Author:
Yi-Chung Lin, Chao-Chun Liang, Kuang-Yi Hsu, Chien-Tsung Huang, Shen-Yun Miao, Wei-Yun Ma, Lun-Wei Ku, Churn-Jung Liau, and Keh-Yih Su
Abstract:
This paper proposes a tag-based statistical framework to solve math word problems with understanding and reasoning. It analyzes the body and question texts into their associated tag-based logic forms, and then performs inference on them. Comparing to those rule-based approaches, the proposed statistical approach alleviates rules coverage and ambiguity resolution problems, and our tag-based approach also provides the flexibility of handling various kinds of related questions with the same body logic form. On the other hand, comparing to those purely statistical approaches, the proposed approach is more robust to the irrelevant information and could more accurately provide the answer. The major contributions of our work are: (1) proposing a tag-based logic representation such that the system is less sensitive to the irrelevant information and could provide answer more precisely; (2) proposing a unified statistical framework for performing reasoning from the given text.
Keywords:
Math Word Problem Solver, Machine Reading, Natural Language Understanding
Author:
Chien-Tsung Huang, Yi-Chung Lin, and Keh-Yih Su
Abstract:
This paper proposes a math operation (e.g., Summation, Addition, Subtraction, Multiplication, Division,
etc.) oriented approach to explain how the answers are obtained for math word problems. Based on the reasoning chain given by the inference engine, we search each math operator involved. For each math operator, we generate one sentence. Since explaining math operation does not require complicated syntax, we adopt a specific template to generate the text for each kind of math operator. To the best of our knowledge, this is the first explanation generation that is specifically tailored to the math word problems.
Keywords:
Explanation Generation, Math Word Problem Explanation, Machine Reading
Author:
Guan-Bin Chen and Hung-Yu Kao
Abstract:
The large amount of text on the Internet cause people hard to understand the meaning in a short limit time. Topic models (e.g. LDA and PLSA) has been proposed to summarize the long text into several topic terms. In the recent years, the short text media such as tweet is very popular. However, directly applies the transitional topic model on the short text corpus usually gating non-coherent topics. Because there is no enough words to discover the word co-occurrence pattern in a short document. The Bi-term topic model (BTM) has been proposed to improve this problem. However, BTM just consider simple bi-term frequency which cause the generated topics are dominated by common words. In this paper, we solve the problem of the frequent bi-term in BTM. Thus, we proposed an improvement of word co-occurrence method to enhance the topic models. We apply the word co-occurrence information to the BTM. The experimental result that show our PMI-β-BTM gets well result in the both of regular short news title text and the noisy tweet text. Moreover, there are two advantages in our method. We do not need any external data and our proposed methods are based on the original topic model that we did not modify the model itself, thus our methods can easily apply to some other existing BTM based models.
Keywords:
Short Text, Topic Model, Document Clustering, Document Classification
Author:
Kai-Wun Shih, Kuan-Yu Chen, Shih-Hung Liu, Hsin-Min Wang and Berlin Chen
Abstract:
The rapidly increasing availability of multimedia associated with spoken documents on the Internet has prompted automatic spoken document summarization to be an important research subject. Thus far, the majority of existing work has focused on extractive spoken document summarization, which selects salient sentences from an original spoken document according to a target summarization ratio and concatenates them to form a summary concisely, in order to convey the most important theme of the document. On the other hand, there has been a surge of interest in developing representation learning techniques for a wide variety of natural language processing (NLP)-related tasks. However, to our knowledge, they are largely unexplored in the context of extractive spoken document summarization. With the above background, this study explores a novel use of both word and sentence representation techniques for extractive spoken document summarization. In addition, three variants of sentence ranking models building on top of such representation techniques are proposed. Furthermore, extra information cues like the prosodic features extracted from spoken documents, apart from the lexical features, are also employed for boosting the summarization performance. A series of experiments conducted on the MATBN broadcast news corpus indeed reveal the performance merits of our proposed summarization methods in relation to several state-of-the-art baselines.
Keywords:
Spoken Document, Extractive Summarization, Word Representation, Sentence Representation, Prosodic Feature
Author:
Ting-Hao Chang, Hsiao-Tsung Hung, Kuan-Yu Chen, Hsin-Min Wang and Berlin Chen
Abstract:
The performance of an automatic speech recognition (ASR) system often deteriorates sharply due to the interference from varying environmental noise. As such, the development of effective and efficient robustness techniques has long been a challenging research subject in the ASR community. In this article, we attempt to obtain noise-robust speech features through modulation spectrum processing of the original speech features. To this end, we explore the use of nonnegative matrix factorization (NMF) and its extensions on the magnitude modulation spectra of speech features so as to distill the most important and noise-resistant information cues that can benefit the ASR performance. The main contributions include three aspects: 1) we leverage the notion of sparseness to obtain more localized and parts-based representations of the magnitude modulation spectra with fewer basis vectors; 2) the prior knowledge of the similarities among training utterances is taken into account as an additional constraint during the NMF derivation; and 3) the resulting encoding vectors of NMF are further normalized so as to further enhance their robustness of representation. A series of experiments conducted on the Aurora-2 benchmark task demonstrate that our methods can deliver remarkable improvements over the baseline NMF method and achieve performance on par with or better than several widely-used robustness methods.
Keywords:
Speech Recognition, Language Model, Concept Information, Model Adaptation
Author:
Po-Hsuan Chen and Chi-Chun Lee
Abstract:
Traditional way of conducting analyses of human behaviors is through manual observation. For example in couple therapy studies, human raters observe sessions of interaction between distressed couples and manually annotate the behaviors of each spouse using established coding manuals. Clinicians then analyze these annotated behaviors to understand the effectiveness of treatment that each couple receives. However, this manual observation approach is very time consuming, and the subjective nature of the annotation process can result in unreliable annotation. Our work aims at using machine learning approach to automate this process, and by using signal processing technique, we can bring in quantitative evidence of human behavior. Deep learning is the current state-of-art machine learning technique. This paper proposes to use stacked sparse autoencoder (SSAE) to reduce the dimensionality of the acoustic-prosodic features used in order to identify the key higher-level features. Finally, we use logistic regression (LR) to perform classification on recognition of
high and low rating of six different codes. The method achieves an overall accuracy of 75% over 6 codes (husband’s average accuracy of 74.9%, wife’s average accuracy of 75%), compared to the previously-published study of 74.1% (husband’s average accuracy of 75%, wife’s average accuracy of 73.2%) (Black
et al., 2013), a total improvement of 0.9%. Our proposed method achieves a higher classification rate by using much fewer number of features (10 times less than the previous work (Black
et al., 2013)).
Keywords:
Deep Learning, Stacked Autoencoders, Couple Therapy, Human Behavior Analysis, Emotion Recognition