Large language models (LLMs) often fail to generate text in the intended target language, particularly in non-English interactions. Concurrently, recent work has explored Language Neuron Intervention (LNI) as a promising technique for steering output language. In this paper, we re-evaluate LNI in more practical scenarios—specifically with instruction-tuned models and prompts that explicitly specify the target language. Our experiments show that while LNI also shows potential in such practical scenarios, its average effect is limited and unstable across models and tasks, with a 0.83% reduction in undesired language output and a 0.1% improvement in performance. Our further analysis identifies two key factors for LNI’s limitation: (1) LNI affects both the output language and the content semantics, making it hard to control one without affecting the other, which explains the weak performance gains. (2) LNI increases the target language token probabilities, but they often remain below the top-1generation threshold, resulting in failure to generate the target language in most cases. Our results highlight both the potential and limitations of LNI, paving the way for future improvements.

Integrated Gradients (IG) serve as a potent tool for explaining the internal structure of a language model. The calculation of IG requires numerical integration, wherein the number of steps serves as a critical hyperparameter. The step count can drastically alter the results, inducing considerable errors in interpretability. To scrutinize the effect of step variation on IG, we measured the difference between theoretical and observed IG totals for each step amount.Our findings indicate that the ideal number of steps to maintain minimal error varies from instance to instance. Consequently, we advocate for customizing the step count for each instance. Our study is the first to quantitatively analyze the variation of IG values with the number of steps.

Low-Rank Adaptation (LoRA) is a widely used Parameter-Efficient Fine-Tuning (PEFT) method that updates an initial weight matrix W₀ with a delta matrix 𝛥 W consisted by two low-rank matrices A and B. A previous study suggested that there is correlation between W₀ and 𝛥 W. In this study, we aim to delve deeper into relationships between W₀ and low-rank matrices A and B to further comprehend the behavior of LoRA. In particular, we analyze a conversion matrix that transform W₀ into low-rank matrices, which encapsulates information about the relationships. Our analysis reveals that the conversion matrices are similar across each layer. Inspired by these findings, we hypothesize that a single linear layer, which takes each layer’s W₀ as input, can yield task-adapted low-rank matrices. To confirm this hypothesis, we devise a method named Conditionally Parameterized LoRA (CondLoRA) that updates initial weight matrices with low-rank matrices derived from a single linear layer. Our empirical results show that CondLoRA maintains a performance on par with LoRA, despite the fact that the trainable parameters of CondLoRA are fewer than those of LoRA. Therefore, we conclude that “a single linear layer yields task-adapted low-rank matrices.” The code used in our experiments is available at https://github.com/CyberAgentAILab/CondLoRA.

This paper describes our system submitted to SemEval-2023 Task 5: Clickbait Spoiling. We work on spoiler generation of the subtask 2 and develop a system which comprises two parts: 1) simple seq2seq spoiler generation and 2) post-hoc model ensembling. Using this simple method, we address the challenge of generating multipart spoiler. In the test set, our submitted system outperformed the baseline by a large margin (approximately 10 points above on the BLEU score) for mixed types of spoilers. We also found that our system successfully handled the challenge of the multipart spoiler, confirming the effectiveness of our approach.

Many recent Short Answer Scoring (SAS) systems have employed Quadratic Weighted Kappa (QWK) as the evaluation measure of their systems. However, we hypothesize that QWK is unsatisfactory for the evaluation of the SAS systems when we consider measuring their effectiveness in actual usage. We introduce a new task formulation of SAS that matches the actual usage. In our formulation, the SAS systems should extract as many scoring predictions that are not critical scoring errors (CSEs). We conduct the experiments in our new task formulation and demonstrate that a typical SAS system can predict scores with zero CSE for approximately 50% of test data at maximum by filtering out low-reliablility predictions on the basis of a certain confidence estimation. This result directly indicates the possibility of reducing half the scoring cost of human raters, which is more preferable for the evaluation of SAS systems.

The idea of subword-based word embeddings has been proposed in the literature, mainly for solving the out-of-vocabulary (OOV) word problem observed in standard word-based word embeddings. In this paper, we propose a method of reconstructing pre-trained word embeddings using subword information that can effectively represent a large number of subword embeddings in a considerably small fixed space. The key techniques of our method are twofold: memory-shared embeddings and a variant of the key-value-query self-attention mechanism. Our experiments show that our reconstructed subword-based embeddings can successfully imitate well-trained word embeddings in a small fixed space while preventing quality degradation across several linguistic benchmark datasets, and can simultaneously predict effective embeddings of OOV words. We also demonstrate the effectiveness of our reconstruction method when we apply them to downstream tasks.

This paper describes our system submitted to the formal run of SemEval-2019 Task 4: Hyperpartisan news detection. Our system is based on a linear classifier using several features, i.e., 1) embedding features based on the pre-trained BERT embeddings, 2) article length features, and 3) embedding features of informative phrases extracted from by-publisher dataset. Our system achieved 80.9% accuracy on the test set for the formal run and got the 3rd place out of 42 teams.

Cross-lingual information retrieval (CLIR) is a document retrieval task where the documents are written in a language different from that of the user’s query. This is a challenging problem for data-driven approaches due to the general lack of labeled training data. We introduce a large-scale dataset derived from Wikipedia to support CLIR research in 25 languages. Further, we present a simple yet effective neural learning-to-rank model that shares representations across languages and reduces the data requirement. This model can exploit training data in, for example, Japanese-English CLIR to improve the results of Swahili-English CLIR.