This is an internal, incomplete preview of a proposed change to the ACL Anthology.
For efficiency reasons, we generate only three BibTeX files per volume, and the preview may be incomplete in other ways, or contain mistakes.
Do not treat this content as an official publication.
This paper details the process of developing the first native large generative language model for the North Germanic languages, GPT-SW3. We cover all parts of the development process, from data collection and processing, training configuration and instruction finetuning, to evaluation, applications, and considerations for release strategies. We discuss pros and cons of developing large language models for smaller languages and in relatively peripheral regions of the globe, and we hope that this paper can serve as a guide and reference for other researchers that undertake the development of large generative models for smaller languages.
We present Superlim, a multi-task NLP benchmark and analysis platform for evaluating Swedish language models, a counterpart to the English-language (Super)GLUE suite. We describe the dataset, the tasks, the leaderboard and report the baseline results yielded by a reference implementation. The tested models do not approach ceiling performance on any of the tasks, which suggests that Superlim is truly difficult, a desirable quality for a benchmark. We address methodological challenges, such as mitigating the Anglocentric bias when creating datasets for a less-resourced language; choosing the most appropriate measures; documenting the datasets and making the leaderboard convenient and transparent. We also highlight other potential usages of the dataset, such as, for instance, the evaluation of cross-lingual transfer learning.
Recent studies in cross-lingual learning using multilingual models have cast doubt on the previous hypothesis that shared vocabulary and joint pre-training are the keys to cross-lingual generalization. We introduce a method for transferring monolingual models to other languages through continuous pre-training and study the effects of such transfer from four different languages to English. Our experimental results on GLUE show that the transferred models outperform an English model trained from scratch, independently of the source language. After probing the model representations, we find that model knowledge from the source language enhances the learning of syntactic and semantic knowledge in English.
We present GTP-SW3, a 3.5 billion parameter autoregressive language model, trained on a newly created 100 GB Swedish corpus. This paper provides insights with regards to data collection and training, while highlights the challenges of proper model evaluation. The results of quantitive evaluation through perplexity indicate that GPT-SW3 is a competent model in comparison with existing autoregressive models of similar size. Additionally, we perform an extensive prompting study which reveals the good text generation capabilities of GTP-SW3.
The long-standing endeavor of relating the textual and the visual domain recently underwent a pivotal breakthrough, as OpenAI released CLIP. This model distinguishes how well an English text corresponds with a given image with unprecedented accuracy. Trained via a contrastive learning objective over a huge dataset of 400M of images and captions, it is a work that is not easily replicated, especially for low resource languages. Capitalizing on the modularization of the CLIP architecture, we propose to use cross-lingual teacher learning to re-train the textual encoder for various non-English languages. Our method requires no image data and relies entirely on machine translation which removes the need for data in the target language. We find that our method can efficiently train a new textual encoder with relatively low computational cost, whilst still outperforming previous baselines on multilingual image-text retrieval.
The introduction of immensely large Causal Language Models (CLMs) has rejuvenated the interest in open-ended text generation. However, controlling the generative process for these Transformer-based models is at large an unsolved problem. Earlier work has explored either plug-and-play decoding strategies, or more powerful but blunt approaches such as prompting. There hence currently exists a trade-off between fine-grained control, and the capability for more expressive high-level instructions. To alleviate this trade-off, we propose an encoder-decoder architecture that enables intermediate text prompts at arbitrary time steps. We propose a resource-efficient method for converting a pre-trained CLM into this architecture, and demonstrate its potential on various experiments, including the novel task of contextualized word inclusion. Our method provides strong results on multiple experimental settings, proving itself to be both expressive and versatile.
We investigate the feasibility of applying standard text categorisation methods to patient text in order to predict treatment outcome in Internet-based cognitive behavioural therapy. The data set is unique in its detail and size for regular care for depression, social anxiety, and panic disorder. Our results indicate that there is a signal in the depression data, albeit a weak one. We also perform terminological and sentiment analysis, which confirm those results.
