Publications

ツール拡張言語モデルを用いた対話管理エージェントは流行しているが、既存ベンチマークデータは実際のビジネスタスクと乖離があり、商用チャットボットに使えない。本研究で提案するデータ作成手法ではデータ作成者が与えた必要最小限の情報に基づいて、要望に沿った大量で高品質なデータを自動的に生成してオーダーメイド対話管理を実現した。作成されたデータは学習データとして対話管理エージェントの学習や改良に使える。

スキルマッピングは職務記述書や履歴書中の文で言及されるオントロジーで定義されたスキルを特定する作業であり労働市場の分析に不可欠である．詳細な分析に必要な細分化されたスキルが付与された学習データを人手で構築するのは高いコストなため既存研究は Large Language Model (LLM) が生成した合成データが bi-encoder の学習に用いる．合成データを活用したさらなる精度改善のため，提案手法 (𝑘NNBE) は推論時に学習に利用したラベル付き合成文を 𝑘-nearest neighbor (𝑘NN) によって取得し入 力文との類似度を bi-encoder のスコアに加える．実験により 𝑘NNBE は bi-encoder の精度改善，さらに既存の最高精度を示す LLM でスキルをリランキングする手法と比較して高いスループットを維持しながら精度改善を確認した．

Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from multiple documents. Since no benchmarks exist for investigating hallucinations in MDS, we use existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, gpt-3.5-turbo and GPT-4o still generate summaries about 79.35% and 44% of the time, raising concerns about their tendency to fabricate content. To understand the characteristics of these hallucinations, we manually evaluate 700+ insights and find that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches to systematically mitigate hallucinations in MDS. We release our dataset and code.

The rapid increase in textual information means we need more efficient methods to sift through, organize, and understand it all. While retrieval-augmented generation (RAG) models excel in accessing information from large document collections, they struggle with complex tasks that require aggregation and reasoning over information spanning across multiple documents–what we call holistic reasoning. Long-context language models (LCLMs) have great potential for managing largescale documents, but their holistic reasoning capabilities remain unclear. In this work, we introduce HoloBench, a novel framework that brings database reasoning operations into text-based contexts, making it easier to systematically evaluate how LCLMs handle holistic reasoning across large documents. Our approach adjusts key factors such as context length, information density, distribution of information, and query complexity to evaluate LCLMs comprehensively. Our experiments show that the amount of information in the context has a bigger influence on LCLM performance than the actual context length. Furthermore, the complexity of queries affects performance more than the amount of information, particularly for different types of queries. Interestingly, queries that involve finding maximum or minimum values are easier for LCLMs and are less affected by context length, even though they pose challenges for RAG systems. However, tasks requiring the aggregation of multiple pieces of information show a noticeable drop in accuracy as context length increases. Additionally, we find that while grouping relevant information generally improves performance, the optimal positioning varies across models. Our findings surface both the advancements and the ongoing challenges in achieving a holistic understanding of long contexts. These can guide future developments in LCLMs and set the stage for creating more robust language models for real-world applications.

Large language models (LLMs), despite their impressive capabilities in natural language understanding tasks in open-domain, often lack effectiveness with similar tasks in enterprise applications due to potential hallucinations, weak multi-hop reasoning ability, and limitations in adapting to heterogeneous data types, among others. Such issues primarily arise due to the absence of private, on-premises enterprises from an LLM’s training corpus. Knowledge-intensive tasks in enterprise often require multi-step reasoning, deep contextual understanding, and integration of information stored and accessed in heterogeneous formats (e.g., tables, graphs, documents, and JSON), which LLMs aren’t inherently equipped to handle without significant adaptation. To this end, retrieval augmented generation (RAG) offers promise in instrumenting such adaptations on demand. While RAG-based approaches focus on controlling the generation and mitigating hallucinations, existing solutions are not sufficient for the requirements of the enterprise settings. In this paper, we outline our approaches toward understanding and implementing a more effective RAG workflow in the wild. To achieve the goal, we draw on the cognitive science concepts of System 1 (fast, intuitive thinking) and System 2 (slow, deliberate, analytical thinking.) In particular, we discuss how existing RAG approaches are more aligned to System 1 and propose to shift from traditional single-model architectures to compound AI systems within a System 2 framework to improve RAG, especially in complex enterprise applications. Such compound AI systems adopt a more systematic approach by assigning specialized tasks to different intelligent agents, optimizing retrieval and generation performance with a retrieval-augmented generation workflow.

大規模言語モデル（LargeLanguageModels;LLMs）の台頭により、自然言語を用いた指示で多岐にわたる言語処理タスクが実行可能になった。しかし、与えられた指示が曖昧性であるためにユーザーの意図と異なるテキストが生成されることがある。特に自然言語生成（NaturalLanguageGeneration;NLG）に見られる曖昧性は広範で、人間が曖昧でない指示を書くことは難しい。そこで本研究では、曖昧性解消ベンチマークデータセットAmbiNLGの導入および自動的かつ網羅的な曖昧性解消を行う。実験より、提案手法による複数の曖昧性の明示的・網羅的な曖昧性解消の有効性を示した。

自然言語処理技術が使われている様々なサービスに、検索が中心的役割を果たしているものが多いが、近年の大規模言語モデルは検索に与えた影響が不明だ。本調査報告書は宿検索に対する文埋め込み計算手法を比較した。文の埋め込みの計算において、従来のBERTモデルとOpenAI社の新しいモデルのそれぞれの強みと弱みを分析した。

Detecting out-of-scope (OOS) utterances is crucial in task-oriented dialogue systems, but obtaining enough annotated OOS dialogues to train a binary classifier directly is difficult in practice. Existing data augmentation methods generate OOS dialogues automatically, but their performance usually depends on an external corpus. Herein we propose SILVER, a self data augmentation method that does not use external data. It improves the accuracy of OOS detection (false positive rate: 90.5% → 47.4%). Furthermore, SILVER successfully generates high-quality in-domain (IND) OOS dialogues in terms of naturalness (percentage: 8% → 68%) and OOS correctness (percentage: 74% → 88%), as evaluated by human workers.