Unsupervised knowledge acquisition for Extracting Named Entities from speech

Abstract

International audienceThis paper presents a Named Entity Recognition (NER) method dedicated to process speech transcriptions. The main principle behind this method is to collect in an unsupervised way lexical knowledge for all entries in the ASR lexicon. This knowledge is gathered with two methods: by automatically extracting NEs on a very large set of textual corpora and by exploiting directly the structure contained in the Wikipedia resource. This lexical knowledge is used to update the statistical models of our NER module based on a mixed approach with generative models (Hidden Markov Models-HMM) and discriminative models (Conditional Random Field-CRF). This approach has been evaluated within the French ESTER 2 evaluation program and obtained the best results at the NER task on ASR transcripts

Combining transcription-based and acoustic-based speaker identifications for broadcast news

“…The proposed systems are evaluated using either manual or automatic segmentations and transcripts. In the two cases, the named entities are automatically detected using [1]. The progress done on all elementary tasks (speaker diarization, speech recognition and the use of belief functions) decreases the speaker identification error rate from 75.15% [8] to 60.59%.…”

“…Our acoustic speaker identification system is based on the wellknown UBM/GMM approach developed in the ALIZE toolkit 1 . First, a pre-processing concerning training, development and test sets is necessary.…”

Toponym recognition in custom-made map titles

“…Toponym identification in such a corpus is a difficult task. Most often, processes for NE identification involve either a symbolic approach based on local grammars (Bontcheva, Dimitrov, Maynard, Tablan, & Cunningham, 2002;Friburger & Maurel, 2004;Poibeau, 2003) or a statistical approach based on automatic learning, or hybrid systems such as Béchet and Charton (2010), and Leidner (2007). Our approach is guided by corpus specifications: located information, spelling variations and a very large number of toponyms.…”

Concept-based embeddings for natural language processing

“…Neighboring words and POS tags: They are acknowledged to be efficient for NER [73,74]. In fact, we use these features not only for training our word embeddings, but also for training the baseline model of NER (Section 3.2.1).…”

“…Nevertheless the adaptation of NER methods to conversational speech remains challenging due to, for example, case insensitivity, lack of punctuations, un-grammatical structure, repetition, and presence of disfluencies inherent to conversations. In addition, there is not much spoken data annotated with named entities to cover the huge variety of named entity instances likely occurring in speech, and simply increasing the amount of manual annotation is not realistic for reasons of cost, evolution of new spoken terms and diversity.Several works on NER from spoken contents have already explored the use of external resources like online gazetteers[73] and Wikipedia[74] to overcome the lack of annotations. Gazetteers, for instance, have successfully boosted NER performance for given entities (e.g., Location), but do not convey the information related to the context words surrounding the entity names that are also important for NER.…”