Easychair Preprint Building a Question-Answering Chatbot Using

EasyChair Preprint № 4979 Building a Question-Answering Chatbot using Forum Data in the Semantic Space Khalil Mrini, Marc Laperrouza and Pierre Dillenbourg EasyChair preprints are intended for rapid dissemination of research results and are integrated with the rest of EasyChair. February 5, 2021 Building a Question-Answering Chatbot using Forum Data in the Semantic Space Khalil Mrini Marc Laperrouza Pierre Dillenbourg CHILI Lab College of Humanities CHILI Lab EPFL EPFL EPFL Switzerland Switzerland Switzerland [email protected] [email protected] [email protected] Abstract 700-dimension vector representation of sentences based on FastText. We build a conversational agent which Goal. In this paper, we attempt to combine both knowledge base is an online forum for conversational agents and conceptual semantic parents of autistic children. We col- representations by creating a Chatbot based on an lect about 35,000 threads totalling some online forum used by the autism community, aim- 600,000 replies, and label 1% of them ing to answer questions of parents of autistic chil- for usefulness using Amazon Mechanical dren. Turk. We train a Random Forest Clas- sifier using sent2vec features to label the Method. The Chatbot is created as follows: remaining thread replies. Then, we use word2vec to match user queries conceptu- 1. We collect the titles and posts of all threads ally with a thread, and then a reply with a in two forums of the same website; predefined context window. 2. Threads are filtered to keep those that have questions as titles; 1 Introduction 3. Part of the threads are selected and their A Chatbot is a software that interacts with users user-provided replies are manually labelled in conversations using natural language. As it can to grade their usefulness on Amazon Me- be trained in different ways, it can serve a vari- chanical Turk; ety of purposes, most notably question answering. Chatbots have become widely used as personal 4. A model is trained on the labels to predict assistants when speech processing techniques are the usefulness of the replies of the remain- included. Known examples include Amazon’s ing unlabelled threads, with features using Alexa and Apple’s Siri. sent2vec; One of the earliest chatbot platforms is MIT’s 5. The Chatbot is built to reply in real time using ELIZA (Weizenbaum, 1966), which used pattern word2vec matching to select the most similar matching and substitution to compose answers, question, and then answers are filtered with thereby giving a sense of understanding of the usefulness labels and word2vec matching; user’s query, even though it lacked contextualiza- tion. ELIZA inspired the widely used chatbot- Organization of the paper. We first review in building platform ALICE (Artificial Linguistic In- Section2 related work on Chatbots based on ternet Computer Entity) (Wallace, 2009). Online and conceptual semantic representations. Conceptual vector-based representations of Then, we describe our collected data set and its la- words such as word2vec (Mikolov et al., 2013) belling in Section3. Afterwards, Section5 details have quickly become ubiquitous in applications of the functioning of the Chatbot. Natural Language Processing. Facebook’s Fast- 2 Related Work Text (Bojanowski et al., 2016) uses compositional n-gram features to enrich word vectors. The Given the large amount of available online tex- sent2vec model (Pagliardini et al., 2017) is a tual data, there are many chatbots created with a knowledge base extracted from online discussion as well as n-grams that compose them, in a similar forums for the purpose of question answering. fashion to FastText. Cong et al.(2008) address the problem of ex- tracting question-answer pairs in online forums. 3 Data Collection To detect questions, sentences are first POS- 3.1 Forum Data Description tagged and then sequential patterns are labelled The forum used in this study is the Wrong Planet1 to indicate whether they correspond to questions. forum. It has been open since 2004 and counts A minimum support is set to mine these labelled more than 25,000 members. A study (Jordan, sequential patterns, along with a minimum con- 2010) described the conversations on these on- fidence. A graph propagation method is used to line autism forums as revealing ”eloquent, empa- select an answer from many candidates. thetic individuals”, with no social awkwardness Noticing that most existing chatbots have hard- like autistic people can express or feel in a face- coded scripts, Wu et al.