Considering wide use of Twitter as the source of information, reaching an interesting tweet for a user among a bunch of tweets is challenging. In this work we propose a Named Entity Recognition (NER) based user profile modeling for Twitter users and employ this model to generate personalized tweet recommendations. E↵ectiveness of the proposed method is shown through a set of experiments.
For named entities to be extracted successfully, the informal writing style in tweets has to be handled. Generally named entities are assumed as words written in uppercase or mixed case phrases where uppercased letters are at the beginning and ending, and almost all of the studies bases on this assumption. However, capitalization is not a strong indicator in tweet-like informal texts, sometimes even misleading. To extract named entities in tweets, the e↵ect of the informality of the tweets has to be minimized as possible. The preprocessing tasks applied can be divided into two logical group:. Pre-segmenting, and Correcting. Removal of links, hashtags, mentions, conjunctives, stop words, vocatives, slang words and elimination of punctuation are considered as presegmentation. It is assumed that parts in the texts before and after a redundant word, or a punctuation mark cannot form a named entity together, therefore every removal of a word is considered as it segments the tweet as well as punctuation does it naturally. Removal of repeating characters that are used to express a feeling such as exaggerating, or yelling, handling mistyping and asciification related problems are considered as correcting and can be thought of conversion of tweets from informal to formal. In the following subsections, we describe the NER and user profile modeling and recommendation steps in more detail.
In this study, the idea of segmenting a tweet text into a set of
phrases, each of which appears more than random occurence
[
Segmentation constitutes the core part of named entity recognition method. The aim here is to split a tweet into consecutive segments. Each segment contains at least one word. For the optimal segmentation, the following objective function is used, where F is the stickiness function, t is an individual tweet, and s represents a segment.
arg max F (t) =
s1...sn
n
X F (si)
i=1
(1)
Although the term stickiness is generally used for
expressing tendency of a user to stay longer on a web page by a
user, Li et. al defined it as the metric of a word group to be
seen together in documents frequently, or not [
arg max F (s) = F (s1) + F (s2)
s1,s2
(2)
Thus far, tweets are segmented making use of the stickiness
function. In the result of this phase, tweet segments, which
are candidate named entities, are obtained. These candidate
named entities have to be validated whether they are real
named entities or not, so that they can be used as an
indicator of the user’s interest. For this purpose, as explained
before, Wikipedia is chosen as a reference for a segment to be
a named entity, and a graph-based knowledge-base based on
Wikipedia is constructed. If the segment, which is actually
a candidate named entity, matches exactly with a Wikipedia
title in the constructed knowledge base, then it is accepted
to be a named entity. In case of inexact match, we use the
Levenshtein distance [
Named Entities At this step, named entities with their frequency counts in a tweet obtained from followees’ posts, and followees’ relative ranking obtained in data gathering phase is processed as shown in Figure 1. Using these data, a user interest model is generated. It is basically a graph based relationship model. Let G = (V, E) be a weighted labelled graph with the node set V and edge set E. Node set V is labelled with the label set L1 where L1 2 { Root, F ollowee, N amedEntity} and Edge set E is labelled with the label set L2 where L2 2 { F ollows, W rites}. In other words, a user interest model graph has three types of nodes; Root, Friend, Named Entity, along with two types of weighted edges; Writes, and Follows. Weight of Writes edge represents the appearance count of a named entity for a followere’s posts where weight of the Follows edge represents relative ranking of a followed. Therefore, a twitter profile is represented as Root node Follows one or many Followees, and a Followee node Writes one or many Named Entities. The structure of the graph is shown in Figure 2.
Determining whether a tweet is interesting or not is achieved by comparing NE representation of the tweet with the generated user interest model. This comparison results in a ranking of candidate tweets. As the first step, candidate tweets are processed to obtain their NE representations. NE representation of a tweet simply includes the NEs, and their frequency counts. In order to compare with the candidate tweet, user interest model has to be interpreted by including the ranking score factor of the friends. Every followee’s named entities and their appearance counts are first multiplied with the friend’s ranking, and then summed. Therefore, a set of named entities with their scores based on the user interest model is obtained. The mathematical interpretation to calculate the score of a single named entity is given in Equation 3, where SCNE represents the overall score of a named entity, C represents the frequency count of a named entity for a user, n represents the count of friends included in the user interest model, RR represents the relative ranking score of a followed, and U represents the user himself. With the same approach, the final score of all of the named entities appearing in the user interest model is calculated.
