1{2 "document_id": "lb_qasper_37f06d3a413beef1661730ae1049e7adccd6308032ae1fc5",3 "family": "public_qasper",4 "split": "test",5 "text": "Introduction\nFrom a group of small users at the time of its inception in 2009, Quora has evolved in the last few years into one of the largest community driven Q&A sites with diverse user communities. With the help of efficient content moderation/review policies and active in-house review team, efficient Quora bots, this site has emerged into one of the largest and reliable sources of Q&A on the Internet. On Quora, users can post questions, follow questions, share questions, tag them with relevant topics, follow topics, follow users apart from answering, commenting, upvoting/downvoting etc. The integrated social structure at the backbone of it and the topical organization of its rich content have made Quora unique with respect to other Q&A sites like Stack Overflow, Yahoo! Answers etc. and these are some of the prime reasons behind its popularity in recent times. Quality question posting and getting them answered are the key objectives of any Q&A site. In this study we focus on the answerability of questions on Quora, i.e., whether a posted question shall eventually get answered. In Quora, the questions with no answers are referred to as “open questions”. These open questions need to be studied separately to understand the reason behind their not being answered or to be precise, are there any characteristic differences between `open' questions and the answered ones. For example, the question “What are the most promising advances in the treatment of traumatic brain injuries?” was posted on Quora on 23rd June, 2011 and got its first answer after almost 2 years on 22nd April, 2013. The reason that this question remained open so long might be the hardness of answering it and the lack of visibility and experts in the domain. Therefore, it is important to identify the open questions and take measures based on the types - poor quality questions can be removed from Quora and the good quality questions can be promoted so that they get more visibility and are eventually routed to topical experts for better answers.\nCharacterization of the questions based on question quality requires expert human interventions often judging if a question would remain open based on factors like if it is subjective, controversial, open-ended, vague/imprecise, ill-formed, off-topic, ambiguous, uninteresting etc. Collecting judgment data for thousands of question posts is a very expensive process. Therefore, such an experiment can be done only for a small set of questions and it would be practically impossible to scale it up for the entire collection of posts on the Q&A site. In this work, we show that appropriate quantification of various linguistic activities can naturally correspond to many of the judgment factors mentioned above (see table 2 for a collection of examples). These quantities encoding such linguistic activities can be easily measured for each question post and thus helps us to have an alternative mechanism to characterize the answerability on the Q&A site.\nThere are several research works done in Q&A focusing on content of posts. BIBREF0 exploit community feedback to identify high quality content on Yahoo! Answers. BIBREF1 use textual features to predict answer quality on Yahoo! Answers. BIBREF2 , investigate predictors of answer quality through a comparative, controlled field study of user responses. BIBREF3 study the problem of how long questions remain unanswered. BIBREF4 propose a prediction model on how many answers a question shall receive. BIBREF5 analyze and predict unanswered questions on Yahoo Answers. BIBREF6 study question quality in Yahoo! Answers.\nDataset description\nWe obtained our Quora dataset BIBREF7 through web-based crawls between June 2014 to August 2014. This crawling exercise has resulted in the accumulation of a massive Q&A dataset spanning over a period of over four years starting from January 2010 to May 2014. We initiated crawling with 100 questions randomly selected from different topics so that different genre of questions can be covered. The crawling of the questions follow a BFS pattern through the related question links. We obtained 822,040 unique questions across 80,253 different topics with a total of 1,833,125 answers to these questions. For each question, we separately crawl their revision logs that contain different types of edit information for the question and the activity log of the question asker.\nLinguistic activities on Quora\nIn this section, we identify various linguistic activities on Quora and propose quantifications of the language usage patterns in this Q&A site. In particular, we show that there exists significant differences in the linguistic structure of the open and the answered questions. Note that most of the measures that we define are simple, intuitive and can be easily obtained automatically from the data (without manual intervention). Therefore the framework is practical, inexpensive and highly scalable.