-sas-2022.github.io/). However, there is no previous tutorial about a similar topic based on the authors’ best knowledge. Due to the growing popularity of SSL, and the shared mission of the areas in bringing speech and language technologies to more use cases with better quality and scaling the technologies for under-represented languages, we propose this tutorial to systematically survey the latest SSL techniques, tools, datasets, and performance achievement in speech processing. The proposed tutorial is highly relevant to the special theme of ACL about language diversity. One of the main focuses of the tutorial is leveraging SSL to reduce the dependence of speech technologies on labeled data, and to scale up the technologies especially for under-represented languages and use cases.

Simultaneous machine translation (SimulMT) speeds up the translation process by starting to translate before the source sentence is completely available. It is difficult due to limited context and word order difference between languages. Existing methods increase latency or introduce adaptive read-write policies for SimulMT models to handle local reordering and improve translation quality. However, the long-distance reordering would make the SimulMT models learn translation mistakenly. Specifically, the model may be forced to predict target tokens when the corresponding source tokens have not been read. This leads to aggressive anticipation during inference, resulting in the hallucination phenomenon. To mitigate this problem, we propose a new framework that decompose the translation process into the monotonic translation step and the reordering step, and we model the latter by the auxiliary sorting network (ASN). The ASN rearranges the hidden states to match the order in the target language, so that the SimulMT model could learn to translate more reasonably. The entire model is optimized end-to-end and does not rely on external aligners or data. During inference, ASN is removed to achieve streaming. Experiments show the proposed framework could outperform previous methods with less latency.

In this paper, we study the possibility of unsupervised Multiple Choices Question Answering (MCQA). From very basic knowledge, the MCQA model knows that some choices have higher probabilities of being correct than others. The information, though very noisy, guides the training of an MCQA model. The proposed method is shown to outperform the baseline approaches on RACE and is even comparable with some supervised learning approaches on MC500.

Most chatbot literature that focuses on improving the fluency and coherence of a chatbot, is dedicated to making chatbots more human-like. However, very little work delves into what really separates humans from chatbots – humans intrinsically understand the effect their responses have on the interlocutor and often respond with an intention such as proposing an optimistic view to make the interlocutor feel better. This paper proposes an innovative framework to train chatbots to possess human-like intentions. Our framework includes a guiding chatbot and an interlocutor model that plays the role of humans. The guiding chatbot is assigned an intention and learns to induce the interlocutor to reply with responses matching the intention, for example, long responses, joyful responses, responses with specific words, etc. We examined our framework using three experimental setups and evaluated the guiding chatbot with four different metrics to demonstrate flexibility and performance advantages. Additionally, we performed trials with human interlocutors to substantiate the guiding chatbot’s effectiveness in influencing the responses of humans to a certain extent. Code will be made available to the public.

Automatic detection of toxic language plays an essential role in protecting social media users, especially minority groups, from verbal abuse. However, biases toward some attributes, including gender, race, and dialect, exist in most training datasets for toxicity detection. The biases make the learned models unfair and can even exacerbate the marginalization of people. Considering that current debiasing methods for general natural language understanding tasks cannot effectively mitigate the biases in the toxicity detectors, we propose to use invariant rationalization (InvRat), a game-theoretic framework consisting of a rationale generator and a predictor, to rule out the spurious correlation of certain syntactic patterns (e.g., identity mentions, dialect) to toxicity labels. We empirically show that our method yields lower false positive rate in both lexical and dialectal attributes than previous debiasing methods.

Deep learning based natural language processing (NLP) has become the mainstream of research in recent years and significantly outperforms conventional methods. However, deep learning models are notorious for being data and computation hungry. These downsides limit the application of such models from deployment to different domains, languages, countries, or styles, since collecting in-genre data and model training from scratch are costly. The long-tail nature of human language makes challenges even more significant. Meta-learning, or ‘Learning to Learn’, aims to learn better learning algorithms, including better parameter initialization, optimization strategy, network architecture, distance metrics, and beyond. Meta-learning has been shown to allow faster fine-tuning, converge to better performance, and achieve amazing results for few-shot learning in many applications. Meta-learning is one of the most important new techniques in machine learning in recent years. There is a related tutorial in ICML 2019 and a related course at Stanford, but most of the example applications given in these materials are about image processing. It is believed that meta-learning has great potential to be applied in NLP, and some works have been proposed with notable achievements in several relevant problems, e.g., relation extraction, machine translation, and dialogue generation and state tracking. However, it does not catch the same level of attention as in the image processing community. In the tutorial, we will first introduce Meta-learning approaches and the theory behind them, and then review the works of applying this technology to NLP problems. This tutorial intends to facilitate researchers in the NLP community to understand this new technology better and promote more research studies using this new technology.

This paper investigates whether the power of the models pre-trained on text data, such as BERT, can be transferred to general token sequence classification applications. To verify pre-trained models’ transferability, we test the pre-trained models on text classification tasks with meanings of tokens mismatches, and real-world non-text token sequence classification data, including amino acid, DNA, and music. We find that even on non-text data, the models pre-trained on text converge faster, perform better than the randomly initialized models, and only slightly worse than the models using task-specific knowledge. We also find that the representations of the text and non-text pre-trained models share non-trivial similarities.

Speech separation is a problem in the field of speech processing that has been studied in full swing recently. However, there has not been much work studying a multi-accent speech separation scenario. Unseen speakers with new accents and noise aroused the domain mismatch problem which cannot be easily solved by conventional joint training methods. Thus, we applied MAML and FOMAML to tackle this problem and obtained higher average Si-SNRi values than joint training on almost all the unseen accents. This proved that these two methods do have the ability to generate well-trained parameters for adapting to speech mixtures of new speakers and accents. Furthermore, we found out that FOMAML obtains similar performance compared to MAML while saving a lot of time.

