PaperlayerView exact revision on arXiv

Source-linked AI summary

Reproducing Omitted Temporal Expressions in Japanese News for Retrieval-Augmented Applications

Tomoaki Yasuda, Shotaro Ishihara

arXiv:2609.09569v1cs.CL

TL;DR

Japanese news often omits temporal information that readers recover from publication dates, creating problems when articles are processed as standalone text for search and RAG. The paper introduces jaROTE, a rule-based pipeline that normalizes such expressions before indexing; experiments show improved time-constrained retrieval and a fast, low-cost alternative that does not require LLM inference.

  • Problem

    Standalone news processing can lose publication-date context, leaving omitted temporal expressions underspecified and causing temporal mismatches, incomplete representations, and unstable LLM interpretation.

  • Method

    jaROTE is a rule-based Japanese-news pipeline that extracts omitted temporal expressions and normalizes them into dates or intervals using publication dates and local context.

  • Results

    Temporal reproduction improves time-constrained retrieval, especially lexical retrieval, while jaROTE remains fast and low-cost without LLM inference.

  • Takeaways & Limitations

    Publication-date-grounded temporal reproduction can serve as a lightweight preprocessing step for indexing Japanese news in search and RAG systems.

  • Takeaways & Limitations

    Generalization remains unverified because corpus analysis and rule development used the Nikkei and Livedoor data, while robustness across publishers, genres, periods, larger collections, and other domains remains unevaluated.

Abstract

from arXiv · show

News articles often contain omitted temporal expressions, such as day-only or month-only mentions, which must be interpreted with reference to the publication date. When such articles are indexed or processed as standalone text in search and retrieval-augmented generation (RAG) systems, these omissions can cause temporal mismatches and unstable interpretation by large language models. We focus on reproducing omitted temporal expressions as concrete dates or intervals using the publication date as external context before the articles are indexed for search and RAG applications. Specifically, building on established temporal-expression extraction and normalization techniques and informed by a manual analysis of Japanese news articles, we propose jaROTE, a rule-based pipeline for Japanese news. Experiments on two news corpora demonstrate that jaROTE achieves high performance, and remains competitive with LLMs while providing a fast, low-cost pipeline. We further show that temporal reproduction improves time-constrained lexical retrieval, demonstrating the practical value of publication-date-grounded normalization for Japanese news retrieval.

1 Introduction

Japanese news often omits parts of temporal expressions that readers infer from publication dates, but standalone indexing can leave them underspecified. The paper proposes jaROTE, a rule-based pipeline that reproduces these expressions before search and RAG processing, and evaluates their retrieval value.

  • Standalone news chunks can lose publication-date context, causing temporal mismatches, incomplete representations, and unstable LLM interpretation.
  • The study analyzes omitted expressions, creates gold reproductions, evaluates jaROTE on Nikkei and Livedoor corpora, and tests time-constrained retrieval.
  • jaROTE reproduces omitted temporal expressions as concrete dates or intervals using publication date and local context.The approach decomposes reproduction into extraction and normalization.
  • jaROTE is an auditable rule-based pipeline designed to be fast, low-cost, and usable without LLM inference.
  • Temporal reproduction improves time-constrained retrieval, especially for lexical retrieval, without modifying the retriever or embedding model.

2 Related Work

The paper builds on established temporal-expression extraction and normalization research, including Japanese resources and methods. Its distinctive focus is publication-date-grounded reproduction of omitted expressions for Japanese news search and RAG applications.

  • Prior work includes temporal-expression recognition and normalization benchmarks, annotation schemes, and rule-based systems such as SUTime.
  • Japanese temporal processing includes BCCWJ-TimeBank and methods addressing DATE–DURATION ambiguity in numeric day expressions.
  • The ambiguity distinction matters because duration expressions can otherwise be incorrectly treated as dates during extraction.
  • jaROTE extends this literature by targeting publication-date-grounded reproduction of omitted temporal expressions for downstream search and RAG.

3 Analyzing Temporal Expressions in Japanese News

The authors define five reproducible categories by analyzing omitted temporal expressions in 97 Nikkei articles and requiring extraction plus normalization into concrete dates or intervals. Most target expressions use publication date as the primary reference, although a small minority require discourse-level reference tracking.

