Standard information retrieval (IR) systems operate on a simple premise: if a document is relevant to a query, it likely shares keywords with that query. However, complex reasoning often requires connecting disparate facts that have no direct lexical overlap. This "lexical gap" or "lexical chasm" is the primary bottleneck in multi-hop question answering (QA) and fact-checking.
When a question requires multiple steps—such as "Which city was the director of Inception born in?"—a search engine might easily find that Christopher Nolan directed Inception. But the second step (finding Nolan's birthplace) requires a query for "Christopher Nolan," a term that may not appear in the original user question.
The Failure of Single-Step Retrieval
Traditional keyword-based methods like BM25 fail in these scenarios because the "bridge" information is hidden. As noted in Multi-Hop Paragraph Retrieval for Open-Domain Question Answering by Yair Feldman, the final supporting document often shares zero significant terminology with the initial query.
This challenge is quantified in the HoVer dataset, introduced by Yichen Jiang. By requiring models to reason across up to four Wikipedia articles, HoVer reveals a massive performance gap: while humans can navigate these lexical chasms with 81% accuracy, standard AI models often struggle to reach 15%, precisely because they cannot "see" the next hop without a direct keyword match.
Identifying the Bridge Entity
To solve this, researchers have moved toward "bridge reasoning." The goal is to identify the intermediate entity—the "missing link"—within the first set of retrieved results and use it to launch a subsequent search.
Wenhan Xiong and colleagues proposed a Bridge Reasoner that treats the first retrieved paragraph as a context for finding a new search term. Instead of just looking for the answer, the model looks for the "bridge" that will lead to the answer. Similarly, Fan Luo introduced STEP, which uses graph theory (specifically Steiner trees) to identify bridge phrases in an unsupervised manner, expanding the query to include terms that logically connect the question to potential evidence.
Iterative Rethinking and Query Reformulation
Rather than a single search, modern systems use an iterative "retrieve-and-reformulate" loop. Yair Feldman developed the MUPPET framework, which uses a neural reformulation layer to "re-think" the search query after each piece of evidence is found. This allows the system to pivot from the original question to a more specific query based on the intermediate entities discovered.
This iterative process mimics human research:
- Search for the initial topic.
- Extract a new lead (the bridge entity) from the results.
- Reformulate the search to follow that lead.
Cognitive Graphs and Structural Reasoning
Beyond simple iteration, some researchers argue that multi-hop reasoning requires a dual-process approach: one system to extract information and another to reason over the structure of that information.
The CogQA framework by Chang Zhou implements this using a "Cognitive Graph." It uses a transformer (System 1) to pull potential entities from text and a Graph Neural Network (System 2) to compute the relationships between them. This allows the model to maintain a "frontier" of reasoning, tracking which entities have been explored and which might lead to the final answer.
A similar hybrid approach was used by Weibin Li and the Baidu PGL team to win the TextGraphs 2020 Shared Task. Their "ERNIESage" architecture performs token-level interactions during graph aggregation, ensuring that the semantic nuances of the language model are preserved while the GNN handles the multi-step logical connectivity.
Constructing Reasoning Chains
The ultimate goal of these methods is the construction of a "reasoning chain"—a sequence of facts where each link provides the context for the next. Weiwen Xu and the CGR team demonstrated that these chains can be built automatically using Abstract Meaning Representation (AMR). By converting text into semantic graphs, the model can identify "indirect" facts that are logically necessary but lexically distant, effectively building a map across the lexical gap.