Query Entity Recognition and Linking

"jaguar top speed" is about an animal or a car, and three words give a linker almost nothing to decide with. A news article mentioning Jaguar has paragraphs of context; a query has a mention, one or two context words, and a deadline, because linking has to finish before retrieval starts. Named entities are pervasive in web queries, which is what motivated treating query NER as its own task (Guo, Xu, Cheng and Li, 2009, Named Entity Recognition in Query, SIGIR), which is why entity linking sits in the query path of every major search engine: it is how a query is routed to a knowledge panel, a product catalogue or a structured answer.

Query intent classification decides what kind of answer a query wants; linking, which usually feeds it, works out which things the query names.

Two subtasks, one pass

Linking decomposes into mention detection, choosing which spans of the query refer to entities, and disambiguation, choosing a knowledge-base identifier for each span, or NIL if the entity is not in the base. A query of \(k\) terms has \(k(k+1)/2\) contiguous spans, so a five-word query has 15 candidate mentions, and most systems score all of them.

Guo et al. treated a query containing an entity as a triple: the entity string \(e\), the remaining context \(t\), and the entity's class \(c\). They scored

learning the class-conditional context distributions from query logs with a weakly supervised topic model. The factorisation shows why query NER is possible at all: "top speed" is far more probable after a car than an animal, and that is learnable from logs even though each query is tiny.

The anchor-text prior and a worked score

Most practical linkers start from statistics on Wikipedia anchor text. For a mention \(m\) and entity \(e\):

Link probability says whether a span is worth linking at all ("the" almost never is). Commonness is a prior over senses. Suppose, as an illustration rather than measured data, that "jaguar" appears as anchor text 10,000 times: 6,000 links to the car maker, 3,000 to the animal, 1,000 to other senses. Commonness gives the car 0.6 and the animal 0.3. A context model then reweights. If a similarity score between "top speed" and each entity yields likelihoods of 0.4 for the car and 0.5 for the animal, the unnormalised posteriors are \(0.6 \times 0.4 = 0.24\) and \(0.3 \times 0.5 = 0.15\), and the car still wins. The prior dominates unless the context is strongly discriminative, which is exactly what you want on head queries and exactly what goes wrong for the less common sense.

Blanco, Ottaviano and Meij built a production linker for Yahoo on this pattern. Their model scores candidate entities independently rather than jointly, which keeps linking at \(O(k^2)\) in the number of query terms; uses hashing and compression to shrink the anchor statistics; and adds distributional similarity between query words and entities for context. They reported more than 14 percent improvement over state-of-the-art baselines at sub-millisecond latency, at least two orders of magnitude faster than existing systems (Blanco, Ottaviano and Meij, 2015, Fast and Space-Efficient Entity Linking in Queries, WSDM).

Interpretations, not just links

Hasibi, Balog and Bratsberg argued that the evaluation tradition borrowed from document linking asks the wrong question of queries. They separated semantic mapping, returning a ranked list of entities related to the query, useful for suggestions and panels, from interpretation finding, returning one or more sets of entities that each form a coherent reading of the query (Hasibi, Balog and Bratsberg, 2015, Entity Linking in Queries: Tasks and Evaluation, ICTIR). "jaguar top speed" has two valid interpretations, and a system that must emit exactly one link has been asked to guess rather than to understand.

Neural linkers changed the machinery rather than the task. ELQ encodes the question and candidate entities with a bi-encoder and performs mention detection and linking in one pass, improving F1 over the previous best by 12.7 points on WebQSP and 19.6 on GraphQuestions (Li et al., 2020, Efficient One-Pass End-to-End Entity Linking for Questions, EMNLP, arXiv:2010.02413). It is more accurate on natural-language questions and much slower per query than an anchor-statistics linker. Whether a dense-retrieval pipeline needs explicit linking at all is itself contested: embeddings capture much of what a link would add, but they cannot trigger a knowledge panel or apply an exact entity filter.

When it breaks

The popular sense wins by default. Commonness priors encode the web's majority reading. A user at a zoo site searching "jaguar diet" can be sent to car content unless context or site-level priors override the global statistics.

New entities are NIL. A product launched this morning, a newly famous person or an internal project name has no anchor statistics. The linker either returns NIL or, worse, links to the nearest existing entity with a similar name.

Over-linking breaks phrases. "the who tour", "best buy hours" and "new york times square" contain spans that are entities in one segmentation and ordinary words in another. Mention detection interacts with the query segmentation covered in its own concept, and errors compound across the two.

Hard routing turns a link error into a recall failure. If a linked entity is used to filter results or pick a vertical, a wrong link removes the right documents entirely. Treating links as soft ranking features, with the confidence exposed, contains the damage.

Evaluation per link overstates quality. Mention-level F1 rewards getting easy entities right. What matters is whether the whole interpretation of the query is right, which is the stricter measure Hasibi et al. proposed.