Thailand

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model&#39;s diverse response styles such as starting with &quot;Sure&quot; or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

ACL 2024

My Answer is C: First-Token Probabilities Do Not Match Text Answers in Instruction-Tuned Language Models

mcqa

llms

bias

evaluation

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

workshop paper

This paper presents an interpretable unsupervised morphological learning model, showing comparable performance to supervised models in learning complex morphological rules of Turkish as evidenced by its application to the problem of morphological inflection within the SIGMORPHON Shared Tasks. The significance of our unsupervised approach lies in its alignment with how humans naturally acquire rules from raw data without supervision. To achieve this, we construct a model with multiple codebooks of VQ-VAE employing continuous and discrete latent variables during word generation. We evaluate the model's performance under high and low-resource scenarios, and use probing techniques to examine encoded information in latent representations. We also evaluate its generalization capabilities by testing unseen suffixation scenarios within the SIGMORPHON-UniMorph 2022 Shared Task 0. Our results demonstrate our model's ability to distinguish word structures into lemmas and suffixes, with each codebook specialized for different morphological features, contributing to the interpretability of our model and effectively performing morphological inflection on both seen and unseen morphological features.

Unsupervised Learning of Turkish Morphology with Multiple Codebook VQ-VAE

The emergence of multilingual large language models has enabled the development of language understanding and generation systems in Azerbaijani. However, most of the production-grade systems rely on cloud solutions, such as GPT-4. While there have been several attempts to develop open foundation models for Azerbaijani, these works have not found their way into common use due to a lack of systemic benchmarking. This paper encompasses several lines of work that promote open-source foundation models for Azerbaijani. We introduce (1) a large text corpus for Azerbaijani, (2) a family of encoder-only language models trained on this dataset, (3) labeled datasets for evaluating these models, and (4) extensive evaluation that covers all major open-source models with Azerbaijani support.

Open foundation models for Azerbaijani language

We introduce ImplicaTR, a linguistically informed diagnostic dataset designed to evaluate semantic and pragmatic reasoning capabilities of Natural Language Inference (NLI) models in Turkish. Existing Turkish NLI datasets treat NLI as determining whether a sentence pair represents $\textit{entailment}$, $\textit{contradiction}$, or a $\textit{neutral}$ relation. Such datasets do not distinguish between $\textit{semantic entailment}$ and $\textit{pragmatic implicature}$, which linguists have long recognized as separate inferences types. ImplicaTR addresses this by testing NLI models' ability to differentiate between $\textit{entailment}$ and $\textit{implicature}$, thus assessing their pragmatic reasoning skills. The dataset consists of 19,350 semi-automatically generated sentence pairs covering $\textit{implicature, entailment, contradiction,}$ and $\textit{neutral}$ relations. We evaluated various models (BERT, Gemma, Llama-2, and Mistral) on ImplicaTR and found out that these models can reach up to 98\% accuracy on semantic and pragmatic reasoning. We also fine tuned various models on subsets of ImplicaTR to test the abilities of NLI models to generalize across unseen implicature contexts. Our results indicate that model performance is highly dependent on the diversity of linguistic expressions within each subset, highlighting a weakness in the abstract generalization capabilities of large language models regarding pragmatic reasoning. We share all the code, models, and the dataset.

ImplicaTR: A Granular Dataset for Natural Language Inference and Pragmatic Reasoning in Turkish

The paper introduces a publicly available corpus of Turkish situated dialogs annotated for coreference. We developed an annotation scheme for coreference annotation in Turkish, a language with pro-drop and rich agglutinating morphology. The annotation scheme is tailored for these aspects of the language, making it potentially applicable to similar languages. The corpus comprises 60 dialogs containing in total 3900 sentences, 18360 words, and 6120 mentions.

A coreference corpus of Turkish situated dialogs

Large language models (LLMs) have shown impressive capabilities in tasks such as machine translation, text summarization, question answering, and solving complex mathematical problems. However, their primary training on data-rich languages like English limits their performance in low-resource languages. This study addresses this gap by focusing on the Indexical Shift problem in Turkish. The Indexical Shift problem involves resolving pronouns in indexical shift contexts, a grammatical challenge not present in high-resource languages like English. We present the first study examining indexical shift in any language, releasing a Turkish dataset specifically designed for this purpose. Our Indexical Shift Dataset consists of 156 multiple-choice questions, each annotated with necessary linguistic details, to evaluate LLMs in a few-shot setting. We evaluate recent multilingual LLMs, including GPT-4, GPT-3.5, Cohere-AYA, Trendyol-LLM, and Turkcell-LLM, using this dataset. Our analysis reveals that even advanced models like GPT-4 struggle with the grammatical nuances of indexical shift in Turkish, achieving only moderate performance. These findings underscore the need for focused research on the grammatical challenges posed by low-resource languages. We released the dataset and code \href{https://anonymous.4open.science/r/indexical_shift_llm-E1B4} {here}.

