Name
Abstract | ||
|---|---|---|
Nowadays there is an increasing interest on deploying spoken conversational agents to provide ubiquitous Question and Answering information to customers about corporate services and commercial products and supporting different users' devices such as PC desktops or mobile phones. Unfortunately, creating an accurate system requires a lot of handwork, where developers must consider several factors such as the performance of the ASR system, the presence of typos in the transcribed queries, the large number of possible variations to ask for the same information using different sentences, or the subtle differences that could exist between similar, but semantically different, questions. In this paper, we propose a methodology for quickly creating robust spoken-based conversational agents with very low resources. Our solution only requires few hand-made query samples, which are automatically expanded to deal with the use of different synonyms and wordings; next, spoken queries are automatically generated using a TTS system and then the audio files are corrupted to simulate different noise conditions and environments that the final users can experiment when they query the system using their voice with means of their mobile devices or by using a kiosk. Then, these audio files are subsequently transcribed using a general purpose ASR which produces an n-best list of recognized results that is first used to retrieve relevant documents and then re-ranked in order to select the final answer. Our tests on a set of 21 different topics proves that our proposal can get a 13% absolute better accuracy than a standard IR using an index with only in-domain answers and 6.3% better than a system including millions of negative out-of-domain candidate answers which is what it is expected for a scalable system. |
Year | Venue | Field |
|---|---|---|
2015 | Asia-Pacific Signal and Information Processing Association Annual Summit and Conference | Ask price,General purpose,Computer science,Corporate services,Signal-to-noise ratio,Robustness (computer science),Feature extraction,Mobile device,Natural language processing,Artificial intelligence,Interactive kiosk |
DocType | ISSN | Citations |
Conference | 2309-9402 | 0 |
PageRank | References | Authors |
0.34 | 15 | 3 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Luis Fernando D'Haro | 1 | 181 | 25.97 |
| Seok-Hwan Kim | 2 | 165 | 23.82 |
| Rafael E. Banchs | 3 | 566 | 63.64 |