  • The analysis covers 97 Nikkei articles and uses publication dates because news expressions are abbreviated and publication metadata is explicit.
  • The five target categories account for 256 of 303 omitted temporal expressions, or 84.5%.
  • Incomplete expressions omit year, month, or day components that can often be completed from the publication date.
  • Relative expressions use calendar arithmetic, while range and composite expressions require completing interval endpoints or combining operations.
  • Deictic, seasonal, and comparison expressions are excluded when they require additional temporal references or cannot become concrete retrieval dates or intervals.
  • Three of 256 target expressions, or 1.2%, required updating the discourse reference time rather than using the publication date.

4 Proposed Method: jaROTE

jaROTE reproduces omitted temporal expressions by combining preprocessing, rule-based extraction, multiple-reference-date analysis, and context-based normalization. It outputs explicit dates or periods grounded in an article’s publication date and local context.

  • 4 Proposed Method: jaROTE: jaROTE takes an article and publication date as input and outputs text with target temporal expressions reproduced as explicit dates or periods.
  • 4 Proposed Method: jaROTE: The pipeline masks non-target numerical expressions, including durations, financial expressions, addresses, scores, and coarse multi-year era expressions, before processing.
  • 4 Proposed Method: jaROTE: It converts Japanese era expressions, expands two-digit years, and completes omitted range endpoints so each endpoint can be normalized independently.
  • 4 Proposed Method: jaROTE: After preprocessing, jaROTE applies ja-timex, a rule-based parser that extracts candidate Japanese temporal expressions and structured temporal annotations.
  • 4 Proposed Method: jaROTE: For each candidate, jaROTE generates analyses using the publication date, the previous month’s last day, and the publication date in the next year.
  • 4 Proposed Method: jaROTE: It selects an analysis using expression granularity and local context, applying future-oriented patterns only to incomplete day-only and month-day expressions.

5 Evaluation: Temporal Reproduction

The evaluation measures extraction, fixed-span normalization, and end-to-end reproduction on Nikkei and Livedoor Japanese news corpora. jaROTE improves precision over ja-timex while retaining high recall, reaches strong normalization accuracy, and remains affected by online-news ambiguity and discourse-level reference shifts.

  • 5 Evaluation: Temporal Reproduction: The evaluation measures extraction accuracy, LLM-comparative normalization, and end-to-end temporal reproduction.
  • 5 Evaluation: Temporal Reproduction: The Nikkei corpus contains 256 annotated omitted temporal expressions from 97 articles, while Livedoor contains 463 expressions from 270 sampled articles.
  • 5 Evaluation: Temporal Reproduction: Livedoor is not completely independent of development, so its evaluation measures performance across a different news medium rather than generalization to unseen data.
  • 5 Evaluation: Temporal Reproduction: jaROTE substantially improves extraction precision over ja-timex while maintaining high recall on both corpora.
  • 5 Evaluation: Temporal Reproduction: 0.9710 accuracy on Nikkei and 0.9431 accuracy on Livedoor are achieved by jaROTE in fixed-span normalization while processing articles in milliseconds rather than seconds.
  • 5 Evaluation: Temporal Reproduction: End-to-end performance remains strong, but Livedoor precision decreases because diverse online-news expressions produce false positives and normalization errors.
  • 5 Evaluation: Temporal Reproduction: Remaining errors include ambiguous month selection, numerical expressions resembling dates, and reference-time shifts that jaROTE may incorrectly anchor to the publication date.

6 Evaluation: Retrieval

The retrieval study tests whether publication-date-grounded temporal reproduction helps answer time-constrained queries under lexical and dense retrieval. Reproduction consistently benefits lexical retrieval, while dense-retrieval effects depend on the embedding model.

  • Evaluation setup: Recall@k evaluates whether the gold chunk appears among the top 1, 3, or 5 retrieved chunks under three text conditions.The conditions are Original, Original + publication date, and Reproduced.
  • Dataset: 212 valid time-constrained questions were constructed from 97 Nikkei articles containing omitted temporal expressions.Question candidates were generated from event attributes and gold reproduced temporal expressions, then manually filtered.
  • Results: For lexical retrieval with TF-IDF and BM25, the Reproduced condition is best for every recall@k value.This indicates improved lexical matching between time-constrained queries and documents.
  • Results: For ruri-v3-310m and plamo-embedding-1b, reproduction consistently yields the best dense-retrieval recall@k across all k.Dense-retrieval outcomes vary across embedding models rather than showing a uniform improvement.
  • Interpretation: Simply prepending the publication date sometimes improves retrieval but is less consistent than reproducing omitted expressions in the article body.The comparison uses the Original + publication date condition against Original and Reproduced.

