Instance disambiguation is to obtain the concept of the target instance in context, which has been attracting much attention from academia. Existing methods are highly dependent on similar or related instances in context. However, the number of instances that can be extracted from a user-generated short text is limited. To tackle this problem, we propose a distributed semantic model (DSM) based method, which consists of three parts. 1) Measuring the correlation between contextual terms and each concept of the ambiguous instance based on DSMs; 2) Filtering out uninformative terms based on the correlations distribution over the concepts, which reduces noise interference; 3) Prioritizing the informative terms to highlight their discriminating capabilities. The concept with the maximum correlation score is considered as the meaning of the target instance. Experiment results demonstrate that the proposed method outperforms baseline methods.
In recent years, user-generated short texts (UGSTs) swept the world at an alarming rate. The study of these data could bring tremendous value for business organizations. To fully exploit these data, we need to understand them better. However, there are some ambiguous instances in UGSTs, which has a great impact on understanding. Therefore, instance disambiguation has been attracting much attention from academia.
Many scholars attempt to eliminate ambiguity based on instances [6] in context. However, an inevitable challenge is the number of instances contained in a UGST is limited. Recently, some eforts have been made to learn knowledge from the context of target instance ∗Yongjun Li is the corresponding author. to improve the performance of disambiguation [1–3]. Generally, there are two strategies. The first one is to use statistical models to obtain the topic of the UGST, and then determine the meaning of the ambiguous instance based on the topic [3]. Due to the sparsity of textual content, building an efective statistical model may not be easy. The second strategy is to use other types of terms for help. Wen et al. [1] found that verbs and adjectives are also helpful for disambiguation. Thus, they constructed a co-occurrence network for typed terms, and then chose the most related contextual term for disambiguation. However, the co-occurrence networks are word-based, which cannot apply to multi-word expressions (MWEs).
In this paper, we propose an Instance Disambiguation method with Context Awareness (IDwCA), which focuses on utilizing various types of contextual terms for disambiguation. Generally, some contextual terms cannot provide us with useful disambiguation information. For convenience, we call them uninformative terms. Otherwise, they are informative terms. To avoid noise interference, we calculate the correlation between contextual terms and each concept of the target instance to filter out uninformative terms. An important basis is the measurement of correlation. The DSMs and Probase are used in the measurement of correlation, which is efective and lightweight. Further, for the remaining contextual terms (informative terms), we prioritize each term to highlight their discrimination. Finally, we recalculated the correlation between informative terms and each concept of the target instance. The concept with the maximum score is considered as the meaning of the target instance. Experiments on ground-truth datasets illustrate the superiority of IDwCA over the-state-of-art methods.
A term t is a word or a MWE. In this paper, we only consider noun terms, verb (v) terms and adjective (ad j) terms, which are very helpful for disambiguation. In addition, for noun terms, we refine them into instances and concepts. While an instance e is a concrete object and a concept c is a general and abstract description of a set of instances. For example, "banana" and "grape" are instances, and they can be explained by the concept "fruit".
Problem Formulation 1. Instance disambiguation. Given a UGST T = {t1, t2, ..., tm }, wherein ti denotes a term. Assume term tk is an ambiguous term, and its candidate concept set is denoted by C = {cj |j = 1, 2, ..., l }. We define tk as the target instance and other terms in T as contextual terms for tk . The task of IDwCA is to identify the most approximate concept of tk from C.
The key issue of Problem 1 is to select related terms that have high discriminating capabilities for disambiguation. The main difference from existing work is that we use the corpus and knowledge information together to measure the semantic correlation of terms and then choose more types of contextual terms for disambiguation rather than solely relying on instances. 2.2
In IDwCA, first, DSMs and Probase are used to measure the correlation between all contextual terms and each concept of the target instance. Second, the Kullback Leiber (KL) divergence is employed to filter out uninformative terms. Then for the remaining informative terms, we prioritize them to highlight their discrimination. Finally, based on these informative terms, we obtain the concept of the target instance.
2.2.1 Correlation calculation between terms and concepts. We
could easily determine the most appropriate concept of the target
instance, if we have the knowledge about the semantic correlation
between contextual terms and concepts. We use DSMs for help,
which focuses on surrounding context of a word and is ideal for
calculating correlation. However, they cannot deal with MWEs. We
use semantic composition to solve this problem. Given a MWE,
denotes as p. Assume there are N words in p. Given the semantic
vector of each word, the vector of p can be calculated by Eq.(
N
Õ
c=1
v(wc )
That is, the vector of p is the sum of the vectors of all the words in
it. However, it ignores the syntactic relation between words and
may introduce too much noise. To solve this problem, we assign
weights to words based on their part-of-speech in p, where the
weights of nouns, verbs and adjectives are set to 1, and the rest is
set to 0. Then, the Eq.(
N Õ c=1 v(p) =
ac ∗ v(wc )
where ac denotes the weight of wc , ac ∈ {0, 1}. Finally, the cosine
metric is used to calculate the correlation, as shown in Eq.(
RD (t , c) = cos(v(t ), v(c))
Preliminary evaluation shows that the DSM-based method works reasonably well for many pairs of terms, but for some noun terms, the results are less satisfactory. We use Probase to fill this gap, which provides isA knowledge for concepts and instances, and two typicality scores for a concept/instance pair <c,e>: P (e |c) = n(c, e)/n(c) and P (c |e) = n(c, e)/n(e), where n(•) refers to the number of occurrences of a given term or a pair of terms in Probase. Following [5], we use the corresponding context of terms to calculate correlation.
