課程名稱︰自然語言處理 課程性質︰系內選修 課程教師︰陳信希 開課學院：電資學院 開課系所︰資訊工程學系 考試日期（年月日）︰2016/04/21 考試時限（分鐘）：180 mins 試題 : 01. Machine translation (MT) is one of practical NLP applications. The development of MT systems has a long history, but still has space to improve. Please address two linguistic phenomena to explain why MT systems are challenging. (10pts) 02. An NLP system can be implemented in a pipeline, including modules of morphological processing, syntactic analysis, semantic interpretation and context analysis. Please use the following news story to describe the concepts behind. You are asked to mention one task in each module. (10pts) 這場地震可能影響日相安倍晉三的施政計畫。安倍十八日說，消費睡調漲的 計畫不會改變。 03. Ambiguity is inherent in natural language. Please describe why ambiguity may happen in each of the following cases. (10pts) (a) Prepositional phrase attachment. (b) Noun-noun compound. (c) Word: bass 04. Why the extraction of multiword expressions is critical for NLP applications? Please propose a method to check if an extracted multiword expression meets the non-compositionality criterion, (10pts) 05. Mutual information and likelihood ratio are commonly used to find collocations in a corpus. Please describe the ideas of these two methods. (10pts) 06. Emoticons are commonly used in social media. They can be regarded as a special vocabulary in a language. Emoticon understanding is helpful to understand the utterances in an interaction. Please propose an "emoticon" embedding approach to represent each emoticons as a vector, and find the most 5 relevant words to each emoticon. (10pts) 07. To deal with unseen n-grams, smoothing techniques are adopted in conventional language modeling approach. They are applied to n-grams to reallocate probability mass from observed n-grams to unobserved n-grams, producing better estimates for unseen data. Please show a smoothing technique for the conventional language model, and discuss why neural network language model (NNLM) can achieve better generalization for unseen n-grams. (10pts) 08. In HMM learning, we aim at inferring the best model parameters, given a skeletal model and an observation sequence. The following two equations are related to compute the state transition probabilities. Σ_{t=1}^{T-1} ξ_t(i, j) \hat{a}_{ij} = --------------------------------------- Σ_{t=1}^{T-1} Σ_{j=1}^{N} ξ_t(i,j) α_t(i) a_{ij} b_j(o_{t+1}) β_{t+1}(j) ξ_t(i, j) = ----------------------------------------- α_T(q_F) Please answer the following questions. (10pts) (a) Intuitively, we can generate all possible paths for the given observation sequence, and compute total times of a transition which the observation passes. Which part in the above equations avoids the generation of all possible paths? (b) Which part in the above equations is related to prorate count to estimate the transition probability of a transition? 09. Many NLP problems can be cast as a sequence labelling problem. Part of speech tagging is a typical example. Given a model and an observation sequence, we aim at finding the most probable state sequence. Please explain why this process is called a decoding process. In addition, please give another application which can be also treated as a sequence labelling problem. (10pts) 10. What is long-distance dependencies or unbounded dependencies? Why such kinds of linguistic phenomena are challenging in NLP? (10pts) 11. Part of speech tagging can be formulated in the following two alternatives: Model 1: \hat{t}_1^n = argmax_{t_1^n} Π_{i=1}^n P(w_i|t_i) P(t_i|t_{i-1}) Model 2: \hat{t}_1^n = argnax_{t_1^n} Π_{i=1}^n P(t_i|w_i, t_{i-1}) Please answer the following questions. (10pts) (a) Which one is discriminative model? (b) Which one can introduce more features? (c) Which one can use Viterbi algorithm to improve the speed? (d) Which one is derived on the basis of Bayes rule? 12. The following parsing tree is selected from Chinese Treebank 8.0. What NP and VP rules can be extracted from this parsing tree to form parts of a treebank grammer? (10pts) ( (IP (IP (NP-SBJ (NN 建築)) | (VP (VC 是) | | (NP-PRD (CP-APP (IP (NP-SBJ (-NONE- *pro*)) | | | | (VP (VV 開發) | | | | | (NP-PN-OBJ (NR 浦東)))) | | | | (DEC 的)) | | | (QP (CD 一) | | | (CLP (M 項))) | | | (ADJP (JJ 主要)) | | | (NP (NN 經濟) | | | (NN 活動))))) | (PU 。) | (IP (NP-SBJ (-NONE- *pro*)) | (VP (DP-TMP (DT 這些) | | | (CLP (M 年))) | | (VP (VE 有) | | (IP-OBJ (NP-SBJ (NP (QP (CD 數百) | | | | | (CLP (M 家))) | | | | | (NP (NN 建築) | | | | | (NN 公司))) | | | | (PU 、) | | | | (NP (QP (CD 四千餘) | | | | | (CLP (M 個))) | | | | | (NP (NN 建築) | | | | | (NN 工地)))) | | | (VP (VV 遍布) | | | | (PP-LOC (P 在) | | | | | (LCP (NP (DP (DT 這) | | | | | | (CLP (M 片))) | | | | | | (NP (NN 熱土))) | | | | | (LC 上)))))))) | (PU 。)) ) --

※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 140.112.30.49 ※ 文章網址: https://webptt.com/m.aspx?n=bbs/NTU-Exam/M.1467630306.A.186.html