D+3 

Today is the fourth day of the project.

Jason, Grace, and Jane make the recipe of Related Work, the Steve's style and Jason attempt to recreate the Remus Rome Coffee for the first time.

Before the noon meeting at a nearby MOS to discuss the dishes Jane and Grace prepared for Jason to examine and offer suggestion ...

1. With making a Rome Coffee in mind, Jason brought from home some coffee and a lemon and a lemon juicer and bough an ice cream cone. He did make a Rome coffee almost as good as 旅沐咖啡 Coffee Remus where Rome Coffee was invented.

2. Initially, we were going to meet at MacDonald opposite the Mackey Memorial Hospital, Hsinchu. But Grace changed it to MOS at the last minutes. M was packed with more people than usual.

3. Jason sugguested that we are going to have only seven paragraph instead of the 10 paragraghs of the original recipe, Why?

Three paragraphs (#1, #2 and #10) are essential and will stay, plus 4 paragraphs for different approaches or related issues.

We will let go 1. a paragraph that talks about the next step of the research (answer extraction after document retrieval) is not relevant for us. 2. A paragraph about independent discovery which we do not have 3, Yet another paragraph for briefity consideration 

4. With that Jason wrote a ONE-SHOT GENERATION prompt to get a draft of the RW.

The prompt goes like this:

I am writing "related work" for my problem statement: 

"""<paste your problem statement here>""" 

I want to mirrow the related work of a model paper with a different problem statement: 

 """ Problem Statement We are given a general purpose search engine SE (e.g., Google) that operates over a document collection (e.g., the Web), and a natural language question QNL. Our goal is to retrieve a set of documents via SE that are likely to contain an answer to QNL. For this, we transform QNL into a set of queries, q1, . . . , qm, such that the top K documents returned by SE for each of the queries are likely to contain an answer to the original question QNL. 

Related Work 

Paragraph #1 Question Answering (QA) has been an area of active research. Recently, the state-of-the-art in QA research has been represented in the Text Retrieval Evaluation Conference (TREC) question answering track evaluations [Voorhees 2001], which involve retrieving a short (50 bytes long) answer to a set of test questions. In our work we address an aspect of question answering that was not a direct focus of the TREC QA track. We also consider a more general class of questions, where the answers may not be short, precise facts, and the user might be interested in multiple answers (e.g., consider the question “What are ways people can be motivated?”). 

 Paragraph #2 More specifically, we focus on the first part of the question answering task, namely, retrieving promising documents that are likely to contain an answer to a given question. Retrieving documents to match an information need has long been an active topic of information retrieval (IR) research. The body of the IR research most closely related to our work focuses on automatically expand- ing queries in order to retrieve additional relevant documents (e.g., Rocchio [1971] and Mitra et al. [1998]). An interesting approach presented by Xu and Croft [2000] describes how to automatically expand a query based on the co- occurrence of query terms with other terms in the top-ranked documents for the original query. In general, automatic query expansion systems modify queries at run-time on a query-by-query basis, using documents returned in response to the original query. In contrast, we will show how to derive a set of query transformations that work best for each type of question. 

 Paragraph #3 Most of the question answering systems presented in the TREC 8 and 9 QA track evaluations used the general approach of retrieving documents or passages that are similar to the original question with variations of standard TF-IDF term weighting schemes [Salton 1989]. The most promising passages are chosen using heuristics and/or hand-crafted regular expressions. This ap- proach is not optimal because documents that are similar to the question are initially retrieved. However, the user is actually looking for documents contain- ing an answer and these documents may contain few of the terms used to ask the original question. Using effective queries is particularly important when document retrieval is expensive or limited to a certain number of documents, as is the case with Web search engines. 

 Paragraph #4 More recently, some of the traditional QA systems have begun to modify queries to improve the chance of retrieving answers. Harabagiu et al. [2000] describe a system that transforms questions using a hierarchy of question types. The hierarchy is built semi-automatically using a bootstrapping tech- nique. Other systems (e.g., Hovy et al. [2000]) used WordNet [Miller 1995] to expand queries with word synonyms and hypernyms to retrieve additional rel- evant documents. However, this approach has not been shown to significantly improve the quality of the retrieved document set. Additionally, the AskMSR system [Brill et al. 2001] used manually-crafted, question-to-query transfor- mations to focus on promising documents. In a question answering document retrieval approach closer to our work, Ittycheriah et al. [2000] used LCA [Xu and Croft 2000], a form of automatic query expansion, to expand the initial query with terms selected from encyclopedia passages matching the initial question. 

 Paragraph #5 Recent work has been done on reranking the candidate set of documents, or phrases, with the goal of locating those most likely to contain answers to a ques- tion. In contrast, we focus on learning how to generate good queries to create a quality candidate set of documents, which can then be postprocessed by such reranking techniques. For example, a machine-learning and term-correlation based approach for reranking candidate answer phrases, or documents, is de- scribed in Berger et al. [2000] and is similar in spirit to our work. However, this work does not address learning effective queries for specific Web search en- gines. Similarly, Mann [2002] describes a method for learning to locate an exact answer in a passage by using co-occurrence statistics derived from question- answer pairs used in trivia games. We use a similar training set (FAQ question- answer pairs in our case) for the different purpose of deriving effective queries to retrieve promising documents. Prager et al. [2002] show how to automati- cally identify the semantic type of the expected answer to a question. All these methods can be used to process the set of documents retrieved by Tritus to extract short and precise answers from them. 

 Paragraph #6 In the research area of specialized search engines, Glover et al. [2001] present a method for learning topic-specific query modifications. The system starts with a set of example documents on the topic of interest, and trains a classifier to recognize such documents. Then, salient features are selected from the example documents and tested on the search engine of interest by retrieving documents using the candidate queries. Our method, which we will describe in the next section, uses a similar approach for a different setting of question answering. 

 Paragraph #7 In contrast to the previous research in question answering and query expan- sion, we present a system that automatically learns multiple query transforma- tions, optimized specifically for each search engine, with the goal of maximizing the probability of an information retrieval system returning documents that contain answers to a given question. We exploit the inherent regularity and power of natural language by transforming natural language questions into sets of effective search engine queries. """