Section: New Results

Named Entity Recognition and Entity Linking

Participants : Rosa Stern, Benoît Sagot.

Identifying named entities is a widely studied issue in Natural Language Processing, because named entities are crucial targets in information extraction or retrieval tasks, but also for preparing further NLP tasks (e.g., parsing). Therefore a vast amount of work has been published that is dedicated to named entity recognition, i.e., the task of identification of named entity mentions (spans of text denoting a named entity), and sometimes types. However, real-life applications need not only identify named entity mentions, but also know which real entity they refer to; this issue is addressed in tasks such as knowledge base population with entity resolution and linking, which require an inventory of entities is required prior to those tasks in order to constitute a reference.

Improvements of the Aleda entity database

Within the Alexina framework, we develop since 2012 the entity database Aleda [124] , aimed at constituting such a reference. Aleda was first developed for French but is under development for English. Aleda is extracted automatically from Wikipedia and Geonames. It is used among others in the Sx Pipe processing chain and its NP named entity recognition, as well as in the Nomos named entity linking system.

In 2011, major efforts have been made for improving the coverage, precision and richness of the French Aleda: improvements in the tool for creating an XML almost-raw-text version of the wikipedia, new method for identifying and typing entities among wikipedia articles, based on infoboxes and wikipedia categories, richer database structure for storing more detailed information about each entity, and many other improvements. A paper about these advances has been sumbitted to LREC 2012.

Cooperation of symbolic and statistical methods for named entity recognition and typing

Named entity recognition and typing is achieved both by symbolic and probabilistic systems. We have performed an experiment [24] for making the rule-based system NP, Sx Pipe's high-precision named entity recognition system developed at Alpage on AFP news corpora and which relies on the Aleda named entity database, interact with LIANE, a high-recall probabilistic system developed by Frédéric Béchet and trained on oral transcriptions from the ESTER corpus. We have shown that a probabilistic system such as LIANE can be adapted to a new type of corpus in a non-supervized way thanks to large-scale corpora automatically annotated by NP. This adaptation does not require any additional manual anotation and illustrates the complementarity between numeric and symbolic techniques for tackling linguistic tasks.

Nomos, a statistical named entity linking system

For information extraction from news wires, entities such as persons, locations or organizations are especially relevant in a knowledge acquisition context. Through a process of named entity recognition and entity linking applied jointly, we aim at the extraction and complete identification of these relevant entities, which are meant to enrich textual content in the form of metadata. In order to store and access extracted knowledge in a structured and coherent way, we aim at populating an ontological reference base with these metadata. We have pursued our efforts in this direction, using an approach where NLP tools have early access to Linked Data resources and thus have the ability to produce metadata integrated in the Linked Data framework. In particular, we have studied how the entity linking process in this task must deal with noisy data, as opposed to the general case where only correct entity identification is provided.

We use the symbolic named entity recognition system NP, a component of Sx Pipe, and use it as a mention detection module. Its output is then processed through our entity linking system, which is based on a supervised model learnt from examples of linkings. Since our named entity recognition is not deterministic, as opposed to other entity linking tasks where the gold named entity recognition results are provided, it is configured to remain ambiguous and non-deterministic, i.e., its output preserves a number of ambiguities which are usually resolved at this level. In particular, no disambiguation is made in the cases of multiple possible mentions boundaries (e.g., {Paris}+{Hilton} vs. {Paris Hilton}). In order to cope with possible false mention matches, which should be discarded as linking queries, the named entity recognition output is made more ambiguous by adding a not-an-entity alternative to each mention detection. The entity linking module's input therefore consists in multiple possible readings of sentences. For each reading, this module must perform entity linking on every possible entity mention by selecting their most probable matching entity. Competing readings are then ranked according to the score of entities (or sequence of entities) ranked first in each of them. The reading with no entity should also receive a score in order to be included in the ranking. The motivation for this joint task lies in the frequent necessity of accessing contextual and referential information in order to complete an accurate named entity recognition; thus the part where named entity recognition usually resolves a number of ambiguities is left for the entity linking module, which uses contextual and referential information about entities.

We have realized a first implementation of our system, as well as experiments and evaluation results. In particular, when using knowledge about entities to perform entity linking, we discuss the usefulness of domain specific knowledge and the problem of domain adaptation.