When is it beneficial for a research community to organize a broader collaborative effort on a topic, and when should we instead promote individual efforts? In this opinion piece, we argue that we are at a stage in the development of large-scale language models where a collaborative effort is desirable, despite the fact that the preconditions for making individual contributions have never been better. We consider a number of arguments for collaboratively developing a large-scale Nordic language model, include environmental considerations, cost, data availability, language typology, cultural similarity, and transparency. Our primary goal is to raise awareness and foster a discussion about our potential impact and responsibility as NLP community.
Advanced NLP models require huge amounts of data from various domains to produce high-quality representations. It is useful then for a few large public and private organizations to join their corpora during training. However, factors such as legislation and user emphasis on data privacy may prevent centralized orchestration and data sharing among these organizations. Therefore, for this specific scenario, we investigate how gossip learning, a massively-parallel, data-private, decentralized protocol, compares to a shared-dataset solution. We find that the application of Word2Vec in a gossip learning framework is viable. Without any tuning, the results are comparable to a traditional centralized setting, with a loss of quality as low as 4.3%. Furthermore, the results are up to 54.8% better than independent local training.
Most work in NLP makes the assumption that it is desirable to develop solutions in the native language in question. There is consequently a strong trend towards building native language models even for low-resource languages. This paper questions this development, and explores the idea of simply translating the data into English, thereby enabling the use of pretrained, and large-scale, English language models. We demonstrate empirically that a large English language model coupled with modern machine translation outperforms native language models in most Scandinavian languages. The exception to this is Finnish, which we assume is due to inferior translation quality. Our results suggest that machine translation is a mature technology, which raises a serious counter-argument for training native language models for low-resource languages. This paper therefore strives to make a provocative but important point. As English language models are improving at an unprecedented pace, which in turn improves machine translation, it is from an empirical and environmental stand-point more effective to translate data from low-resource languages into English, than to build language models for such languages.
This paper introduces a long-range multiple-choice Question Answering (QA) dataset, based on full-length fiction book texts. The questions are formulated as 10-way multiple-choice questions, where the task is to select the correct character name given a character description, or vice-versa. Each character description is formulated in natural text and often contains information from several sections throughout the book. We provide 20,000 questions created from 10,000 manually annotated descriptions of characters from 177 books containing 152,917 words on average. We address the current discourse regarding dataset bias and leakage by a simple anonymization procedure, which in turn enables interesting probing possibilities. Finally, we show that suitable baseline algorithms perform very poorly on this task, with the book size itself making it non-trivial to attempt a Transformer-based QA solution. This leaves ample room for future improvement, and hints at the need for a completely different type of solution.
This paper problematizes the reliance on documents as the basic notion for defining term interactions in standard topic models. As an alternative to this practice, we reformulate topic distributions as latent factors in term similarity space. We exemplify the idea using a number of standard word embeddings built with very wide context windows. The embedding spaces are transformed to sparse similarity spaces, and topics are extracted in standard fashion by factorizing to a lower-dimensional space. We use a number of different factorization techniques, and evaluate the various models using a large set of evaluation metrics, including previously published coherence measures, as well as a number of novel measures that we suggest better correspond to real-world applications of topic models. Our results clearly demonstrate that term-based models outperform standard document-based models by a large margin.
We cast the problem of event annotation as one of text categorization, and compare state of the art text categorization techniques on event data produced within the Uppsala Conflict Data Program (UCDP). Annotating a single text involves assigning the labels pertaining to at least 17 distinct categorization tasks, e.g., who were the attacking organization, who was attacked, and where did the event take place. The text categorization techniques under scrutiny are a classical Bag-of-Words approach; character-based contextualized embeddings produced by ELMo; embeddings produced by the BERT base model, and a version of BERT base fine-tuned on UCDP data; and a pre-trained and fine-tuned classifier based on ULMFiT. The categorization tasks are very diverse in terms of the number of classes to predict as well as the skeweness of the distribution of classes. The categorization results exhibit a large variability across tasks, ranging from 30.3% to 99.8% F-score.
