WO2015053607A1

WO2015053607A1 - System and method for semantic-level sentiment analysis of text

Info

Publication number: WO2015053607A1
Application number: PCT/MY2014/000182
Authority: WO
Inventors: Anand Sadanandan ARUN; Weng Onn Kow; Benjamin CHU MIN XIAN; Lukose Dickson; Daniel Bahls
Original assignee: Mimos Bhd
Current assignee: Mimos Bhd
Priority date: 2013-10-10
Filing date: 2014-06-12
Publication date: 2015-04-16
Anticipated expiration: 2016-04-10
Also published as: MY167959A

Abstract

The present invention relates to a system and method for semantic level sentiment analysis. The system (100) comprises of a graph generator component (10), a semantic sentiment analyser component (20), a sentiment processor component (30), a sentiment dictionary (40), a sentiment taxonomy (50), a semantic sentiment patterns repository (60) and a propagation rules repository (70). The system (100) accepts text data as input and analyses sentiment in the text. The method enables semantically valid sentiment in terms of the entire text as well as the individual entities in the text.

Description

SYSTEM AND METHOD FOR SEMANTIC LEVEL SENTIMENT ANALYSIS

FIELD OF INVENTION

The present invention relates to a system and method for semantic level sentiment analysis.

BACKGROUND OF THE INVENTION

Sentiment analysis uses natural language processing which aims to determine the attitude of a speaker or writer about some topic or the overall contextual polarity of a document. Natural language processing deals with the actual text element and transforms it into a machine usable format. By using sentiment analysis, it allows a system to track products, brands and also people to determine whether they are viewed positively or negatively. Sentiment analysis is important as it allows a system to track any new product or brand perception besides tracking business reputations. It also allows individuals to get an opinion or review on something on a global scale. Recently, a lot of researches have been carried out to improve the sentiment analysis of text. An example of a system and method for sentiment analysis is disclosed in

United States Patent No. 2011/0137906. It involves the steps of extracting a snippet related to a subject from a content database; calculating a sentiment score for the snippet; classifying the snippet into sentiment category and classifying the snippets as positive, negative or neutral.

The common existing sentiment analysis techniques only classify whole chunks of text with respect to certain sentiment classes. There is no disclosure of finding out the sentiment towards the entities contained in the text. Therefore, there is a need to provide a system and method that addresses the above mentioned drawbacks of the existing sentiment analysis techniques.

SUMMARY OF INVENTION

The present invention relates to a system (100) and method for semantic level sentiment analysis. The system (100) comprises of a graph generator component (10) for constructing conceptual representation of a text; a sentiment dictionary (40); a sentiment taxonomy (50); and a propagation rules repository (70), wherein the system (100) is further characterised in that a semantic sentiment analyser component (20) for semantically analysing the conceptual structures and generates sentiments; a sentiment processor component (30) for receiving results from the semantic sentiment analyser component (20) and processing the received results; and a semantic sentiment patterns repository (60) which is a library of sentiment patterns stored as graph.

The method for semantic sentiment analysis is characterised by the steps of constructing conceptual representation of a text by a graph generator component (10); loading a terminology-box which consists of concepts and relations, concept- hierarchy and relation-hierarchy by a semantic sentiment analyser component (20); performing classification of the entire fact graph by the semantic sentiment analyser component (20); performing classification of objects involved in the fact graph by the semantic sentiment analyser component (20); performing in-graph attribute propagation graph by the semantic sentiment analyser component (20); processing the classification results graph by the semantic sentiment analyser component (20); and performing sentiment processing by a sentiment processor component (30). Preferably, performing classification of the entire fact graph by the semantic sentiment analyser (20) includes the steps of selecting pattern from the sentiment patterns repository (60); projecting the fact graph onto a pattern graph; extracting sentiment weight of the pattern graph; identifying the number of pattern graphs projected that the project graph can project to if no more patterns are found; and determining the classification of the entire fact graph by calculating the weighted majority vote.

