Named entity disambiguation

Wikipedia has an article on: Named entity disambiguation

Formalization

Mapping $\Gamma: M \rightarrow E$ where $M = \{m_1, m_2,..., m_k\}$ are name mentions, characterized by name m.S, context m.C and document m.D and $E = \{e_1, e_2,..., e_n\}$ are entities in a knowledge base.

Approaches

Individual

This approach assumes that the disambiguation of different mentions are independent. It solves the optimization problem:

$\Gamma^*_{local} = \arg \max_{\Gamma} \sum^k_{i=1}\phi(m_i, \gamma_i)$

where $\gamma_i = \Gamma(m_i), \phi(m_i, \gamma_i)$ measure the compatibility between the mention $m_i$ and the entity chosen by $\Gamma$.

Features used in this approach may come from surrounding words, sentence or paragraph (e.g. DBpedia Spotlight).

• Bag of Words (BoW), cosine similarity: Mihalcea & Csomai
• BoW + Categories: Cucerzan (2007), Bunescu & Pasca, Fader et al.

Some employed algorithms:

• Classification: Zhang et al. and Mihalcea & Csomai.
• Learning-to-rank techniques: Zheng et al., Dredze et al. and Zhou et al.

TODO: a survey: http://www.jair.org/media/4129/live-4129-7870-jair.pdf

Relational heuristics

The idea was that the referent entity of a name mention should be coherent with its unambiguous contextual entities.

• Medelyan et al. determined the compatibility using the semantic relatedness between the candidate entity and the contextual entities.
• Milne and Witten extended the method of Medelyan et al. by adopting learning-based techniques to balance the semantic relatedness, the commonness and the context quality.

Collective

Main article: Collective named entity disambiguation

Some algorithms disambiguate based on collective topic coherence only, for example BabelFy.

Individual+Collective

An example of referent graph from Han et al. (2011)

Collective approaches combine local information with global inference to achieve better results. Given a coherence function $\psi$ that assigns a number for each assignment of the whole document, the optimization goal is now:

$\Gamma^*_{local} = \arg \max_{\Gamma} \left[\sum^k_{i=1}\phi(m_i, \gamma_i)\right] + \psi(\Gamma)$

For some choices of coherence function, the problem has been proved to be NP-Hard.

Some references:

• Pair-wise interdependence: Kulkarni et al..
• Directly connected entities: Han et al., Hoffart et al. (2011)
• All semantically related entities: Guo & Barbosa (2014)

Reranking

• Language model: Dalvi et al. (2014)^{(doesn't fit neatly in individual-collective spectrum?)}

Features

Local features

Probability of context

Encodes the context of the named entities, i.e. $P(c|e)$, where c is context of the named entity e. For a specific context, a higher probability will be assigned to the named entity which frequently appears with that context. Han and Sun (2011) proposed an entity context model to estimate the distribution $P(c|e)$ by encoding the context of an entity e in a unigram language model. They define the context as the surrounding window of 50 terms, and used this formula to find the entity context probability:

$P(c|e) = P_e(t_1)P_e(t_2)...P_e(t_n)$

where $P_e(t) = \frac{Count_e(t)}{\sum_tCount_e(t)}$, and $Count_e(t)$ is the frequency of occurrence of term t in the context of the named entity e.

Evaluation

Datasets

• Wikipedia articles: traditionally used but Kulkarni et al. (2009) pointed out that it is unsuitable to the evaluation of high-recall entity linking tasks because it annotates name mentions very sparsely (only the important name mentions are annotated).
• TAC 2009: focuses on individual EL tasks in different documents, unsuitable for our collective EL settings.
• IITB dataset: a set of documents (107 documents in total) collected from the web sites belonging to a handful of domains. For each document, its name mentions’ referent entities in Wikipedia are manually annotated to be as exhaustive as possible. In total, 17,200 name mentions are annotated, 161 name mentions per document on average.
• KORE50 (Hoffart et al., 2012), which consists of 50 short English sentences (mean length of 14 words) with a total number of 144 mentions manually annotated using YAGO2, for which a Wikipedia mapping is available. This dataset was built with the idea of testing against a high level of ambiguity for the EL task.
• AIDA-CoNLL (Hoffart et al., 2011), which consists of 1392 English articles, for a total of roughly 35K named entity mentions annotated with YAGO concepts separated in development, training and test sets.

Evaluation criteria

• F-score

Application

• Information extraction
• Knowledge base population
• Semantic search

See also

• Resources for entity linking