Instruction following is a key capability for LLMs. However, recent studies have shown that LLMs often struggle with instructions containing multiple constraints (e.g. a request to create a social media post “in a funny tone” with “no hashtag”). Despite this, most evaluations focus solely on synthetic data. To address this, we introduce RealInstruct, the first benchmark designed to evaluate LLMs’ ability to follow real-world multi-constrained instructions by leveraging queries real users asked AI assistants. We also investigate model-based evaluation as a cost-effective alternative to human annotation for this task. Our findings reveal that even the proprietary GPT-4 model fails to meet at least one constraint on over 21% of instructions, highlighting the limitations of state-of-the-art models. To address the performance gap between open-source and proprietary models, we propose the Decompose, Critique and Refine (DeCRIM) self-correction pipeline, which enhances LLMs’ ability to follow constraints. DeCRIM works by decomposing the original instruction into a list of constraints and using a Critic model to decide when and where the LLM’s response needs refinement. Our results show that DeCRIM improves Mistral’s performance by 7.3% on RealInstruct and 8.0% on IFEval even with weak feedback. Moreover, we demonstrate that with strong feedback, open-source LLMs with DeCRIM can outperform GPT-4 on both benchmarks.

Creating children’s stories through text generation is a creative task that requires stories to be both entertaining and suitable for young audiences. However, since current story generation systems often rely on pre-trained language models fine-tuned with limited story data, they may not always prioritize child-friendliness. This can lead to the unintended generation of stories containing problematic elements such as violence, profanity, and biases. Regrettably, despite the significance of these concerns, there is a lack of clear guidelines and benchmark datasets for ensuring content safety for children. In this paper, we introduce a taxonomy specifically tailored to assess content safety in text, with a strong emphasis on children’s well-being. We present PG-Story, a dataset that includes detailed annotations for both sentence-level and discourse-level safety. We demonstrate the potential of identifying unsafe content through self-diagnosis and employing controllable generation techniques during the decoding phase to minimize unsafe elements in generated stories.

Automatic melody-to-lyric generation is a task in which song lyrics are generated to go with a given melody. It is of significant practical interest and more challenging than unconstrained lyric generation as the music imposes additional constraints onto the lyrics. The training data is limited as most songs are copyrighted, resulting in models that underfit the complicated cross-modal relationship between melody and lyrics. In this work, we propose a method for generating high-quality lyrics without training on any aligned melody-lyric data. Specifically, we design a hierarchical lyric generation framework that first generates a song outline and second the complete lyrics. The framework enables disentanglement of training (based purely on text) from inference (melody-guided text generation) to circumvent the shortage of parallel data. We leverage the segmentation and rhythm alignment between melody and lyrics to compile the given melody into decoding constraints as guidance during inference. The two-step hierarchical design also enables content control via the lyric outline, a much-desired feature for democratizing collaborative song creation. Experimental results show that our model can generate high-quality lyrics that are more on-topic, singable, intelligible, and coherent than strong baselines, for example SongMASS, a SOTA model trained on a parallel dataset, with a 24% relative overall quality improvement based on human ratings. Our code is available at https://github.com/amazon-science/unsupervised-melody-to-lyrics-generation.

The tasks of humor understanding and generation are challenging and subjective even for humans, requiring commonsense and real-world knowledge to master. Puns, in particular, add the challenge of fusing that knowledge with the ability to interpret lexical-semantic ambiguity. In this paper, we present the ExPUNations (ExPUN) dataset, in which we augment an existing dataset of puns with detailed crowdsourced annotations of keywords denoting the most distinctive words that make the text funny, pun explanations describing why the text is funny, and fine-grained funniness ratings. This is the first humor dataset with such extensive and fine-grained annotations specifically for puns. Based on these annotations, we propose two tasks: explanation generation to aid with pun classification and keyword-conditioned pun generation, to challenge the current state-of-the-art natural language understanding and generation models’ ability to understand and generate humor. We showcase that the annotated keywords we collect are helpful for generating better novel humorous texts in human evaluation, and that our natural language explanations can be leveraged to improve both the accuracy and robustness of humor classifiers.

Previous work on pun generation commonly begins with a given pun word (a pair of homophones for heterographic pun generation and a polyseme for homographic pun generation) and seeks to generate an appropriate pun. While this may enable efficient pun generation, we believe that a pun is most entertaining if it fits appropriately within a given context, e.g., a given situation or dialogue. In this work, we propose a new task, context-situated pun generation, where a specific context represented by a set of keywords is provided, and the task is to first identify suitable pun words that are appropriate for the context, then generate puns based on the context keywords and the identified pun words. We collect a new dataset, CUP (Context-sitUated Pun), containing 4.5k tuples of context words and pun pairs. Based on the new data and setup, we propose a pipeline system for context-situated pun generation, including a pun word retrieval module that identifies suitable pun words for a given context, and a pun generation module that generates puns from context keywords and pun words. Human evaluation shows that 69% of our top retrieved pun words can be used to generate context-situated puns, and our generation module yields successful puns 31% of the time given a plausible tuple of context words and pun pair, almost tripling the yield of a state-of-the-art pun generation model. With an end-to-end evaluation, our pipeline system with the top-1 retrieved pun pair for a given context can generate successful puns 40% of the time, better than all other modeling variations but 32% lower than the human success rate. This highlights the difficulty of the task, and encourages more research in this direction.