Speech translation (ST) aims to learn transformations from speech in the source language to the text in the target language. Previous works show that multitask learning improves the ST performance, in which the recognition decoder generates the text of the source language, and the translation decoder obtains the final translations based on the output of the recognition decoder. Because whether the output of the recognition decoder has the correct semantics is more critical than its accuracy, we propose to improve the multitask ST model by utilizing word embedding as the intermediate.

While behaviors of pretrained language models (LMs) have been thoroughly examined, what happened during pretraining is rarely studied. We thus investigate the developmental process from a set of randomly initialized parameters to a totipotent language model, which we refer to as the embryology of a pretrained language model. Our results show that ALBERT learns to reconstruct and predict tokens of different parts of speech (POS) in different learning speeds during pretraining. We also find that linguistic knowledge and world knowledge do not generally improve as pretraining proceeds, nor do downstream tasks’ performance. These findings suggest that knowledge of a pretrained model varies during pretraining, and having more pretrain steps does not necessarily provide a model with more comprehensive knowledge. We provide source codes and pretrained models to reproduce our results at https://github.com/d223302/albert-embryology.

Pre-training Transformer from large-scale raw texts and fine-tuning on the desired task have achieved state-of-the-art results on diverse NLP tasks. However, it is unclear what the learned attention captures. The attention computed by attention heads seems not to match human intuitions about hierarchical structures. This paper proposes Tree Transformer, which adds an extra constraint to attention heads of the bidirectional Transformer encoder in order to encourage the attention heads to follow tree structures. The tree structures can be automatically induced from raw texts by our proposed “Constituent Attention” module, which is simply implemented by self-attention between two adjacent words. With the same training procedure identical to BERT, the experiments demonstrate the effectiveness of Tree Transformer in terms of inducing tree structures, better language modeling, and further learning more explainable attention scores.

Data-driven, knowledge-grounded neural conversation models are capable of generating more informative responses. However, these models have not yet demonstrated that they can zero-shot adapt to updated, unseen knowledge graphs. This paper proposes a new task about how to apply dynamic knowledge graphs in neural conversation model and presents a novel TV series conversation corpus (DyKgChat) for the task. Our new task and corpus aids in understanding the influence of dynamic knowledge graphs on responses generation. Also, we propose a preliminary model that selects an output from two networks at each time step: a sequence-to-sequence model (Seq2Seq) and a multi-hop reasoning model, in order to support dynamic knowledge graphs. To benchmark this new task and evaluate the capability of adaptation, we introduce several evaluation metrics and the experiments show that our proposed approach outperforms previous knowledge-grounded conversation models. The proposed corpus and model can motivate the future research directions.

Because it is not feasible to collect training data for every language, there is a growing interest in cross-lingual transfer learning. In this paper, we systematically explore zero-shot cross-lingual transfer learning on reading comprehension tasks with language representation model pre-trained on multi-lingual corpus. The experimental results show that with pre-trained language representation zero-shot learning is feasible, and translating the source data into the target language is not necessary and even degrades the performance. We further explore what does the model learn in zero-shot setting.

Paraphrase generation is an interesting and challenging NLP task which has numerous practical applications. In this paper, we analyze datasets commonly used for paraphrase generation research, and show that simply parroting input sentences surpasses state-of-the-art models in the literature when evaluated on standard metrics. Our findings illustrate that a model could be seemingly adept at generating paraphrases, despite only making trivial changes to the input sentence or even none at all.

Auto-encoders compress input data into a latent-space representation and reconstruct the original data from the representation. This latent representation is not easily interpreted by humans. In this paper, we propose training an auto-encoder that encodes input text into human-readable sentences, and unpaired abstractive summarization is thereby achieved. The auto-encoder is composed of a generator and a reconstructor. The generator encodes the input text into a shorter word sequence, and the reconstructor recovers the generator input from the generator output. To make the generator output human-readable, a discriminator restricts the output of the generator to resemble human-written sentences. By taking the generator output as the summary of the input text, abstractive summarization is achieved without document-summary pairs as training data. Promising results are shown on both English and Chinese corpora.

Although transfer learning has been shown to be successful for tasks like object and speech recognition, its applicability to question answering (QA) has yet to be well-studied. In this paper, we conduct extensive experiments to investigate the transferability of knowledge learned from a source QA dataset to a target dataset using two QA models. The performance of both models on a TOEFL listening comprehension test (Tseng et al., 2016) and MCTest (Richardson et al., 2013) is significantly improved via a simple transfer learning technique from MovieQA (Tapaswi et al., 2016). In particular, one of the models achieves the state-of-the-art on all target datasets; for the TOEFL listening comprehension test, it outperforms the previous best model by 7%. Finally, we show that transfer learning is helpful even in unsupervised scenarios when correct answers for target QA dataset examples are not available.

In this paper, we propose new methods to learn Chinese word representations. Chinese characters are composed of graphical components, which carry rich semantics. It is common for a Chinese learner to comprehend the meaning of a word from these graphical components. As a result, we propose models that enhance word representations by character glyphs. The character glyph features are directly learned from the bitmaps of characters by convolutional auto-encoder(convAE), and the glyph features improve Chinese word representations which are already enhanced by character embeddings. Another contribution in this paper is that we created several evaluation datasets in traditional Chinese and made them public.