(2008) devise an automatic to-face conversation. method to extract a chatbot knowledge base from We collected all threads from two forums on an online forum. They use rough set classifiers Wrong Planet: Parents’ Discussion and General with manually defined attributes, and experiment Autism Discussion. A thread can be defined as a as well with ensemble learning, yielding high re- discussion having a title, a first post, a number of call and precision scores. replies and a number of views. A reply can be de- Qu and Liu(2011) attempt to predict the useful- fined as a text response to the first post of a thread, ness of threads and posts in an online forum using and has textual content (a message), a timestamp labelled data. They train a Naive Bayes classifier (date of publication) and information about the au- and a Hidden Markov Model, with the latter re- thor (user name, age, gender, date of joining, loca- sulting in the highest F1 scores. Likewise, Huang tion, number of posts written, and a user level). et al.(2007) use labelled online forum data in the We filter the threads based on their titles, such training of an SVM model to extract the best pairs that only threads which titles are questions are of thread titles and replies. kept. A question is defined here by rules: a sen- Finally, Boyanov et al.(2017) fine-tune tence is a question if it ends with a question mark, word2vec embeddings to select question-answer starts with a verb (modal or not) or an interroga- pairs based on cosine similarity. They train a tive adverb, and is followed by at least one other seq2seq model and evaluate it with the Machine verb. Translation evaluation system BLEU and with After this filtering, we remove threads with- MAP. out replies and obtain 35,807 threads, totalling Word embeddings can be used in chatbots to 603,185 replies. model questions and answers at the conceptual level. They have become widespread in NLP ap- 3.2 Partial Labelling of Data plications. To provide the most problem-solving answers to The word2vec embeddings (Mikolov et al., each question, we need some sort of judgment. We 2013) model words as vectors of real numbers. use Amazon Mechanical Turk to obtain labels for These vectors model the contexts of words, such a part of the data set. that cosine similarity between two vectors stems For each task, like in Figure1, we present the from the context similarity of their respective thread title, the rst post of the thread and one re- words. Ultimately, cosine similarity between two ply. The worker is asked to rate the reply in rela- word vectors represents how conceptually similar tion with the thread title and its rst post in terms of they are. usefulness. There are five varying levels of useful- Facebook’s FastText (Bojanowski et al., 2016) ness: Useless, Somewhat Useless, Neutral, Some- enrich word2vec embeddings by adding composi- what Useful, Useful. tional n-gram features, that assume that parts of a Since we were limited to a budget of USD 300.- word can determine its conceptual representation. and that each task costs USD 0.03, we can only The sent2vec model (Pagliardini et al., 2017) label a maximum of 10,000 replies. Therefore, we are trained in an unsupervised manner, and build had to optimise the choice of threads to label, as sentence representations by looking at unigrams, 1Available at http://wrongplanet.net/forums/ Figure 1: Example of an Amazon Mechanical Turk task for data labelling. Label Count Percentage bedding of the reply Useful 4,127 42.29% Somewhat Useful 3,411 34.96% • Cosine similarity between the sent2vec em- Neutral 1,143 11.71% bedding of the title and the first post and the Somewhat Useless 670 6.87% sent2vec embedding of the reply Useless 407 4.17% • The number of characters of the reply Total 9,758 100.00% • The number of sentences in the reply Table 1: Distribution of the labels obtained with • Amazon Mechanical Turk. The author’s user level • The number of posts of the author they would become our training set. We estimated the number of threads which replies we can label 4.2 Results and Discussion at 1% of the total. We trained a variety of models using the above To get this sample, we have to get thread titles features on a training set of 80% and a test set that are not statistically exceptions. Thus we limit of 20%. We obtained the results in Table2. We the candidate threads to the ones that have at least notice that the Random Forest Classifier performs 8 replies, and which first post does not exceed a best, except in precision where it is slightly out- threshold of 1054 characters. Then, we have to performed by SVM. Accuracy and F1-Scores are get thread titles that are as far apart conceptually not so high as it is a multi-label classification, and as possible To do so, we model each of the thread there must also be errors in labelling, or at least titles as a sent2vec vector, and make 373 clusters subjectivity bias. We use therefore the Random of threads using k-means.

Load more