Zhiguo Wang. IBM Watson

Understanding and modeling the sentences.

Abstract

Sentence modeling is a crucial procedure for question answering. In this presentation, I will talk about applying deep learning methods for sentence modeling, including sentence clustering, sentence matching, sentence classification and sequential labeling. I will also present how to apply these technologies into real world question answering systems.

Overview

factorid:

similarity / overlap: question & passages

architecture:

Q -> lucene search(wiki) -> passage generator -> passage ranker(DL, sen-matching) -> chunk extractor(DL, seq labeling) -> feature generator(human) -> chunk ranker(DL)

chunk: sentence contains the correct answer

others

all of other non-factoid q, use FAQ-based

FAQ: classify -> similarity -> ensemble

sen-clusering

use companies’ custem servies log, to build FAQ training set (with human intervering)

traditional: bag-of-words / TF-IDF, k-means
drawback:

  • BoW / TF-IDF intrinsic problems
  • human knowledge to clustering better than pure unsupervised

so semi-supervised. let the clustering be like what we want

modeling: CNN (for text representation ??) / LSTM

loss func: (k-means term) + (max labeled margin term)

sen-matching

motivation:

  • lexically different representation for the same sentence
  • not only word-word level but also phrase-level and syntax-level

model: blabla

sen-classification

???

seq labeling

chunk -> seq labeling -> answer begin, w, w

better than NER, avoiding the large ranking set