Croatia

Numerous types of social biases have been identified in pre-trained language models (PLMs), and various intrinsic bias evaluation measures have been proposed for quantifying those social biases.
Prior works have relied on human annotated examples to compare existing intrinsic bias evaluation measures.
However, this approach is not easily adaptable to different languages nor amenable to large scale evaluations due to the costs and difficulties when recruiting human annotators.
To overcome this limitation, we propose a method to compare intrinsic gender bias evaluation measures without relying on human-annotated examples.
Specifically, we create multiple bias-controlled versions of PLMs using varying amounts of male vs. female gendered sentences, mined automatically from an unannotated corpus using gender-related word lists. 
Next, each bias-controlled PLM is evaluated using an intrinsic bias evaluation measure, and the rank correlation between the computed bias scores and the gender proportions used to fine-tune the PLMs is computed.
Experiments on multiple corpora and PLMs repeatedly show that the correlations reported by our proposed method that does not require human annotated examples are comparable to those computed using human annotated examples in prior work.

EACL 2023

Comparing Intrinsic Gender Bias Evaluation Measures without using Human Annotated Examples

ethics

fairness

evaluation

poster

### Welcome to the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL) 
Continuing its mission of expanding and involving the science community of all European countries, EACL had selected the Ukrainian community for the 16th EACL, which took place online due to the COVID pandemic. Unfortunately, the ongoing war made the organisation in Kyiv impossible. Considering the importance of physical interaction among researchers, especially after the restrictions imposed by the COVID pandemic, in addition to an online mode, the conference will be held in Dubrovnik, Croatia, from 2 to 6 of May, 2023. The original aim of strengthening the connection with the Ukrainian community will not change as our program will feature a dedicated session and a workshop to highlight work on Ukrainian language technologies. As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.

You need to register for the conference in order to access this site. Please visit https://2023.eacl.org/registration for more information.

As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.

Masked language models (MLMs) conventionally mask 15\% of tokens due to the belief that more masking would leave insufficient context to learn good representations; this masking rate has been widely used, regardless of model sizes or masking strategies. In this work, we revisit this important choice of MLM pre-training. We first establish that 15\% is not universally optimal, and larger models should adopt a higher masking rate. Specifically, we find that masking 40\% outperforms 15\% for BERT-large size models on GLUE and SQuAD. Interestingly, an extremely high masking rate of 80\% can still preserve 95\% fine-tuning performance and most of the accuracy in linguistic probing, challenging the conventional wisdom about the role of the masking rate. We then examine the interplay between masking rates and masking strategies and find that uniform masking requires a higher masking rate compared to sophisticated masking strategies such as span or PMI masking. Finally, we argue that increasing the masking rate has two distinct effects: it leads to more corruption, which makes the prediction task more difficult; it also enables more predictions, which benefits optimization. Using this framework, we revisit BERT's 80-10-10 corruption strategy. Together, our results contribute to a better understanding of MLM pre-training.

Should You Mask 15% in Masked Language Modeling?

Long video content understanding poses a challenging set of research questions as it involves long-distance, cross-media reasoning and knowledge awareness. In this paper, we present a new benchmark for this problem domain, targeting the task of deep movie/TV question answering (QA) beyond previous work's focus on simple plot summary and short video moment settings. We define several baselines based on direct retrieval of relevant context for long-distance movie QA. Observing that real-world QAs may require higher-order multi-hop inferences, we further propose a novel framework, called the DeepMaven, which extracts events, entities, and relations from the rich multimedia content in long videos to pre-construct movie knowledge graphs (movieKGs), and at the time of QA inference, complements general semantics with structured knowledge for more effective information retrieval and knowledge reasoning. We also introduce our recently collected DeepMovieQA dataset, including 1,000 long-form QA pairs from 41 hours of videos, to serve as a new and useful resource for future work. Empirical results show the DeepMaven performs competitively for both the new DeepMovieQA and the pre-existing MovieQA dataset.

DeepMaven: Deep Question Answering on Long-Distance Movie/TV Show Videos with Multimedia Knowledge Extraction and Synthesis

Identifying the difference between two versions of the same article is useful to update knowledge bases and to understand how articles evolve. 
Paired texts occur naturally in diverse situations: reporters write similar news stories and maintainers of authoritative websites must keep their information up to date. 
We propose representing factual changes between paired documents as question-answer pairs, where the answer to the same question differs between two versions. We find that question-answer pairs can flexibly and concisely capture the updated contents. Provided with paired documents, annotators identify questions that are answered by one passage but answered differently or cannot be answered by the other. We release DiffQG which consists of 759 QA pairs and 1153 examples of paired passages with no factual change. These questions are intended to be both unambiguous and information-seeking and involve complex edits, pushing beyond the capabilities of current question generation and factual change detection systems. Our dataset summarizes the changes between two versions of the document as questions and answers, studying automatic update summarization in a novel way.