Preferably, performing classification of objects involved in the fact graph by the semantic sentiment analyser (20) includes the steps of selecting all predefined classification targets; selecting sub-graphs involving the classification targets; determining sentiment classification of the sub-graphs by calculating the weighted majority vote; repeating the step of performing classification of the entire fact graph by the semantic sentiment analyser (20); and using the results of the sentiment classification to annotate the fact graph in creating a modified fact graph. Preferably, performing in-graph attribute propagation graph by the semantic sentiment analyser (20) includes the steps of selecting propagation rules from a propagation rules repository (70); selecting the matching rules based on the concepts in the modified fact graph; and applying rule to assign sentiment classification to the corresponding concepts.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.

FIG. 1 illustrates a system (100) for semantic level sentiment analysis according to an embodiment of the present invention. FIG. 2 (a) illustrates a flow chart of a conceptual structure generation from text according to an embodiment of the present invention.

FIG. 2 (b) illustrates an example of a conceptual structure generation from text. FIG. 3 illustrates a flow chart of a semantic sentiment analysis process according to an embodiment of the present invention.

FIG. 4 (a) illustrates a flow chart of the sentiment classification of the entire fact graph according to an embodiment of the present invention.

FIG. 4 (b) illustrates an example of the sentiment classification of the entire fact graph.

FIG. 5 (a) illustrates a flow chart of the sentiment classification of objects involved in the fact graph according to an embodiment of the present invention.

FIG. 5 (b) illustrates an example of the sentiment classification of objects involved in the fact graph. FIG. 5 (c) illustrates the result of the example from FIG. 5 (b). FIG. 6 (a) illustrates a flow chart of the in-graph attribute propagation according to an embodiment of the present invention. FIG. 6 (b) illustrates an example of the in-graph attribute propagation.

FIG. 6 (c) illustrates the result of the example from FIG. 6 (b).

FIG. 7 illustrates a flow chart of the sentiment processing according to an embodiment of the present invention.

DESCRIPTION OF THE PREFFERED EMBODIMENT

A preferred embodiment of the present invention will be described herein below with reference to the accompanying drawings. In the following description, well known functions or constructions are not described in detail since they would obscure the description with unnecessary detail.

Reference is made initially to FIG. 1 which illustrates a system (100) for semantic level sentiment analysis according to an embodiment of the present invention. The system (100) accepts text data as input and analyses sentiment in the text. It generates semantically valid sentiment in terms of the entire text as well as the individual entities in the text. The system (100) comprises of a graph generator component (10), a semantic sentiment analyser component (20), a sentiment processor component (30), a sentiment dictionary (40), a sentiment taxonomy (50), a semantic sentiment patterns repository (60) and a propagation rules repository (70).

The graph generator component (10) is used to construct conceptual representation of the text. The semantic sentiment analyser component (20) semantically analyses the conceptual structures and generates sentiments. The sentiment processor component (30) receives the results from the semantic sentiment analyser component (20) and processes the received results.

The sentiment dictionary (40) comprises of an enriched list if keywords that are associated to polarity values. Each keyword is mapped to one or more sentiment categories such as positive, negative, sad, angry, happy, confident, etc. The list may be enhances with synonyms which are extracted from a lexical database. The sentiment taxonomy (50) comprises of taxonomical representation of common terms to describe sentiments. The semantic sentiment pattern repository (60) is a library of sentiment patterns stored as graphs. These sentiment patterns tell how sentiment terms are related to a particular noun. A generated annotated semantic sentiment graph is composed of a conceptual graph query, g; a sentiment class, c e C; and a weight, w e R+. A collection of these graphs are stored in the semantic sentiment patterns repository (60). Finally, the propagation rules repository (70) comprises of a set of propagation rules representing how the sentiment classes may be assigned to related concepts.

Referring now to FIG. 2 (a), it illustrates a flow chart of the conceptual structure generation from text according to an embodiment of the present invention. Initially the graph generator component (10) accepts text as input as in step 101 before it generates conceptual representation of the text as in step 102. An example of a text and its corresponding conceptual representation is shown in FIG. 2 (b). Since the text "Mary gave John a boring book" contains a set of facts, a fact graph is generated to represent the conceptual structure.