We (Team Skurt) propose a simple method to detect lexical semantic change by clustering contextualized embeddings produced by XLM-R, using K-Means++. The basic idea is that contextualized embeddings that encode the same sense are located in close proximity in the embedding space. Our approach is both simple and generic, but yet performs relatively good in both sub-tasks of SemEval-2020 Task 1. We hypothesize that the main shortcoming of our method lies in the simplicity of the clustering method used.
This paper investigates data-driven segmentation using Re-Pair or Byte Pair Encoding-techniques. In contrast to previous work which has primarily been focused on subword units for machine translation, we are interested in the general properties of such segments above the word level. We call these segments r-grams, and discuss their properties and the effect they have on the token frequency distribution. The proposed approach is evaluated by demonstrating its viability in embedding techniques, both in monolingual and multilingual test settings. We also provide a number of qualitative examples of the proposed methodology, demonstrating its viability as a language-invariant segmentation procedure.
This paper investigates the presence of gender bias in pretrained Swedish embeddings. We focus on a scenario where names are matched with occupations, and we demonstrate how a number of standard pretrained embeddings handle this task. Our experiments show some significant differences between the pretrained embeddings, with word-based methods showing the most bias and contextualized language models showing the least. We also demonstrate that the previously proposed debiasing method does not affect the performance of the various embeddings in this scenario.
This paper discusses the question whether it is possible to learn a generic representation that is useful for detecting various types of abusive language. The approach is inspired by recent advances in transfer learning and word embeddings, and we learn representations from two different datasets containing various degrees of abusive language. We compare the learned representation with two standard approaches; one based on lexica, and one based on data-specific n-grams. Our experiments show that learned representations do contain useful information that can be used to improve detection performance when training data is limited.
Sentiment and topic analysis are common methods used for social media monitoring. Essentially, these methods answers questions such as, “what is being talked about, regarding X”, and “what do people feel, regarding X”. In this paper, we investigate another venue for social media monitoring, namely issue ownership and agenda setting, which are concepts from political science that have been used to explain voter choice and electoral outcomes. We argue that issue alignment and agenda setting can be seen as a kind of semantic source similarity of the kind “how similar is source A to issue owner P, when talking about issue X”, and as such can be measured using word/document embedding techniques. We present work in progress towards measuring that kind of conditioned similarity, and introduce a new notion of similarity for predictive embeddings. We then test this method by measuring the similarity between politically aligned media and political parties, conditioned on bloc-specific issues.
Automatic detection of five language components, which are all relevant for expressing opinions and for stance taking, was studied: positive sentiment, negative sentiment, speculation, contrast and condition. A resource-aware approach was taken, which included manual annotation of 500 training samples and the use of limited lexical resources. Active learning was compared to random selection of training data, as well as to a lexicon-based method. Active learning was successful for the categories speculation, contrast and condition, but not for the two sentiment categories, for which results achieved when using active learning were similar to those achieved when applying a random selection of training data. This difference is likely due to a larger variation in how sentiment is expressed than in how speakers express the other three categories. This larger variation was also shown by the lower recall results achieved by the lexicon-based approach for sentiment than for the categories speculation, contrast and condition.
This paper presents the Gavagai Living Lexicon, which is an online distributional semantic model currently available in 20 different languages. We describe the underlying distributional semantic model, and how we have solved some of the challenges in applying such a model to large amounts of streaming data. We also describe the architecture of our implementation, and discuss how we deal with continuous quality assurance of the lexicon.
This paper discusses evaluation methodologies for a particular kind of meaning models known as word-space models, which use distributional information to assemble geometric representations of meaning similarities. Word-space models have received considerable attention in recent years, and have begun to see employment outside the walls of computational linguistics laboratories. However, the evaluation methodologies of such models remain infantile, and lack efforts at standardization. Very few studies have critically assessed the methodologies used to evaluate word spaces. This paper attempts to fill some of this void. It is the central goal of this paper to answer the question how can we determine whether a given word space is a good word space?