DiffQG: Generating Questions to Summarize Factual Changes

We propose to use image captions from the Web as a previously underutilized resource for paraphrases (i.e., texts with the same "message") and to create and analyze a corresponding dataset. When an image is reused on the Web, an original caption is often assigned. We hypothesize that different captions for the same image naturally form a set of mutual paraphrases. To demonstrate the suitability of this idea, we analyze captions in the English Wikipedia, where editors frequently relabel the same image for different articles. The paper introduces the underlying mining technology, the resulting Wikipedia-IPC dataset, and compares known paraphrase corpora with respect to their syntactic and semantic paraphrase similarity to our new resource. In this context, we introduce characteristic maps along the two similarity dimensions to identify the style of paraphrases coming from different sources. An annotation study demonstrates the high reliability of the algorithmically determined characteristic maps.

Paraphrase Acquisition from Image Captions

Hate speech has unfortunately become a significant phenomenon on social media platforms, and it can cover various topics (misogyny, sexism, racism, xenophobia, etc.) and targets (e.g., black people, women). Various hate speech detection datasets have been proposed, some annotated for specific topics, and others for hateful speech in general. 
In either case, they often employ different annotation guidelines, which can lead to inconsistencies, even in datasets focusing on the same topics. This can cause issues in models trying to generalize across more data and more topics in order to improve detection accuracy. In this paper, we propose, for the first time, a topic-oriented approach to study generalization across popular hate speech datasets. 
We first perform a comparative analysis of the performances of Transformer-based models in capturing topic-generic and topic-specific knowledge when trained on different datasets. We then propose a novel, simple yet effective approach to study more precisely which topics are best captured in implicit manifestations of hate, showing that selecting combinations of datasets with better out-of-domain topical coverage improves the reliability of automatic hate speech detection.

What Did You Learn To Hate? A Topic-Oriented Analysis of Generalization in Hate Speech Detection

Inspired by retrieval-augmented language generation and pretrained Vision and Language (V\&L) encoders, we present a new approach to image captioning that generates sentences given the input image and a set of captions retrieved from a datastore, as opposed to the image alone. The encoder in our model jointly processes the image and retrieved captions using a pretrained V\&L BERT, while the decoder attends to the multimodal encoder representations, benefiting from the extra textual evidence from the retrieved captions. Experimental results on the COCO dataset show that image captioning can be effectively formulated from this new perspective. Our model, named EXTRA, benefits from using captions retrieved from the training dataset, and it can also benefit from using an external dataset without the need for retraining. Ablation studies show that retrieving a sufficient number of captions (e.g., k=5)
can improve captioning quality. Our work contributes towards using pretrained V\&L encoders for generative tasks, instead of standard classification tasks.

Retrieval-augmented Image Captioning

Large multilingual models have inspired a new class of word alignment methods, which work well for the model's pretraining languages. However, the languages most in need of automatic alignment are low-resource and, thus, not typically included in the pretraining data. In this work, we ask: How do modern aligners perform on unseen languages, and are they better than traditional methods? We contribute gold-standard alignments for Bribri–Spanish, Guarani–Spanish, Quechua–Spanish, and Shipibo-Konibo–Spanish. With these, we evaluate state-of-the-art aligners with and without model adaptation to the target language. Finally, we also evaluate the resulting alignments extrinsically through two downstream tasks: named entity recognition and part-of-speech tagging. We find that although transformer-based methods generally outperform traditional models, the two classes of approach remain competitive with each other.

Meeting the Needs of Low-Resource Languages: The Value of Automatic Alignments via Pretrained Models

Cross-Lingual Dialogue Dataset Creation via Outline-Based Generation

We conducted a human subject study of named entity recognition on a noisy corpus of conversational music recommendation queries, with many irregular and novel named entities. We evaluated the human NER linguistic behaviour in these challenging conditions and compared it with the most common NER systems nowadays, fine-tuned transformers. Our goal was to learn about the task to guide the design of better evaluation methods and NER algorithms. The results showed that NER in our context was quite hard for both human and algorithms under a strict evaluation schema; humans had higher precision, while the model higher recall because of entity exposure especially during pre-training; and entity types had different error patterns (e.g. frequent typing errors for artists). The released corpus goes beyond predefined frames of interaction and can support future work in conversational music recommendation.

A Human Subject Study of Named Entity Recognition in Conversational Music Recommendation Queries

A modular design encourages neural models to disentangle and recombine different facets of knowledge to generalise more systematically to new tasks. In this work, we assume that each task is associated with a subset of latent skills from an (arbitrary size) inventory. In turn, each skill corresponds to a parameter-efficient (sparse / low-rank) model adapter. By jointly learning adapters and a routing function that allocates skills to each task, the full network is instantiated as the average of the parameters of active skills. We propose several inductive biases that encourage re-usage and composition of the skills, including variable-size skill allocation and a dual-speed learning rate. We evaluate our latent-skill model in two main settings: 1) multitask reinforcement learning for instruction following on 8 levels of the BabyAI platform; and 2) few-shot fine-tuning of language models on 160 NLP tasks of the CrossFit benchmark. We find that the modular design of our network enhances sample efficiency in reinforcement learning and few-shot generalisation in supervised learning, compared to a series of baselines. These include models where parameters are fully shared, task-specific, conditionally generated (HyperFormer), or sparse mixture-of-experts (TaskMoE).

Downloads

Next from EACL 2023

Should You Mask 15% in Masked Language Modeling?