TIQ: A Benchmark for Temporal Question Answering with Implicit Time Constraints

Temporal question answering (QA) involves explicit (e.g., "...before 2024") or implicit (e.g., "...during the Cold War period") time constraints. Implicit constraints are more challenging; yet benchmarks for temporal QA largely disregard such questions. This shortcoming spans three aspects. First, implicit questions are scarce in existing benchmarks. Second, questions are created based on hand-crafted rules, thus lacking diversity in formulations. Third, the source for answering is either a KB or a text corpus, disregarding cues from multiple sources. We propose a benchmark, called TIQ (Temporal Implicit Questions), based on novel techniques for constructing questions with implicit time constraints. First, questions are created automatically, with systematic control of topical diversity, timeframe, head vs. tail entities, etc. Second, questions are formulated using diverse snippets and further paraphrasing by a large language model. Third, snippets for answering come from a variety of sources including KB, text, and infoboxes. The TIQ benchmark contains 10,000 questions with ground-truth answers and underlying snippets as supporting evidence.