Moving on to FIG. 3, it illustrates a flow chart of a sentiment analysis process according to an embodiment of the present invention. After the conceptual structure is constructed in step 102, the system (100) loads a terminology-box from the semantic sentiment analyser component (20) which comprises of concepts and relations, concept-hierarchy and relation-hierarchy as in step 210. It then selects the conceptual representation i.e. the fact graph generated in step 102 to be analysed from multiple dimensions as in step 220. The semantic sentiment analyser component (20) initially performs classification of the entire fact graphs as in step 230, wherein the entire fact graph is analysed to determine the sentiment classification before it performs classification of objects involved in the fact graph as in step 240. Next, the semantic sentiment analyser component (20) performs in- graph attribute propagation as in step 250 by propagating sentiments which are associated with a particular concept to other concept that are related. Lastly, the semantic sentiment analyser component (20) sends the three classification results i.e. the sentiment classification of the entire fact graph, the sentiment classification objects involved in the fact graph and the sentiment classification by the in-graph attribute propagation to the sentiment processor component (30) for further processing as in step 260.

FIG. 4 (a) illustrates a flow chart of the sentiment classification of the entire fact graph according to an embodiment of the present invention which further explains step 230 of FIG. 3. In analysing the entire fact graph to determine the sentiment classification, the fact graph is projected on to the sets of graphs in the semantic sentiment pattern repository (60) as in step 231 , wherein the sets of graphs are the sentiment patterns which are stored as graphs. This is achieved through graph projection technique which results in one or more pattern graphs, p that match the fact graph as in step 232. The graph projection technique measures the similarity between two graphs, wherein if a graph can be projected fully on to another graph or vice versa, it is defined as very similar. When a graph can be partially projected, it is defined as a partial match.

The weight of the pattern graph, w is retrieved from each of the matched pattern graph as in step 233, wherein the matched pattern graphs are the results of the graph projection technique which can either be fully projected or partially projected. If more patterns are found as in step 234, the steps are repeated from step 231 until no more patterns are found. The numbers of projections, f found are then identified for each pattern graph as in step 235. Finally, the sentiment classification of the input graph is determined using the equation below,

Cfacts = arg max ) _Wifj

Vi=i

where c represents the classification of the fact graph, w represents the weight of the pattern graph and f represents the number of projections or matches found as in step 236.

An example of this process is shown in FIG. 4 (b) wherein the example graph represents the sentence "Ben travelled from London to Paris in an ugly bus." Upon projecting this graph on to the semantic sentiment pattern repository, a matching pattern i.e. entity->attr->ugly is found. In other words, the attribute of the concept "entity" is ugly. By using the sentiment class and weight, the entire graph can be classified as negative. Entities are the objects of interest and each entity instance is described by a set of attributes that define its qualities, characteristics or properties.

The step 240 of FIG. 3 is further described in FIG. 5 (a), which illustrates a flow chart of the sentiment classification of objects involved in the fact graph according to an embodiment of the present invention. In this step, the sentiment classification is performed on the sub-graphs in the fact graph where a sub-graph is composed of at least one concept. Initially, a set of predefined classification target is selected as in step 241. For instance, in the case of entity being a classification target, the process is conducted to determine the sentiment classification of all the entities found in the fact graph, by determining the sub-graphs that represent the classification targets as in step 243. Sub-graphs are smaller graphs which are derived from the fact graph. These sub-graphs are extracted by selecting the classification target and the concepts that are directly linked to it. Next, the sentiment classification process is repeated from step 231 to step 236 until all the pattern graphs in the semantic sentiment patterns repository (60) are processed to identify the classification of that particular sub-graph. Finally, the results of the sentiment classification are used to annotate the fact graph in creating a modified fact graph as in step 244, wherein a modified fact graph is a fact graph that is associated with a sentiment classification.

An example of this process is shown in FIG. 5 (b). The conceptual graph for the sentence "The beautiful Anna travelled in an uncomfortable and ugly bus" is used as projection. Assuming that entity is a classification target, multiple pattern graphs can be identified and used for determining the sentiment classification. The results of the classification are shown in FIG. 5 (c) where it can be seen that the sentiments related to the entity Bus are negative i.e. ugly and uncomfortable whereas the one related to the entity Anna is positive i.e. beautiful. Referring now to FIG. 6 (a) which illustrates a flow chart of the in-graph attribute propagation according to an embodiment of the present invention, it further describes step 250 of FIG. 3. For this step, sentiment classifications are applied automatically by propagation based on propagation rules which are used to perform the in-graph attribute propagation as in step 251. There are no restrictions as to what properties are to be used in the propagation rule. It can use any of the properties in the relation hierarchy. For example in FIG. 6 (b), if concepts A and B are linked together with a "hasPart" relation, then the sentiment for B is assigned to A as well.