Natural Language Generation (NLG) for task-oriented dialogue systems focuses on communicating specific content accurately, fluently, and coherently. While these attributes are crucial for a successful dialogue, it is also desirable to simultaneously accomplish specific stylistic goals, such as response length, point-of-view, descriptiveness, sentiment, formality, and empathy. In this work, we focus on stylistic control and evaluation for schema-guided NLG, with joint goals of achieving both semantic and stylistic control. We experiment in detail with various controlled generation methods for large pretrained language models: specifically, conditional training, guided fine-tuning, and guided decoding. We discuss their advantages and limitations, and evaluate them with a broad range of automatic and human evaluation metrics. Our results show that while high style accuracy and semantic correctness are easier to achieve for more lexically-defined styles with conditional training, stylistic control is also achievable for more semantically complex styles using discriminator-based guided decoding methods. The results also suggest that methods that are more scalable (with less hyper-parameters tuning) and that disentangle context generation and stylistic variations are more effective at achieving semantic correctness and style accuracy.

Neural network based approaches to data-to-text natural language generation (NLG) have gained popularity in recent years, with the goal of generating a natural language prompt that accurately realizes an input meaning representation. To facilitate the training of neural network models, researchers created large datasets of paired utterances and their meaning representations. However, the creation of such datasets is an arduous task and they mostly consist of simple meaning representations composed of slot and value tokens to be realized. These representations do not include any contextual information that an NLG system can use when trying to generalize, such as domain information and descriptions of slots and values. In this paper, we present the novel task of Schema-Guided Natural Language Generation (SG-NLG). Here, the goal is still to generate a natural language prompt, but in SG-NLG, the input MRs are paired with rich schemata providing contextual information. To generate a dataset for SG-NLG we re-purpose an existing dataset for another task: dialog state tracking, which includes a large and rich schema spanning multiple different attributes, including information about the domain, user intent, and slot descriptions. We train different state-of-the-art models for neural natural language generation on this dataset and show that in many cases, including rich schema information allows our models to produce higher quality outputs both in terms of semantics and diversity. We also conduct experiments comparing model performance on seen versus unseen domains, and present a human evaluation demonstrating high ratings for overall output quality.

Natural language generators (NLGs) for task-oriented dialogue typically take a meaning representation (MR) as input, and are trained end-to-end with a corpus of MR/utterance pairs, where the MRs cover a specific set of dialogue acts and domain attributes. Creation of such datasets is labor intensive and time consuming. Therefore, dialogue systems for new domain ontologies would benefit from using data for pre-existing ontologies. Here we explore, for the first time, whether it is possible to train an NLG for a new larger ontology using existing training sets for the restaurant domain, where each set is based on a different ontology. We create a new, larger combined ontology, and then train an NLG to produce utterances covering it. For example, if one dataset has attributes for family friendly and rating information, and the other has attributes for decor and service, our aim is an NLG for the combined ontology that can produce utterances that realize values for family friendly, rating, decor and service. Initial experiments with a baseline neural sequence-to-sequence model show that this task is surprisingly challenging. We then develop a novel self-training method that identifies (errorful) model outputs, automatically constructs a corrected MR input to form a new (MR, utterance) training pair, and then repeatedly adds these new instances back into the training data. We then test the resulting model on a new test set. The result is a self-trained model whose performance is an absolute 75.4% improvement over the baseline model. We also report a human qualitative evaluation of the final model showing that it achieves high naturalness, semantic coherence and grammaticality.

Neural natural language generation (NNLG) from structured meaning representations has become increasingly popular in recent years. While we have seen progress with generating syntactically correct utterances that preserve semantics, various shortcomings of NNLG systems are clear: new tasks require new training data which is not available or straightforward to acquire, and model outputs are simple and may be dull and repetitive. This paper addresses these two critical challenges in NNLG by: (1) scalably (and at no cost) creating training datasets of parallel meaning representations and reference texts with rich style markup by using data from freely available and naturally descriptive user reviews, and (2) systematically exploring how the style markup enables joint control of semantic and stylistic aspects of neural model output. We present YelpNLG, a corpus of 300,000 rich, parallel meaning representations and highly stylistically varied reference texts spanning different restaurant attributes, and describe a novel methodology that can be scalably reused to generate NLG datasets for other domains. The experiments show that the models control important aspects, including lexical choice of adjectives, output length, and sentiment, allowing the models to successfully hit multiple style targets without sacrificing semantics.