\nContent of a question text is important to attract people and make them engage more toward it. The linguistic structure (i.e., the usage of POS tags, the use of Out-of-Vocabulary words, character usage etc.) one adopts are key factors for answerability of questions. We shall discuss the linguistic structure that often represents the writing style of a question asker.\nIn fig 1 (a), we observe that askers of open questions generally use more no. of words compared to answered questions. To understand the nature of words (standard English words or chat-like words frequently used in social media) used in the text, we compare the words with GNU Aspell dictionary to see whether they are present in the dictionary or not. We observe that both open questions and answered questions follow similar distribution (see fig 1 (b)). Part-of-Speech (POS) tags are indicators of grammatical aspects of texts. To observe how the Part-of-Speech tags are distributed in the question texts, we define a diversity metric. We use the standard CMU POS tagger BIBREF8 for identifying the POS tags of the constituent words in the question. We define the POS tag diversity (POSDiv) of a question $q_i$ as follows: $POSDiv(q_i) = -\\sum _{j \\in pos_{set}}p_j\\times \\log (p_j)$ where $p_j$ is the probability of the $j^{th}$ POS in the set of POS tags. Fig 1 (c) shows that the answered questions have lower POS tag diversity compared to open questions. Question texts undergo several edits so that their readability and the engagement toward them are enhanced. It is interesting to identify how far such edits can make the question different from the original version of it. To capture this phenomena, we have adopted ROUGE-LCS recall BIBREF9 from the domain of text summarization. Higher the recall value, lesser are the changes in the question text. From fig 1 (d), we observe that open questions tend to have higher recall compared to the answered ones which suggests that they have not gone through much of text editing thus allowing for almost no scope of readability enhancement.\nPsycholinguistic analysis:\nThe way an individual talks or writes, give us clue to his/her linguistic, emotional, and cognitive states. A question asker's linguistic, emotional, cognitive states are also revealed through the language he/she uses in the question text. In order to capture such psycholinguistic aspects of the asker, we use Linguistic Inquiry and Word Count (LIWC) BIBREF10 that analyzes various emotional, cognitive, and structural components present in individuals' written texts. LIWC takes a text document as input and outputs a score for the input for each of the LIWC categories such as linguistic (part-of-speech of the words, function words etc.) and psychological categories (social, anger, positive emotion, negative emotion, sadness etc.) based on the writing style and psychometric properties of the document. In table 1 , we perform a comparative analysis of the asker's psycholinguistic state while asking an open question and an answered question.\nAskers of open questions use more function words, impersonal pronouns, articles on an average whereas asker of answered questions use more personal pronouns, conjunctions and adverbs to describe their questions. Essentially, open questions lack content words compared to answered questions which, in turn, affects the readability of the question. As far as the psychological aspects are concerned, answered question askers tend to use more social, family, human related words on average compared to an open question asker. The open question askers express more positive emotions whereas the answered question asker tend to express more negative emotions in their texts. Also, answered question askers are more emotionally involved and their questions reveal higher usage of anger, sadness, anxiety related words compared to that of open questions. Open questions, on the other hand, contains more sexual, body, health related words which might be reasons why they do not attract answers.\nIn table 2 , we show a collection of examples of open questions to illustrate that many of the above quantities based on the linguistic activities described in this section naturally correspond to the factors that human judges consider responsible for a question remaining unanswered. This is one of the prime reasons why these quantities qualify as appropriate indicators of answerability.\nPrediction model\nIn this section, we describe the prediction framework in detail. Our goal is to predict whether a given question after a time period $t$ will be answered or not.\nLinguistic styles of the question asker\nThe content and way of posing a question is important to attract answers. We have observed in the previous section that these linguistic as well as psycholinguistic aspects of the question asker are discriminatory factors. For the prediction, we use the following features:",6 "char_len": 9853,7 "document_hash": "c80e6ea2a8791382763603e4bebf2f505c20c72902dfe037d5594021bc540bcc",8 "questions": [9 {10 "question_id": "q00",11 "qtype": "qasper",12 "question": "Do the answered questions measure for the usefulness of the answer?"13 }14 ]15}