The modified fact graph generated from step 244 is analysed further by applying matching rules from the propagation rules repository (70) as in step 252. The rules are matched based on the concepts present in the fact graph. Upon applying the rules, further concepts other than the previously tagged ones can be classified as in step 253. Special markers, which are labels to indicate whether a particular concept is a target concept, are used to determine "Target" placeholders, wherein a target concept is a concept to which propagation rules are to be applied. The system uses these special markers to decide what should be carried forward or where to find the sentiment in the match i.e. the "Sentiment Class" and to which object it should be propagated to i.e. the "target." An example of the attribute propagation is shown in FIG. 6 (b) while the results of applying the above mentioned propagation rule on an annotated facts are shown in FIG. 6 (c). The concept storyline is annotated as positive according to step 240. Referring now to FIG. 6 (c), it can be deducted that "Movie A" can also be classified as positive by using the propagation rule due to the "hasPart" property.

Finally, FIG. 7 illustrates a flow chart of the sentiment processing according to an embodiment of the present invention. Initially, the sentiment processor component (30) receives the results from the semantic sentiment analyser as in step 301. It then consolidates the results to show the sentiment classification in terms of the whole sentence as well as for objects of interest in the sentence to produce the multidimensional sentiment analysis results as in step 302.

While embodiments of the invention have been illustrated and described, it is not intended that these embodiments illustrate and describe all possible forms of the invention. Rather, the words used in the specifications are words of description rather than limitation and various changes may be made without departing from the scope of the invention.

Claims

1. A system (100) for semantic level sentiment analysis comprises of:

a) a graph generator component (10) for constructing conceptual representation of a text;

b) a sentiment dictionary (40);

c) a sentiment taxonomy (50); and

d) a propagation rules repository (70),

wherein the system (100) is further characterised in that:

i. a semantic sentiment analyser component (20) for semantically analysing the conceptual structures and generates sentiments;

ii. a sentiment processor component (30) for receiving results from the semantic sentiment analyser component (20) and processing the received results; and

iii. a semantic sentiment patterns repository (60) which is a library of sentiment patterns stored as graph.

2. A method for semantic sentiment analysis is characterised by the steps of:

a) constructing conceptual representation of a text by a graph generator component (10);

b) loading a terminology-box which consists of concepts and relations, concept-hierarchy and relation-hierarchy by a semantic sentiment analyser component (20);

c) performing classification of the entire fact graph by the semantic sentiment analyser component (20);

d) performing classification of objects involved in the fact graph by the semantic sentiment analyser component (20);

e) performing in-graph attribute propagation graph by the semantic sentiment analyser component (20);

f) processing the classification results graph by the semantic sentiment analyser component (20); and

g) performing sentiment processing by a sentiment processor component (30). The method as claimed as claim 2, wherein performing classification of the entire fact graph by the semantic sentiment analyser (20) includes the steps of:

a) selecting pattern from the sentiment patterns repository (60);

b) projecting the fact graph onto a pattern graph;

c) extracting sentiment weight of the pattern graph;

d) identifying the number of pattern graphs projected that the project graph can project to if no more patterns are found; and

e) determining the classification of the entire fact graph by calculating the weighted majority vote.

The method as claimed as claim 2, wherein performing classification of objects involved in the fact graph by the semantic sentiment analyser (20) includes the steps of:

a) selecting all predefined classification targets;

b) selecting sub-graphs involving the classification targets;

c) determining sentiment classification of the sub-graphs by calculating the weighted majority vote;

d) repeating the step of performing classification of the entire fact graph by the semantic sentiment analyser (20); and

e) using the results of the sentiment classification to annotate the fact graph in creating a modified fact graph.

The method as claimed as claim 2, wherein performing in-graph attribute propagation graph by the semantic sentiment analyser (20) includes the steps of:

a) selecting propagation rules from a propagation rules repository (70); b) selecting the matching rules based on the concepts in the modified fact graph; and

c) applying rule to assign sentiment classification to the corresponding concepts.