Neural generation methods for task-oriented dialogue typically generate from a meaning representation that is populated using a database of domain information, such as a table of data describing a restaurant. While earlier work focused solely on the semantic fidelity of outputs, recent work has started to explore methods for controlling the style of the generated text while simultaneously achieving semantic accuracy. Here we experiment with two stylistic benchmark tasks, generating language that exhibits variation in personality, and generating discourse contrast. We report a huge performance improvement in both stylistic control and semantic accuracy over the state of the art on both of these benchmarks. We test several different models and show that putting stylistic conditioning in the decoder and eliminating the semantic re-ranker used in earlier models results in more than 15 points higher BLEU for Personality, with a reduction of semantic error to near zero. We also report an improvement from .75 to .81 in controlling contrast and a reduction in semantic error from 16% to 2%.

Natural language generators for task-oriented dialogue must effectively realize system dialogue actions and their associated semantics. In many applications, it is also desirable for generators to control the style of an utterance. To date, work on task-oriented neural generation has primarily focused on semantic fidelity rather than achieving stylistic goals, while work on style has been done in contexts where it is difficult to measure content preservation. Here we present three different sequence-to-sequence models and carefully test how well they disentangle content and style. We use a statistical generator, Personage, to synthesize a new corpus of over 88,000 restaurant domain utterances whose style varies according to models of personality, giving us total control over both the semantic content and the stylistic variation in the training data. We then vary the amount of explicit stylistic supervision given to the three models. We show that our most explicit model can simultaneously achieve high fidelity to both semantic and stylistic goals: this model adds a context vector of 36 stylistic parameters as input to the hidden state of the encoder at each time step, showing the benefits of explicit stylistic supervision, even when the amount of training data is large.

Responses in task-oriented dialogue systems often realize multiple propositions whose ultimate form depends on the use of sentence planning and discourse structuring operations. For example a recommendation may consist of an explicitly evaluative utterance e.g. Chanpen Thai is the best option, along with content related by the justification discourse relation, e.g. It has great food and service, that combines multiple propositions into a single phrase. While neural generation methods integrate sentence planning and surface realization in one end-to-end learning framework, previous work has not shown that neural generators can: (1) perform common sentence planning and discourse structuring operations; (2) make decisions as to whether to realize content in a single sentence or over multiple sentences; (3) generalize sentence planning and discourse relation operations beyond what was seen in training. We systematically create large training corpora that exhibit particular sentence planning operations and then test neural models to see what they learn. We compare models without explicit latent variables for sentence planning with ones that provide explicit supervision during training. We show that only the models with additional supervision can reproduce sentence planning and discourse operations and generalize to situations unseen in training.

Many of the creative and figurative elements that make language exciting are lost in translation in current natural language generation engines. In this paper, we explore a method to harvest templates from positive and negative reviews in the restaurant domain, with the goal of vastly expanding the types of stylistic variation available to the natural language generator. We learn hyperbolic adjective patterns that are representative of the strongly-valenced expressive language commonly used in either positive or negative reviews. We then identify and delexicalize entities, and use heuristics to extract generation templates from review sentences. We evaluate the learned templates against more traditional review templates, using subjective measures of convincingness, interestingness, and naturalness. Our results show that the learned templates score highly on these measures. Finally, we analyze the linguistic categories that characterize the learned positive and negative templates. We plan to use the learned templates to improve the conversational style of dialogue systems in the restaurant domain.

Effective models of social dialog must understand a broad range of rhetorical and figurative devices. Rhetorical questions (RQs) are a type of figurative language whose aim is to achieve a pragmatic goal, such as structuring an argument, being persuasive, emphasizing a point, or being ironic. While there are computational models for other forms of figurative language, rhetorical questions have received little attention to date. We expand a small dataset from previous work, presenting a corpus of 10,270 RQs from debate forums and Twitter that represent different discourse functions. We show that we can clearly distinguish between RQs and sincere questions (0.76 F1). We then show that RQs can be used both sarcastically and non-sarcastically, observing that non-sarcastic (other) uses of RQs are frequently argumentative in forums, and persuasive in tweets. We present experiments to distinguish between these uses of RQs using SVM and LSTM models that represent linguistic features and post-level context, achieving results as high as 0.76 F1 for “sarcastic” and 0.77 F1 for “other” in forums, and 0.83 F1 for both “sarcastic” and “other” in tweets. We supplement our quantitative experiments with an in-depth characterization of the linguistic variation in RQs.

Informal first-person narratives are a unique resource for computational models of everyday events and people’s affective reactions to them. People blogging about their day tend not to explicitly say I am happy. Instead they describe situations from which other humans can readily infer their affective reactions. However current sentiment dictionaries are missing much of the information needed to make similar inferences. We build on recent work that models affect in terms of lexical predicate functions and affect on the predicate’s arguments. We present a method to learn proxies for these functions from first-person narratives. We construct a novel fine-grained test set, and show that the patterns we learn improve our ability to predict first-person affective reactions to everyday events, from a Stanford sentiment baseline of .67F to .75F.