7 Conclusion

The paper concludes that jaROTE provides fast, low-cost, publication-date-grounded temporal reproduction for Japanese news. Across Nikkei and Livedoor evaluations, reproduction improves time-constrained retrieval and can serve as preprocessing before search and RAG indexing.

  • jaROTE is a rule-based pipeline that reproduces omitted temporal expressions using publication date as the primary reference.
  • Experiments on Nikkei and Livedoor news corpora show fast, low-cost performance without LLM inference.
  • Publication-date-grounded reproduction improves time-constrained retrieval.
  • Temporal reproduction can serve as a lightweight preprocessing step for indexing Japanese news archives in search and RAG systems.

Limitations

The limitations constrain jaROTE’s temporal scope, contextual reasoning, and evaluation generalization. The system excludes expressions requiring additional reference resolution, uses local heuristics for some ambiguities, and has not been validated broadly outside the controlled study setting.

  • Scope of temporal expressions: jaROTE covers expressions extractable by ja-timex and concretely normalizable using publication date as the primary reference.It excludes deictic, seasonal, and expressions requiring comparison with an in-article reference time.
  • Complex processing: jaROTE lacks global propagation of newly introduced reference times and may fail when the effective reference shifts away from publication date.Ambiguous day-only and month-day expressions use a local future-oriented surface-pattern heuristic.
  • Evaluation coverage and generalization: Generalization to unseen data remains unverified because the Nikkei corpus informed analysis and rule design, while additional suppression rules were developed on Livedoor.Expression distributions may also differ across publishers, genres, and time periods.
  • Evaluation coverage and generalization: The retrieval evaluation uses a limited, controlled setup that does not capture all behaviors in end-to-end RAG systems.It focuses on time-constrained queries, recall@k, fixed chunking, and single-chunk gold evidence.

Ethical considerations

The paper discusses data-use constraints, reproducibility limits, and potential misuse of automated temporal normalization. It also documents implementation and evaluation details relevant to transparency, while noting that some code and auxiliary resources are unavailable.

  • Data rights and permitted use: The experiments use an open research corpus and additional data that cannot be redistributed because of contractual and licensing constraints.The authors state that the data were used within permitted terms and not obtained through unauthorized scraping or access-control circumvention.
  • Transparency and reproducibility: Processing code and certain auxiliary datasets were not released at submission, limiting full reproducibility.The paper specifies rules and procedures, reports results separately, and analyzes failure modes and sensitivity to assumptions.
  • Misuse potential: Automated normalization outputs may be misused if treated as authoritative without verification.The paper recommends preserving provenance, using well-scoped pipelines, and validating outputs in high-stakes contexts.
  • Use of generative AI tools: The study uses generative AI tools for manuscript drafting, polishing, or some synthetic experimental text, with author review of generated outputs.

D End-to-End Category Results

Table 8 reports jaROTE’s end-to-end temporal reproduction performance separately for each expression category.

  • Table 8 presents category-wise end-to-end temporal reproduction performance for jaROTE.
  • The reported evaluation concerns jaROTE’s end-to-end temporal reproduction performance.

E Prompts for constructing the evaluation QA dataset

The evaluation QA dataset is constructed by extracting explicit event attributes from masked temporal-expression contexts, generating constrained questions, manually filtering candidates, and enforcing chunk-based inclusion criteria.

  • E Prompts for constructing the evaluation QA dataset: GPT-5 mini performs information extraction and query generation in the two-stage QA dataset construction process.
  • E.1 Information extraction from sentences: The extraction stage uses a target sentence plus one preceding and one following sentence to identify six event attributes.
  • E.1 Information extraction from sentences: The six attributes are event/action, subject, object, location, quantity, and change.
  • E.2 Constructing an evaluation QA dataset: Generated questions target one of five answer types: who, what, where, amount, or delta.
  • E.2 Constructing an evaluation QA dataset: The information-extraction prompt masks the omitted temporal expression as [TIME] and requests one main event using explicit textual evidence.
  • E.2 Constructing an evaluation QA dataset: Extraction rules define event triggers, agents, patients, locations, amounts, and changes as structured fields, allowing null when an element is unstated.
  • E.2 Constructing an evaluation QA dataset: Query generation combines extracted attributes with gold reproduced temporal expressions and templates that constrain the requested answer.
  • E.2 Constructing an evaluation QA dataset: After manual filtering and random selection per temporal expression, queries are retained only when the target and adjacent sentences fit within one gold chunk, yielding 133 evaluation queries.
Loading 2609.09569v1…