This work studies mitigating fact-conflicting hallucinations for large language model (LLM) at inference time.Particularly, we propose a self-endorsement framework that leverages the fine-grained fact-level comparisons across multiple sampled responses.Compared with prior ensemble methods (e.g., self-consistency) that perform response-level selection, our approach can better alleviate hallucinations for knowledge-intensive tasks.Our approach can broadly benefit smaller and open-source LLMs as it mainly conducts simple content-based comparisons.Experiments on Biographies show that our method can effectively improve the factuality of generations with simple and intuitive prompts across different scales of LLMs.Besides, comprehensive analyses on TriviaQA and GSM8K demonstrate the potential of self-endorsement for broader application.

Despite showing impressive abilities, large language models (LLMs) often struggle with factual inaccuracies, i.e., ”hallucinations”, even when they hold relevant knowledge. To mitigate these hallucinations, current approaches typically necessitate high-quality human factuality annotations. In this work, we explore Self-Alignment for Factuality, where we leverage the self-evaluation capability of an LLM to provide training signals that steer the model towards factuality. Specifically, we incorporate Self-Eval, a self-evaluation component, to prompt an LLM to validate the factuality of its own generated responses solely based on its internal knowledge. Additionally, we design Self-Knowledge Tuning (SK-Tuning) to augment the LLM’s self-evaluation ability by improving the model’s confidence estimation and calibration. We then utilize these self-annotated responses to fine-tune the model via Direct Preference Optimization algorithm. We show that the proposed self-alignment approach substantially enhances factual accuracy over Llama family models across three key knowledge-intensive tasks on TruthfulQA and BioGEN.

Continuous learning from free-text human feedback, such as error corrections, new knowledge, or alternative responses, is essential for today’s chatbots and virtual assistants to stay up-to-date, engaging, and socially acceptable. However, for research on methods for learning from such data, annotated data is scarce. To address this, we examine the error and user response types of six popular dialogue datasets from various types, including MultiWoZ, PersonaChat, Wizards-of-Wikipedia, and others, to assess their extendibility with the needed annotations. For this corpus study, we manually annotate a subset of each dataset with error and user response types using an improved version of the Integrated Error Taxonomy and a newly proposed user response type taxonomy. We provide the resulting dataset (EURTAD) to the community. Our findings provide new insights into dataset composition, including error types, user response types, and the relations between them.

Idioms such as “call it a day” and “piece of cake,” are prevalent in natural language. How do Transformer language models process idioms? This study examines this question by analysing three models - BERT, Multilingual BERT, and DistilBERT. We compare the embeddings of idiomatic and literal expressions across all layers of the networks at both the sentence and word levels. Additionally, we investigate the attention directed from other sentence tokens towards a word within an idiom as opposed to in a literal context. Results indicate that while the three models exhibit slightly different internal mechanisms, they all represent idioms distinctively compared to literal language, with attention playing a critical role. These findings suggest that idioms are semantically and syntactically idiosyncratic, not only for humans but also for language models.

This paper describes Huawei Artificial Intelligence Application Research Center’s neural machine translation system (“BabelTar”). Our submission to the WMT22 biomedical translation shared task covers language directions between English and the other seven languages (French, German, Italian, Spanish, Portuguese, Russian, and Chinese). During the past four years, our participation in this domain-specific track has witnessed a paradigm shift of methodology from a purely data-driven focus to embracing diversified techniques, including pre-trained multilingual NMT models, homograph disambiguation, ensemble learning, and preprocessing methods. We illustrate practical insights and measured performance improvements relating to how we further improve our domain-specific NMT system.

In meta-learning, the knowledge learned from previous tasks is transferred to new ones, but this transfer only works if tasks are related. Sharing information between unrelated tasks might hurt performance, and it is unclear how to transfer knowledge across tasks that have a hierarchical structure. Our research extends a meta-learning model, MAML, by exploiting hierarchical task relationships. Our algorithm, TreeMAML, adapts the model to each task with a few gradient steps, but the adaptation follows the hierarchical tree structure: in each step, gradients are pooled across tasks clusters and subsequent steps follow down the tree. We also implement a clustering algorithm that generates the tasks tree without previous knowledge of the task structure, allowing us to make use of implicit relationships between the tasks. We show that TreeMAML successfully trains natural language processing models for cross-lingual Natural Language Inference by taking advantage of the language phylogenetic tree. This result is useful since most languages in the world are under-resourced and the improvement on cross-lingual transfer allows the internationalization of NLP models.