Given a term t , we first extract its context St from Probase
according to its type. The context of term t is detailed as follows.
- If t is a concept, its context is all the instances that can be
explained by it.
(
(
After then, we transfer the context St into a vector It as shown in
Eq.(
( It = {P (ci1 |t )|i1 = 1, ..., m1}, t .type = e {P (ei2 |t )|i2 = 1, ..., m2}, t .type = c
Then, the measurement of correlation based on Probase can be
expressed as Eq.(
RP (t , c) = ( Íei2∈St ∩Sc P (ei2 |c)∗P (ei2 |t ) , t .type = c
| |It | |∗| |Ic | | Íci1 ∈St P (ci1 |t ) ∗ RP (ci1, c), t .type = e where || • || denotes the norm of a vector.
Finally, we use a strategy to integrate two parts linearly. In
summary, the semantic correlation between terms and concepts
can be calculated by Eq.(
( t .type ∈ {v, ad j } R(t , c) =
RD (t , c), θ ∗ RD (t , c) + (1 − θ ) ∗ RP (t , c), t .type ∈ {e, c } where θ is a tuning parameter.
2.2.2 Contextual term filtering. Normally, some contextual terms do not contains useful disambiguation information, so we filter them out to avoid noise interference. For clarity, we take "the apple is really delicious" as an example. Based on "delicious", we know "apple" is "a kind of fruit". This is because "delicious" is more related to "fruit" than to "company". However, if we filter out the uninformative terms directly according to the correlation scores, we need to set a threshold dynamically, which poses a big challenge. Following [1], we employ the KL divergence. First, we assume that the probabilities of concepts of the target instance are the same. That is, it fits a uniform distribution. Second we calculate the correlation between contextual terms and each concept, and normalize the scores to get a new distribution. Then, the KL divergence is used to measure the divergence between two distributions. The greater the divergence is, the more important the role of the term is. Finally, based on KL divergence, we set a threshold to filter out uninformative terms and obtain a new set of informative terms, denotes as ICT .
2.2.3 Weights of informative terms. Generally, the concept of
the target instance depends heavily on the choice of contextual
terms. Take "the engineer is eating the apple" as an example, the
ICT is {"engineer","eating"}, the concept of "apple" is "company"
according to "engineer", while its concept is "fruit" if based on
"eating". However, an ambiguous instance cannot has diferent
concepts simultaneously. To solve this problem, we prioritize each
informative term to highlight their contributions. Intuition is that
the closer the informative is to the target instance, the greater its
contribution. We propose a weighting function based on sigmoid,
which is described in Eq.(
1
weiдht (ti ) = 1.5 − 1 + e−x
(
Based on Eq.(
Õ R(ICT , c) = weiдht (tp ) ∗ R(tp , c) (8) tp ∈ICT The concept with the maximum score is the result of IDwCA. 3 3.1
As we know, there is no gold standard metric for evaluating instance disambiguation methods. Therefore, we evaluate our method in terms of classification. To verify the validity and generality of the method, we chose Foursquare, Twitter and Facebook as data sources. These social networking sites are popular sites and provide us with open data acquisition APIs. Then, we randomly selected UGSTs from the acquired data contained ambiguous instance "apple", "Harry Potter" and "python". We classified the data manually. For convenience, three datasets are abbreviated as FS, FB and TW, respectively. Table 1 shows the statistics of the ambiguous instance "apple" on three datasets.And the continuous Bag-of-Words model is used in our experiments to obtain the semantic vector of words, which is the one of the most commonly used DMSs. The wiki1 dataset is used for training the model. We compare our approach with the following representative methods: STC-NB [6] and TD [4].
Table 1: Details of FS, FB and TW
Datasets category
fruit company
FS We illustrate the results on three datasets in Figure 1. From the results, we reach the following conclusions. IDwCA outperforms all baselines, which validates its efectiveness. It is reasonable since IDwCA 1) utilizes information from DSMs and Probase to measure the semantic correlation, and then chooses various types of contextual terms for disambiguation, not just relying on instances; 2) assigns weights to informative terms based on their context distances, which reduces noise interference.
The STC-NB performs worse than other methods, because it only considers similar instances, and the correlation between terms are calculated by their co-occurrence times in Probase. Compared with IDwCA, TD achieves worse performances. This is because it divides terms into two types: instances and concepts, which may lead to wrong judgements. And its correlation calculation method does not work well in oral expressions. 3.3
Further, we explore the performance our correlation calculation method. We utilize two datasets in the following experiments: one
well-known dataset WordSim353 2 (WS) for words and one labeled
data WP for MWEs created by [5]. We compare our method with
the baseline algorithms. To evaluate the experiment, we computed
the Pearson Correlation Coeficient (PCC) to measure the machine
ratings and the human ratings over the two datasets. From the
results shown in Figure 3, we observe that IDwCA performs the
best on all datasets. This is because knowledge bases are more
suitable for noun-based terms than for other types of terms, and
IDwCA uses a combination of DSMs to solve ts problem. Meanwhile,
as shown in Eq.(
In this paper, we use DSMs and Probase to measure the correlation of terms and then choose various types of contextual terms for disambiguation. Experiments on ground-truth datasets validate the efectiveness of the proposed method.