Do LLMs Recognize me, When I is not me: Assessment of LLMs Understanding of Turkish Indexical Pronouns in Indexical Shift Contexts

Ottoman Turkish, as a historical variant of modern Turkish, suffers from a scarcity of available corpora and NLP models. This paper outlines our pioneering endeavors to address this gap by constructing a clean text corpus of Ottoman Turkish materials. We detail the challenges encountered in this process and offer potential solutions. Additionally, we present a case study wherein the created corpus is employed in continual pre-training of BERTurk, followed by evaluation of the model's performance on the named entity recognition task for Ottoman Turkish. Preliminary experimental results suggest the effectiveness of our corpus in adapting existing models developed for modern Turkish to historical Turkish.

Towards a Clean Text Corpus for Ottoman Turkish

Euphemisms are a form of figurative language relatively understudied in natural language processing. This research extends the current computational work on potentially euphemistic terms (PETs) to Turkish. We introduce the Turkish PET dataset, the first available of its kind in the field. By creating a list of euphemisms in Turkish, collecting example contexts, and annotating them, we provide both euphemistic and non-euphemistic examples of PETs in Turkish. We describe the dataset and methodologies, and also experiment with transformer-based models on Turkish euphemism detection by using our dataset for binary classification. We compare performances across models using F1, accuracy, and precision as evaluation metrics.

Turkish Delights: a Dataset on Turkish Euphemisms

We conducted a systematic evaluation of seven large language models (LLMs) on tasks in Kazakh. Kazakh is a Turkic language spoken by approximately 13 million native speakers in Kazakhstan and abroad. We used six datasets corresponding to different tasks -- questions answering, causal reasoning, middle school math problems, machine translation, and spelling correction. Three of the datasets were prepared for this study. As expected, the quality of the LLMs on the Kazakh tasks is lower than on the parallel English tasks. GPT-4 shows the best results, followed by Gemini and Aya. In general, LLMs perform better on classification tasks and struggle with generative tasks. Our results provide valuable insights into the applicability of currently available LLMs for Kazakh. We will publish the data collected for this study, which will be a good start for an LLM benchmark focused on Kazakh.

Do LLMs Speak Kazakh? A Pilot Evaluation of Seven Models

This paper presents our work on tools to support the Tatar language, using Revita, a web-based Intelligent Tutoring System for language learning and teaching. The system allows the users — teachers and learners — to upload arbitrary authentic texts, and automatically creates exercises based on these texts that engage the learners in active production of language. It provides graduated feedback when they make mistakes, and performs continuous assessment, which helps the students maintain their learning pace, and helps the teachers to monitor their progress. 
The paper describes the functionality currently implemented for Tatar, which enables learners — who possess basic proficiency beyond the beginner level — to improve their competency, using texts of their choice as learning content. Support for Tatar is being developed to increase public interest in learning the language of this important regional minority, as well as to to provide tools for improving fluency to "heritage speakers" — those who have substantial passive competency, but lack active fluency and need support for regular practice.

An Intelligent Tutor to Support Teaching and Learning of Tatar

This paper reports on the performance of SRCB’s system in the Social Media Mining for Health (#SMM4H) 2024 Shared Task 1: extrac- tion and normalization of adverse drug events (ADEs) in English tweets. We develop a sys- tem composed of an ADE extraction module and an ADE normalization module which fur- ther includes a retrieval module and a filtering module. To alleviate the data imbalance and other issues introduced by the dataset, we em- ploy 4 data augmentation techniques based on Large Language Models (LLMs) across both modules. Our best submission achieves an F1 score of 53.6 (49.4 on the unseen subset) on the ADE normalization task and an F1 score of 52.1 on ADE extraction task.

Downloads

Next from ACL 2024

Unsupervised Learning of Turkish Morphology with Multiple Codebook VQ-VAE

Similar lecture

"My Answer is C": First-Token Probabilities Do Not Match Text Answers in Instruction-Tuned Language Models