Natural conversations are filled with disfluencies. This study investigates if and how BERT understands disfluency with three experiments: (1) a behavioural study using a downstream task, (2) an analysis of sentence embeddings and (3) an analysis of the attention mechanism on disfluency. The behavioural study shows that without fine-tuning on disfluent data, BERT does not suffer significant performance loss when presented disfluent compared to fluent inputs (exp1). Analysis on sentence embeddings of disfluent and fluent sentence pairs reveals that the deeper the layer, the more similar their representation (exp2). This indicates that deep layers of BERT become relatively invariant to disfluency. We pinpoint attention as a potential mechanism that could explain this phenomenon (exp3). Overall, the study suggests that BERT has knowledge of disfluency structure. We emphasise the potential of using BERT to understand natural utterances without disfluency removal.

Most data selection research in machine translation focuses on improving a single domain. We perform data selection for multiple domains at once. This is achieved by carefully introducing instance-level domain-relevance features and automatically constructing a training curriculum to gradually concentrate on multi-domain relevant and noise-reduced data batches. Both the choice of features and the use of curriculum are crucial for balancing and improving all domains, including out-of-domain. In large-scale experiments, the multi-domain curriculum simultaneously reaches or outperforms the individual performance and brings solid gains over no-curriculum training.

The system presented here took part in the 2018 Trolling, Aggression and Cyberbullying shared task (Forest and Trees team) and uses a Gated Recurrent Neural Network architecture (Cho et al., 2014) in an attempt to assess whether combining pre-trained English and Hindi fastText (Mikolov et al., 2018) word embeddings as a representation of the sequence input would improve classification performance. The motivation for this comes from the fact that the shared task data for English contained many Hindi tokens and therefore some users might be doing code-switching: the alternation between two or more languages in communication. To test this hypothesis, we also aligned Hindi and English vectors using pre-computed SVD matrices that pulls representations from different languages into a common space (Smith et al., 2017). Two conditions were tested: (i) one with standard pre-trained fastText word embeddings where each Hindi word is treated as an OOV token, and (ii) another where word embeddings for Hindi and English are loaded in a common vector space, so Hindi tokens can be assigned a meaningful representation. We submitted the second (i.e., multilingual) system and obtained the scores of 0.531 weighted F1 for the EN-FB dataset and 0.438 weighted F1 for the EN-TW dataset.

The current most popular method for automatic Natural Language Generation (NLG) evaluation is comparing generated text with human-written reference sentences using a metrics system, which has drawbacks around reliability and scalability. We draw inspiration from second language (L2) assessment and extract a set of linguistic features to predict human judgments of sentence naturalness. Our experiment using a small dataset showed that the feature-based approach yields promising results, with the added potential of providing interpretability into the source of the problems.

Emojis are used frequently in social media. A widely assumed view is that emojis express the emotional state of the user, which has led to research focusing on the expressiveness of emojis independent from the linguistic context. We argue that emojis and the linguistic texts can modify the meaning of each other. The overall communicated meaning is not a simple sum of the two channels. In order to study the meaning interplay, we need data indicating the overall sentiment of the entire message as well as the sentiment of the emojis stand-alone. We propose that Facebook Reactions are a good data source for such a purpose. FB reactions (e.g. “Love” and “Angry”) indicate the readers’ overall sentiment, against which we can investigate the types of emojis used the comments under different reaction profiles. We present a data set of 21,000 FB posts (57 million reactions and 8 million comments) from public media pages across four countries.

We present the DUEL corpus, consisting of 24 hours of natural, face-to-face, loosely task-directed dialogue in German, French and Mandarin Chinese. The corpus is uniquely positioned as a cross-linguistic, multimodal dialogue resource controlled for domain. DUEL includes audio, video and body tracking data and is transcribed and annotated for disfluency, laughter and exclamations.