SCNE = n X RRi · Ci + RRU · CU i=1 After overall score is calculated for all of the named entities in the user interest model, final scores for candidate tweets are calculated in the following approach: Overall score of named entities in NE representation of a candidate tweet are multiplied with the frequency count in the NE representation of itself. This operation is done for every named entity in the tweet representation, and then by summing these values, final score of a candidate tweet is obtained. If a named entity in a candidate tweet’s NE representation, does not appear in the user interest model, its overall score is accepted as 0 and not taken into consideration assuming the user is not interested in the subject that particular named entity represents. Once final scores for all candidate tweets are calculated, candidate tweets are sorted in descending order, and hence, they are ranked.
SCT = m X SCNEi · CNEi i=1
To evaluate the system from recommendation point of view, two types of datasets as candidate tweets for recommendation and two types of user groups to recommend tweets are formed. The first dataset of candidate tweets, GN RL, is a general dataset containing 100 tweets crawled from newspapers’ Twitter accounts. The second dataset, P SN L is a personal dataset containing 100 tweets that are crawled from the followees of followees of the selected users. There are 10 users volunteered for this experiment where half of them are active Twitter users, whereas the other half are inactive Twitter users. Active Users are the users that use Twitter frequently, have retweeting and mentioning habits, and update followed list when necessary where Inactive Users do not post, retweet, or mention often, and do not update followee list frequently. Volunteered users are categorized on the basis of the information they provided about their Twitter usage habits.
For each user, user interest model is constructed under SCP measure on Wikipedia Corpus along with length normalization for stickiness function, which gives the best results according to the validation experiments. In addition, the best NT and NF values are experimentally obtained, therefore 20 followees and 10 tweets of each followed are included in the model. Candidate tweets are scored by comparing with user’s model as explained in Section 2.3 and then ranked. Meanwhile, each user is asked to classify and score tweets in GN RL and P SN L datasets. Volunteered users made a two-step evaluation on each tweet for each dataset. They are asked to mark the tweet as interesting or uninteresting, and then if the tweet is interesting, they are asked to score the tweet in the range of [1 3] where 1 is the least score, and 3 is the highest score for interestingness. In the (3) (4) Active Users Inactive Users Active Users
Classification Acc. (%) GNRL P SNL 47 49 42 39 36 37 43 36 49 47 43.40 41.60 68 64 66 61 62 56 71 72 72 65 67.80 63.60
54.10 Classification Acc. (%) GNRL P SNL 69 66 62 58 52 55 67 52 72 69 64.40 60.00 88 86 79 74 74 68 88 85 80 77 81.80 78 71.05
Ranking Acc. (nDCG)
GNRL P SNL 0.520 0.612 0.573 0.654 0.433 0.478 0.322 0.301 0.567 0.514 0.483 0.512 0.777 0.909 0.699 0.768 0.760 0.782 0.720 0.815 0.601 0.677 0.711 0.790 baseline method, followee rankings are neglected and hence every named entity has equal weight. Generated recommendations are compared against the user preferences in terms of classification, and ranking.
The results in Table 1 show that the baseline method is able to decide whether a tweet is interesting for a user or not with the accuracy of 54,10% on average with classification and 0,624 nDCG value on average with ranking, which are lower than the results of our system. The performance of the baseline method in some cases decreases down to 36% correct prediction at classification, and 0,322 nDCG value at ranking quality. On the other hand, the results shown in Table 2 shows that the proposed system is able to decide whether a tweet is interesting for a user or not with the accuracy of 71,05% on average for classification and 0,767 nDCG value on average for ranking. Given the suitable user habits, performance of the system increases up to the 88% correct prediction for classification, and 0,958 nDCG value at ranking quality. The comparison of two tables show that the proposed user interest modeling approach increases the performance.
This paper proposes a new approach to Twitter user modeling and tweet recommendation by making use of named entities extracted from tweets. A powerful aspect of NER approach adopted in this study, tweet segmentation, is that it does not require an annotated large volume of training data to extract named entities, therefore a huge overload of annotation is avoided. In addition, this approach is not dependent on the morphology of the language. Experimental results show that the proposed method is capable of deciding on tweets to be recommended according to the user’s interest. Experimental results show the applicability of the approach for recommending tweets.