Language model adaptation for fixed phrases by amplifying partial n‐gram sequences

Abstract

SUMMARYWe propose a method for creating an N-gram language model for use in a speech-operated question-answering system. We note that input questions to such a system frequently consist of an initial section, relating to the query topic, and a formulaic sentence final expression that is used in questions (a fixed phrase). While we are able to model the initial sections adequately using the target query newspaper corpus, we are not able to model the fixed phrases adequately with this data source. In this paper we frame the problem as one of adapting a language model created using a generic corpus to fixed phrases and propose a language model adaptation method that makes use only of a list of fixed phrases created by hand, rather than attempting the more difficult task of collecting an adaptation corpus. In the proposed method we determine which sections in the generic corpus correspond to N-gram sequences on the list of fixed phrases, and perform language model adaptation by amplifying the probabilities of those N-grams; this is equivalent to performing maximum a posteriori (MAP) estimation treating these partial N-gram sequences from the generic corpus itself as posterior information. We perform recognition experiments with spoken questions consisting of input to a question-answering system and confirm the effectiveness of the proposed method.

Relevant document retrieval using a spoken document

“…To this end, we developed a similarity considering "possibility of containing OOV words" as (9) This value means that the document has similar word distribution to the transcription and has many words that are not included in the automatic transcription of the spoken document. The correlation coefficient between this metric and the OOV cover rate was 0.45, which is fairly high considering that this value is calculated without knowing the real OOV set.…”

“…For example, Akiba et al proposed a method to use an utterance with fixed expressions as a retrieval key [9]. This method is an interface of information retrieval where a user utters a query instead of typing it using a keyboard.…”

On the Voice-Activated Question Answering

“…In [5], the authors investigate how to learn a language model that models both parts efficiently. The method proposed in this paper consists of adapting a N-gram language model learned using the document corpus by amplifying the N-gram sequences from a list of interrogative expressions defined by hand.…”

Language Modelization and Categorization for Voice-Activated QA

“…NEs are key elements for the search process [5], so misrecognition of a spoken NE can produce serious errors in the search results. Some works related to the NE recognition problem are [1,7,6]. …”